Figure 1: ConlangCrafter outputs, showing random examples of sample sentences and glosses in diverse conlangs generated with our multi-hop LLM pipeline.Each language is designed to be internally consistent, while being typologically unique in its phonology and morpho-syntax.

Large vision-language models (VLMs) commonly process images at native or high resolution to remain effective across tasks.This inflates visual tokens up to to 99% of total tokens of the prefill stage, resulting in high compute and latency, even when lowresolution images would suffice.We introduce CARES-a Context-Aware Resolution Selector, a lightweight preprocessing module that, given an image-query pair, predicts the minimal sufficient input resolution.CARES uses a compact VLM (350M) to extract features and predict when a target pretrained VLM's response converges to its peak ability to answer correctly.Though trained as a discrete classifier over a set of optional resolutions, CARES interpolates continuous resolutions at inference for fine-grained control.Across nine multimodal benchmarks spanning documents and natural images, as well as diverse target VLMs, CARES preserves task performance while reducing compute by up to 78% on average across 9 benchmarks.

In Deep Image Prior (DIP), a Convolutional Neural Network (CNN) is fitted to map a latent space to a degraded (e.g. noisy) image but in the process learns to reconstruct the clean image. This phenomenon is attributed to CNN's internal image prior. We revisit the DIP framework, examining it from the perspective of a neural implicit representation. Motivated by this perspective, we replace the random latent with Fourier-Features (Positional Encoding). We empirically demonstrate that the convolution layers in DIP can be replaced with simple pixel-level MLPs thanks to the Fourier features properties. We also prove that they are equivalent in the case of linear networks. We name our scheme "Positional Encoding Image Prior" (PIP) and exhibit that it performs very similar to DIP on various image-reconstruction tasks with much fewer parameters. Furthermore, we demonstrate that PIP can be easily extended to videos, an area where methods based on image-priors and certain INR approaches face challenges with stability. Code and additional examples for all tasks, including videos, are available on the project page nimrodshabtay.github.io/PIP.

Anomaly detection in network traffic is critical for identifying deviations from normal behavior—including sophisticated cyber threats and previously unseen attacks—especially when anomalous examples are absent from the training data. The escalating complexity of cyber-attacks necessitates developing methods that not only identify low-likelihood traffic but also provide insights into its anomalous nature and deviations from normal behavior, enabling effective response and troubleshooting. In this work, we leverage the unique capabilities of normalizing flows (NF), a state-of-the-art reversible generative model for exact density estimation, to detect anomalies using only normal traffic. Our approach fundamentally differs from previous methods by utilizing NF’s exact likelihood computation for unsupervised detection and combining it with Shapley values to introduce a novel feature selection framework for guiding the selection of discriminative features in anomaly detection, while also providing statistically grounded enhanced explanations for detected anomalies, pinpointing potential root causes. Through experiments on CICIoT-2023, ISCXTor2016, and CICIDS2017, we demonstrate that our NF-based approach outperforms existing state-of-the-art methods for unsupervised anomaly detection. Notably, on the CICIoT-2023 dataset, we achieve an accuracy of 0.9951, comparable or higher than supervised methods, despite being trained solely on normal data.

Vision-language models (VLMs) excel at extracting and reasoning about information from images.Yet, their capacity to leverage internal knowledge about specific entities remains underexplored.This work investigates the disparity in model performance when answering factual questions about an entity described in text versus depicted in an image.Our results reveal a significant accuracy drop -reaching 18% for some models -when the entity is presented visually instead of textually.To study this gap we present POPVQA, a dataset which allows separating entity recognition and question answering, and use it to benchmark several models.We hypothesize that this decline arises from limitations in how information flows from image tokens to query tokens.Thus, we use mechanistic interpretability tools to reveal that, although image tokens are preprocessed by the vision encoder, meaningful information flow from these tokens occurs only in the much deeper layers.Furthermore, critical image processing happens in the language model's middle layers, allowing few layers for consecutive reasoning, highlighting a potential inefficiency in how the model utilizes its layers for reasoning.These insights shed light on the internal mechanics of VLMs and offer pathways for enhancing their reasoning capabilities.POPVQA can be found at this link.

Nir Endy, Idan Daniel Grosbard, Yuval Ran-Milo, Yonatan Slutzky, Itay Tshuva, Raja Giryes. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The increasing computational and memory demands in deep learning present significant challenges, especially in resource-constrained environments. We introduce a zero-order quantized optimization (ZOQO) method designed for training models with quantized parameters and operations. Our approach leverages zero-order approximations of the gradient sign and adapts the learning process to maintain the parameters’ quantization without the need for full-precision gradient calculations. We demonstrate the effectiveness of ZOQO through experiments in fine-tuning of large language models and black-box adversarial attacks. Despite the limitations of zero-order and quantized operations training, our method achieves competitive performance compared to full-precision methods, highlighting its potential for low-resource environments.

Vision-Language Models (VLMs) have shown remarkable capabilities across diverse visual tasks, including image recognition, video understanding, and Visual Question Answering (VQA) when explicitly trained for these tasks. Despite these advances, we find that present-day VLMs (including the proprietary GPT-4o) lack a fundamental cognitive ability: learning to localize specific objects in a scene by taking into account the context. In this work, we focus on the task of few-shot personalized localization, where a model is given a small set of annotated images (in-context examples) -- each with a category label and bounding box -- and is tasked with localizing the same object type in a query image. Personalized localization can be particularly important in cases of ambiguity of several related objects that can respond to a text or an object that is hard to describe with words. To provoke personalized localization abilities in models, we present a data-centric solution that fine-tunes them using carefully curated data from video object tracking datasets. By leveraging sequences of frames tracking the same object across multiple shots, we simulate instruction-tuning dialogues that promote context awareness. To reinforce this, we introduce a novel regularization technique that replaces object labels with pseudo-names, ensuring the model relies on visual context rather than prior knowledge. Our method significantly enhances the few-shot localization performance of recent VLMs ranging from 7B to 72B in size, without sacrificing generalization, as demonstrated on several benchmarks tailored towards evaluating personalized localization abilities. This work is the first to explore and benchmark personalized few-shot localization for VLMs -- exposing critical weaknesses in present-day VLMs, and laying a foundation for future research in context-driven vision-language applications.

Long-range sequence processing poses a significant challenge for Transformers due to their quadratic complexity in input length. A promising alternative is Mamba, which demonstrates high performance and achieves Transformer-level capabilities while requiring substantially fewer computational resources. In this paper we explore the length-generalization capabilities of Mamba, which we find to be relatively limited. Through a series of visualizations and analyses we identify that the limitations arise from a restricted effective receptive field, dictated by the sequence length used during training. To address this constraint, we introduce DeciMamba, a context-extension method specifically designed for Mamba. This mechanism, built on top of a hidden filtering mechanism embedded within the S6 layer, enables the trained model to extrapolate well even without additional training. Empirical experiments over real-world long-range NLP tasks show that DeciMamba can extrapolate to context lengths that are significantly longer than the ones seen during training, while enjoying faster inference. We will release our code and models.

The large-scale training of multi-modal models on data scraped from the web has shown outstanding utility in infusing these models with the required world knowledge to perform effectively on multiple downstream tasks. However, one downside of scraping data from the web can be the potential sacrifice of the benchmarks on which the abilities of these models are often evaluated. To safeguard against test data contamination and to truly test the abilities of these foundation models we propose LiveXiv: A scalable evolving live benchmark based on scientific ArXiv papers. LiveXiv accesses domain-specific manuscripts at any given timestamp and proposes to automatically generate visual question-answer pairs (VQA). This is done without any human-in-the-loop, using the multi-modal content in the manuscripts, like graphs, charts, and tables. Moreover, we introduce an efficient evaluation approach that estimates the performance of all models on the evolving benchmark using evaluations of only a subset of models. This significantly reduces the overall evaluation cost. We benchmark multiple open and proprietary Large Multi-modal Models (LMMs) on the first version of our benchmark, showing its challenging nature and exposing the models’ true abilities, avoiding contamination. Lastly, in our commitment to high quality, we have collected and evaluated a manually verified subset. By comparing its overall results to our automatic annotations, we have found that the performance variance is indeed minimal (<2.5%). Our dataset is available online on HuggingFace, and our code is available here.

Moran Yanuka, Assaf Ben-Kish, Yonatan Bitton, Idan Szpektor, Raja Giryes. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Proper initialization is critical for Recurrent Neural Networks (RNNs), particularly in long-range reasoning tasks, where repeated application of the same weight matrix can cause vanishing or exploding signals. A common baseline for linear recurrences is Glorot initialization, designed to ensure stable signal propagation---but derived under the infinite-width, fixed-length regime—an unrealistic setting for RNNs processing long sequences. In this work, we show that Glorot initialization is in fact unstable: small positive deviations in the spectral radius are amplified through time and cause the hidden state to explode. Our theoretical analysis demonstrates that sequences of length $t = O(\sqrt{n})$, where $n$ is the hidden width, are sufficient to induce instability. To address this, we propose a simple, dimension-aware rescaling of Glorot that shifts the spectral radius slightly below one, preventing rapid signal explosion or decay. These results suggest that standard initialization schemes may break down in the long-sequence regime, motivating a separate line of theory for stable recurrent initialization.

We propose a smooth regularization technique that instills a strong temporal inductive bias in video recognition models, particularly benefiting lightweight architectures. Our method encourages smoothness in the intermediate-layer embeddings of consecutive frames by modeling their changes as a Gaussian Random Walk (GRW). This penalizes abrupt representational shifts, thereby promoting low- acceleration solutions that better align with the natural temporal coherence inherent in videos. By leveraging this enforced smoothness, lightweight models can more effectively capture complex temporal dynamics. Applied to such models, our technique yields a 3.8%–6.4% accuracy improvement on Kinetics-600. Notably, the MoViNets model family trained with our smooth regularization improves the current state-of-the-art by 3.8%–6.1% within their respective FLOP constraints, while MobileNetV3 and the MoViNets-Stream family achieve gains of 4.9%–6.4% over prior state-of-the-art models with comparable memory footprints. Our code and models are available at https://github.com/cmusatyalab/grw-smoothing.

The lottery ticket hypothesis (LTH) has increased attention to pruning neural networks at initialization. We study this problem in the linear setting. We show that finding a sparse mask at initialization is equivalent to the sketching problem introduced for efficient matrix multiplication. This gives us tools to analyze the LTH problem and gain insights into it. Specifically, using the mask found at initialization, we bound the approximation error of the pruned linear model at the end of training. We theoretically justify previous empirical evidence that the search for sparse networks may be data independent. By using the sketching perspective, we suggest a generic improvement to existing algorithms for pruning at initialization, which we show to be beneficial in the data-independent case.

Abstract. Neural networks are widespread due to their powerful performance. Yet they degrade in the presence of noisy labels at training time. Inspired by the setting of learning with expert advice, where multiplicative weight (MW) updates were recently shown to be robust to moderate data corruptions in expert advice, we propose to use MW for reweighting examples during neural network optimization. We theoretically establish the convergence of our method when used with gradient descent and prove its advantages in one-dimensional cases. We then validate empirically our findings for the general case by showing that MW improves the accuracy of neural networks in the presence of label noise on CIFAR-10, CIFAR-100, and Clothing1M. We also show the impact of our approach on adversarial robustness.

Vision-Language models (VLMs) have proven to be effective at aligning image and text representations, producing superior zero-shot results when transferred to many downstream tasks. However, these representations suffer from some key shortcomings in understanding Compositional Language Concepts (CLC), such as recognizing objects' attributes, states, and relations between different objects. Moreover, VLMs typically have poor interpretability, making it challenging to debug and mitigate compositional-understanding failures. In this work, we introduce the architecture and training technique of Tree-augmented Vision-Language (3VL) model accompanied by our proposed Anchor inference method and Differential Relevance (DiRe) interpretability tool. By expanding the text of an arbitrary image-text pair into a hierarchical tree structure using language analysis tools, 3VL allows the induction of this structure into the visual representation learned by the model, enhancing its interpretability and compositional reasoning. Additionally, we show how Anchor, a simple technique for text unification, can be used to filter nuisance factors while increasing CLC understanding performance, e.g., on the fundamental VL-Checklist benchmark. We also show how DiRe, which performs a differential comparison between VLM relevancy maps, enables us to generate compelling visualizations of the reasons for a model's success or failure.

While recent years have seen rapid progress in image-conditioned text generation, image captioning still suffers from the fundamental issue of hallucinations, namely, the generation of spurious details that cannot be inferred from the given image.Existing methods largely use closed-vocabulary object lists to mitigate or evaluate hallucinations in image captioning, ignoring the long-tailed nature of hallucinations that occur in practice.To this end, we propose a framework for addressing hallucinations in image captioning in the open-vocabulary setting.Our framework includes a new benchmark, OpenCHAIR, that leverages generative foundation models to evaluate open-vocabulary object hallucinations for image captioning, surpassing the popular and similarly-sized CHAIR benchmark in both diversity and accuracy.Furthermore, to mitigate open-vocabulary hallucinations without using a closed object list, we propose MOCHa, an approach harnessing advancements in reinforcement learning.Our multi-objective reward function explicitly targets the trade-off between fidelity and adequacy in generations without requiring any strong supervision.MOCHa improves a large variety of image captioning models, as captured by our OpenCHAIR benchmark and other existing metrics.Code and models can be found in:

Phase-coded imaging is a computational imaging scheme designed to tackle tasks such as passive depth estimation and extended depth of field (EDOF) by inserting depth cues during image capture. Most of the current deep learning-based methods for depth estimation or all-in-focus imaging require a training dataset with high-quality depth maps and an optimal focus point at infinity for all-in-focus images. These datasets are difficult to create, often synthetic, and require external graphic software. Thus, limited datasets are being used, limiting the diversity of applicable domains. We propose a new method named “Deep Phase Coded Image Prior” (DPCIP) for jointly recovering the depth map and all-in-focus image from a coded-phase image using solely the captured image and the optical information of the imaging system. Our approach does not depend on any specific dataset and surpasses prior supervised techniques utilizing the same imaging system. This improvement is achieved through the utilization of a problem formulation based on implicit neural representation (INR) and deep image prior (DIP). Due to our zero-shot method, we overcome the barrier of acquiring accurate ground-truth data of depth maps and all-in-focus images for each new phase-coded system introduced. This allows focusing mainly on developing the imaging system, and not on ground-truth data collection.

In modern machine learning, models can often fit training data in numerous ways, some of which perform well on unseen (test) data, while others do not. Remarkably, in such cases gradient descent frequently exhibits an implicit bias that leads to excellent performance on unseen data. This implicit bias was extensively studied in supervised learning, but is far less understood in optimal control (reinforcement learning). There, learning a controller applied to a system via gradient descent is known as policy gradient, and a question of prime importance is the extent to which a learned controller extrapolates to unseen initial states. This paper theoretically studies the implicit bias of policy gradient in terms of extrapolation to unseen initial states. Focusing on the fundamental Linear Quadratic Regulator (LQR) problem, we establish that the extent of extrapolation depends on the degree of exploration induced by the system when commencing from initial states included in training. Experiments corroborate our theory, and demonstrate its conclusions on problems beyond LQR, where systems are non-linear and controllers are neural networks. We hypothesize that real-world optimal control may be greatly improved by developing methods for informed selection of initial states to train on.

Transformers have achieved state-of-the-art performance in language modeling tasks. However, the reasons behind their tremendous success are still unclear. In this paper, towards a better understanding, we train a Transformer model on a simple next token prediction task, where sequences are generated as a first-order autoregressive process $s_{t+1} = W s_t$. We show how a trained Transformer predicts the next token by first learning $W$ in-context, then applying a prediction mapping. We call the resulting procedure *in-context autoregressive learning*. More precisely, focusing on commuting orthogonal matrices $W$, we first show that a trained one-layer linear Transformer implements one step of gradient descent for the minimization of an inner objective function, when considering augmented tokens. When the tokens are not augmented, we characterize the global minima of a one-layer diagonal linear multi-head Transformer. Importantly, we exhibit orthogonality between heads and show that positional encoding captures trigonometric relations in the data. On the experimental side, we consider the general case of non-commuting orthogonal matrices and generalize our theoretical findings.

Denoising Diffusion Probabilistic Models (DDPM) have recently gained significant attention. DDPMs compose a Markovian process that begins in the data domain and gradually adds noise until reaching pure white noise. DDPMs generate high-quality samples from complex data distributions by defining an inverse process and training a deep neural network to learn this mapping. However, these models are inefficient because they require many diffusion steps to produce aesthetically pleasing samples. Additionally, unlike generative adversarial networks (GANs), the latent space of diffusion models is less interpretable. In this work, we propose to generalize the denoising diffusion process into an Upsampling Diffusion Probabilistic Model (UDPM). In the forward process, we reduce the latent variable dimension through downsampling, followed by the traditional noise perturbation. As a result, the reverse process gradually denoises and upsamples the latent variable to produce a sample from the data distribution. We formalize the Markovian diffusion processes of UDPM and demonstrate its generation capabilities on the popular FFHQ, AFHQv2, and CIFAR10 datasets. UDPM generates images with as few as three network evaluations, whose overall computational cost is less than a single DDPM or EDM step while achieving an FID score of 6.86. This surpasses current state-of-the-art efficient diffusion models that use a single denoising step for sampling. Additionally, UDPM offers an interpretable and interpolable latent space, which gives it an advantage over traditional DDPMs. Our code is available online: \url{https://github.com/shadyabh/UDPM/}

Structured state space models (SSMs), the core engine behind prominent neural networks such as S4 and Mamba, are linear dynamical systems adhering to a specified structure, most notably diagonal. In contrast to typical neural network modules, whose parameterizations are real, SSMs often use complex parameterizations. Theoretically explaining the benefits of complex parameterizations for SSMs is an open problem. The current paper takes a step towards its resolution, by establishing formal gaps between real and complex diagonal SSMs. Firstly, we prove that while a moderate dimension suffices in order for a complex SSM to express all mappings of a real SSM, a much higher dimension is needed for a real SSM to express mappings of a complex SSM. Secondly, we prove that even if the dimension of a real SSM is high enough to express a given mapping, typically, doing so requires the parameters of the real SSM to hold exponentially large values, which cannot be learned in practice. In contrast, a complex SSM can express any given mapping with moderate parameter values. Experiments corroborate our theory, and suggest a potential extension of the theory that accounts for selectivity, a new architectural feature yielding state of the art performance.

Although Deep Neural Networks (DNNs) achieve excellent performance on many real-world tasks, they are highly vulnerable to adversarial attacks. A leading defense against such attacks is adversarial training, a technique in which a DNN is trained to be robust to adversarial attacks by introducing adversarial noise to its input. This procedure is effective but must be done during the training phase. In this work, we propose Augmented Random Forest (ARF), a simple and easy-to-use strategy for robustifying an existing pretrained DNN without modifying its weights. For every image, we generate randomized test time augmentations by applying diverse color, blur, noise, and geometric transforms. Then we use the DNN’s logits output to train a simple random forest to predict the real class label. Our method achieves state-of-the-art adversarial robustness on a diversity of white and black box attacks with minimal compromise on the natural images’ classification. We test ARF also against numerous adaptive white-box attacks and it shows excellent results when combined with adversarial training. https://github.com/giladcohen/ARF.

Member inference (MI) attacks aim to determine if a specific data sample was used to train a machine learning model. Thus, MI is a major privacy threat to models trained on private sensitive data, such as medical records. In MI attacks one may consider the black-box settings, where the model’s parameters and activations are hidden from the adversary, or the white-box case where they are available to the attacker. In this work, we focus on the latter and present a novel MI attack for it that employs influence functions, or more specifically the samples’ self-influence scores, to perform MI prediction. The proposed method is evaluated on CIFAR-10, CIFAR-100, and Tiny ImageNet datasets using various architectures such as AlexNet, ResNet, and DenseNet. Our new attack method achieves new state-of-the-art (SOTA) results for MI even with limited adversarial knowledge, and is effective against MI defense methods such as data augmentation and differential privacy. Our code is available at https://github.com/giladcohen/sif_mi_attack.

Abstract We present SENS, a novel method for generating and editing 3D models from hand‐drawn sketches, including those of abstract nature. Our method allows users to quickly and easily sketch a shape, and then maps the sketch into the latent space of a part‐aware neural implicit shape architecture. SENS analyzes the sketch and encodes its parts into ViT patch encoding, subsequently feeding them into a transformer decoder that converts them to shape embeddings suitable for editing 3D neural implicit shapes. SENS provides intuitive sketch‐based generation and editing, and also succeeds in capturing the intent of the user's sketch to generate a variety of novel and expressive 3D shapes, even from abstract and imprecise sketches. Additionally, SENS supports refinement via part reconstruction, allowing for nuanced adjustments and artifact removal. It also offers part‐based modeling capabilities, enabling the combination of features from multiple sketches to create more complex and customized 3D shapes. We demonstrate the effectiveness of our model compared to the state‐of‐the‐art using objective metric evaluation criteria and a user study, both indicating strong performance on sketches with a medium level of abstraction. Furthermore, we showcase our method's intuitive sketch‐based shape editing capabilities, and validate it through a usability study.

Vision and Language ($VL$) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision & Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best$VL$models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing$VL$models' understanding of SVLCs that makes more effective use of existing$VL$pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching$VL$models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired$VL$datasets.$VL$models trained with the updated data exhibit a significant improvement of up to 15% in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model. Our code and pretrained models are available at: https://github.com/SivanDoveh/TSVLC

Text-guided image generation has progressed rapidly in recent years, inspiring major breakthroughs in text-guided shape generation. Recently, it has been shown that using score distillation, one can successfully text-guide a NeRF model to generate a 3D object. We adapt the score distillation to the publicly available, and computationally efficient, Latent Diffusion Models, which apply the entire diffusion process in a compact latent space of a pretrained autoencoder. As NeRFs operate in image space, a naive solution for guiding them with latent score distillation would require encoding to the latent space at each guidance step. Instead, we propose to bring the NeRF to the latent space, resulting in a Latent-NeRF. Analyzing our Latent-NeRF, we show that while Text-to-3D models can generate impressive results, they are inherently unconstrained and may lack the ability to guide or enforce a specific 3D structure. To assist and direct the 3D generation, we propose to guide our Latent-NeRF using a Sketch-Shape: an abstract geometry that defines the coarse structure of the desired object. Then, we present means to integrate such a constraint directly into a Latent-NeRF. This unique combination of text and shape guidance allows for increased control over the generation process. We also show that latent score distillation can be successfully applied directly on 3D meshes. This allows for generating high-quality textures on a given geometry. Our experiments validate the power of our different forms of guidance and the efficiency of using latent rendering.

Vision and Language (VL) models offer an effective method for aligning representation spaces of images and text allowing for numerous applications such as cross-modal retrieval, visual and multi-hop question answering, captioning, and many more. However, the aligned image-text spaces learned by all the popular VL models are still suffering from the so-called 'object bias' - their representations behave as 'bags of nouns' mostly ignoring or downsizing the attributes, relations, and states of objects described/appearing in texts/images. Although some great attempts at fixing these `compositional reasoning' issues were proposed in the recent literature, the problem is still far from being solved. In this paper, we uncover two factors limiting the VL models' compositional reasoning performance. These two factors are properties of the paired VL dataset used for finetuning (or pre-training) the VL model: (i) the caption quality, or in other words 'image-alignment', of the texts; and (ii) the 'density' of the captions in the sense of mentioning all the details appearing on the image. We propose a fine-tuning approach for automatically treating these factors on a standard collection of paired VL data (CC3M). Applied to CLIP, we demonstrate its significant compositional reasoning performance increase of up to $\sim27$\% over the base model, up to $\sim20$\% over the strongest baseline, and by $6.7$\% on average. Our code is provided in the Supplementary and would be released upon acceptance.

This work bridges two important concepts: the Neural Tangent Kernel (NTK), which captures the evolution of deep neural networks (DNNs) during training, and the Neural Collapse (NC) phenomenon, which refers to the emergence of symmetry and structure in the last-layer features of well-trained classification DNNs. We adopt the natural assumption that the empirical NTK develops a block structure aligned with the class labels, i.e., samples within the same class have stronger correlations than samples from different classes. Under this assumption, we derive the dynamics of DNNs trained with mean squared (MSE) loss and break them into interpretable phases. Moreover, we identify an invariant that captures the essence of the dynamics, and use it to prove the emergence of NC in DNNs with block-structured NTK. We provide large-scale numerical experiments on three common DNN architectures and three benchmark datasets to support our theory.

Abstract Despite recent advances in geometric modelling, 3D mesh modelling still involves a considerable amount of manual labour by experts. In this paper, we introduce Mesh Draping: a neural method for transferring existing mesh structure from one shape to another. The method drapes the source mesh over the target geometry and at the same time seeks to preserve the carefully designed characteristics of the source mesh. At its core, our method deforms the source mesh using progressive positional encoding (PE). We show that by leveraging gradually increasing frequencies to guide the neural optimization, we are able to achieve stable and high‐quality mesh transfer. Our approach is simple and requires little user guidance, compared to contemporary surface mapping techniques which rely on parametrization or careful manual tuning. Most importantly, Mesh Draping is a parameterization‐free method, and thus applicable to a variety of target shape representations, including point clouds, polygon soups and non‐manifold meshes. We demonstrate that the transferred meshing remains faithful to the source mesh design characteristics, and at the same time fits the target geometry well.

This work suggests using sampling theory to analyze the function space represented by interpolating mappings. While the analysis in this paper is general, we focus it on neural networks with bounded weights that are known for their ability to interpolate (fit) the training data. First, we show, under the assumption of a finite input domain, which is the common case in training neural networks, that the function space generated by multi-layer networks with bounded weights, and non-expansive activation functions are smooth. This extends over previous works that show results for the case of infinite width ReLU networks. Then, under the assumption that the input is band-limited, we provide novel error bounds for univariate neural networks. We analyze both deterministic uniform and random sampling showing the advantage of the former.

Deep networks usually require a massive amount of labeled data for their training. Yet, such data may include some mistakes in the labels. Interestingly, networks have been shown to be robust to such errors. This work uses spectral analysis of their learned mapping to provide an explanation for their robustness. In particular, we relate the smoothness regularization that usually exists in conventional training to the attenuation of high frequencies, which mainly characterize noise. By using a connection between the smoothness and the spectral norm of the network weights, we suggest that one may further improve robustness via spectral normalization. Empirical experiments validate our claims and show the advantage of this normalization for classification with label noise.

The ability to generalize learned representations across significantly different visual domains, such as between real photos, clipart, paintings, and sketches, is a fundamental capacity of the human visual system. In this paper, different from most cross-domain works that utilize some (or full) source domain supervision, we approach a relatively new and very practical Unsupervised Domain Generalization (UDG) setup of having no training supervision in neither source nor target domains. Our approach is based on self-supervised learning of a Bridge Across Domains (BrAD) - an auxiliary bridge domain accompanied by a set of semantics preserving visual (image-to-image) mappings to BrAD from each of the training domains. The BrAD and mappings to it are learned jointly (end-to-end) with a contrastive self-supervised representation model that semantically aligns each of the domains to its BrAD-projection, and hence implicitly drives all the domains (seen or unseen) to semantically align to each other. In this work, we show how using an edge-regularized BrAD our approach achieves significant gains across multiple benchmarks and a range of tasks, including UDG, Few-shot UDA, and unsupervised generalization across multi-domain datasets (including generalization to unseen domains and classes).

In this work, we suggest Kernel Filtering Linear Overparameterization (KFLO), where a linear cascade of filtering layers is used during training to improve network performance in test time. We implement this cascade in a kernel filtering fashion, which prevents the trained architecture from becoming unnecessarily deeper. This also allows using our approach with almost any network architecture and let combining the filtering layers into a single layer in test time. Thus, our approach does not add computational complexity during inference. We demonstrate the advantage of KFLO on various network models and datasets in supervised learning.

Ill-posed inverse problems appear in many image processing applications, such as deblurring and super-resolution. In recent years, solutions that are based on deep Convolutional Neural Networks (CNNs) have shown great promise. Yet, most of these techniques, which train CNNs using external data, are restricted to the observation models that have been used in the training phase. A recent alternative that does not have this drawback relies on learning the target image using internal learning. One such prominent example is the Deep Image Prior (DIP) technique that trains a network directly on the input image with the least-squares loss. In this paper, we propose a new image restoration framework that is based on minimizing a loss function that includes a "projected-version" of the Generalized Stein Unbiased Risk Estimator (GSURE) and parameterization of the latent image by a CNN. We propose two ways to use our framework. In the first one, where no explicit prior is used, we show that the proposed approach outperforms other internal learning methods, such as DIP. In the second one, we show that our GSURE-based loss leads to improved performance when used within a plug-and-play priors scheme.

Abstract In this work, we present our strategy for camera control in dynamic scenes with multiple people (sports teams). We learn a generic model of the player dynamics offline in simulation. We use only a few sparse demonstrations of a user's camera control policy to learn a reward function to drive camera motion in an ongoing dynamic scene. Key to our approach is the creation of a low‐dimensional representation of the scene dynamics which is independent of the environment action and rewards, which enables learning the reward function using only a small number of examples. We cast the user‐specific control objective as an inverse reinforcement learning problem, aiming to learn an expert's intention from a small number of demonstrations. The learned reward function is used in combination with a visual model predictive controller (MPC). We learn a generic scene dynamics model that is agnostic to the user‐specific reward, enabling reusing the same dynamics model for different camera control policies. We show the effectiveness of our method on simulated and real soccer matches.

Abstract We present a technique for visualizing point clouds using a neural network. Our technique allows for an instant preview of any point cloud, and bypasses the notoriously difficult surface reconstruction problem or the need to estimate oriented normals for splat‐based rendering. We cast the preview problem as a conditional image‐to‐image translation task, and design a neural network that translates point depth‐map directly into an image, where the point cloud is visualized as though a surface was reconstructed from it. Furthermore, the resulting appearance of the visualized point cloud can be, optionally, conditioned on simple control variables (e.g., color and light). We demonstrate that our technique instantly produces plausible images, and can, on‐the‐fly effectively handle noise, non‐uniform sampling, and thin surfaces sheets.

The recent success of learning-based algorithms can be greatly attributed to the immense amount of annotated data used for training. Yet, many datasets lack annotations due to the high costs associated with labeling, resulting in degraded performances of deep learning methods. Self-supervised learning is frequently adopted to mitigate the reliance on massive labeled datasets since it exploits unlabeled data to learn relevant feature representations. In this work, we propose SS-StyleGAN, a self-supervised approach for image annotation and classification suitable for extremely small annotated datasets. This novel framework adds self-supervision to the StyleGAN architecture by integrating an encoder that learns the embedding to the StyleGAN latent space, which is well-known for its disentangled properties. The learned latent space enables the smart selection of representatives from the data to be labeled for improved classification performance. We show that the proposed method attains strong classification results using small labeled datasets of sizes 50 and even 10. We demonstrate the superiority of our approach for the tasks of COVID-19 and liver tumor pathology identification.

Neural implicit fields are quickly emerging as an attractive representation for learning based techniques. However, adopting them for 3D shape modeling and editing is challenging. We introduce a method for E diting I mplicit S hapes T hrough P art A ware G enera T ion, permuted in short as SPAGHETTI. Our architecture allows for manipulation of implicit shapes by means of transforming, interpolating and combining shape segments together, without requiring explicit part supervision. SPAGHETTI disentangles shape part representation into extrinsic and intrinsic geometric information. This characteristic enables a generative framework with part-level control. The modeling capabilities of SPAGHETTI are demonstrated using an interactive graphical interface, where users can directly edit neural implicit shapes. Our code, editing user interface demo and pre-trained models are available at github.com/amirhertz/spaghetti.

Recently, several deep learning approaches have been proposed for point cloud registration. These methods train a network to generate a representation that helps finding matching points in two 3D point clouds. Finding good matches allows them to calculate the transformation between the point clouds accurately. Two challenges of these techniques are dealing with occlusions and generalizing to objects of classes unseen during training. This work proposes DeepBBS, a novel method for learning a representation that takes into account the best buddy distance between points during training. Best Buddies (i.e., mutual nearest neighbors) are pairs of points nearest to each other. The Best Buddies criterion is a strong indication for correct matches that, in turn, leads to accurate registration. Our experiments show improved performance compared to previous methods. In particular, our learned representation leads to an accurate registration for partial shapes and in unseen categories. Our code is publicly available1

Few-shot detection and classification have advanced significantly in recent years. Yet, detection approaches require strong annotation (bounding boxes) both for pre-training and for adaptation to novel classes, and classification approaches rarely provide localization of objects in the scene. In this paper, we introduce StarNet - a few-shot model featuring an end-to-end differentiable non-parametric star-model detection and classification head. Through this head, the backbone is meta-trained using only image-level labels to produce good features for jointly localizing and classifying previously unseen categories of few-shot test tasks using a star-model that geometrically matches between the query and support images (to find corresponding object instances). Being a few-shot detector, StarNet does not require any bounding box annotations, neither during pre-training nor for novel classes adaptation. It can thus be applied to the previously unexplored and challenging task of Weakly Supervised Few-Shot Object Detection (WS-FSOD), where it attains significant improvements over the baselines. In addition, StarNet shows significant gains on few-shot classification benchmarks that are less cropped around the objects (where object localization is key).

Few-shot learning methods offer pre-training techniques optimized for easier later adaptation of the model to new classes (unseen during training) using one or a few examples. This adaptivity to unseen classes is especially important for many practical applications where the pre-trained label space cannot remain fixed for effective use and the model needs to be "specialized" to support new categories on the fly. One particularly interesting scenario, essentially overlooked by the few-shot literature, is Coarse-to-Fine Few-Shot (C2FS), where the training classes (e.g. animals) are of much ‘coarser granularity’ than the target (test) classes (e.g. breeds). A very practical example of C2FS is when the target classes are sub-classes of the training classes. Intuitively, it is especially challenging as (both regular and few-shot) supervised pre-training tends to learn to ignore intra-class variability which is essential for separating sub-classes. In this paper, we introduce a novel ’Angular normalization’ module that allows to effectively combine supervised and self-supervised contrastive pre-training to approach the proposed C2FS task, demonstrating significant gains in a broad study over multiple baselines and datasets. We hope that this work will help to pave the way for future research on this new, challenging, and very practical topic of C2FS classification.

Nowadays, there is an abundance of data involving images and surrounding free-form text weakly corresponding to those images. Weakly Supervised phrase-Grounding (WSG) deals with the task of using this data to learn to localize (or to ground) arbitrary text phrases in images without any additional annotations. However, most recent SotA methods for WSG assume an existence of a pre-trained object detector, relying on it to produce the ROIs for localization. In this work, we focus on the task of Detector-Free WSG (DF-WSG) to solve WSG without relying on a pre-trained detector. The key idea behind our proposed Grounding by Separation (GbS) method is synthesizing ‘text to image-regions’ associations by random alpha-blending of arbitrary image pairs and using the corresponding texts of the pair as conditions to recover the alpha map from the blended image via a segmentation network. At test time, this allows using the query phrase as a condition for a non-blended query image, thus interpreting the test image as a composition of a region corresponding to the phrase and the complement region. Our GbS shows an 8.5% accuracy improvement over previous DF-WSG SotA, for a range of benchmarks including Flickr30K, Visual Genome, and ReferIt, as well as a complementary improvement (above 7%) over the detector-based approaches for WSG.

The vast majority of image recovery tasks are ill-posed problems. As such, methods that are based on optimization use cost functions that consist of both fidelity and prior (regularization) terms. A recent line of works imposes the prior by the Regularization by Denoising (RED) approach, which exploits the good performance of existing image denoising engines. Yet, the relation of RED to explicit prior terms is still not well understood, as previous work requires too strong assumptions on the denoisers. In this paper, we make two contributions. First, we show that the RED gradient can be seen as a (sub)gradient of a prior function—but taken at a denoised version of the point. As RED is typically applied with a relatively small noise level, this interpretation indicates a similarity between RED and traditional gradients. This leads to our second contribution: We propose to combine RED with the Back-Projection (BP) fidelity term rather than the common Least Squares (LS) term that is used in previous works. We show that the advantages of BP over LS for image deblurring and super-resolution, which have been demonstrated for traditional gradients, carry on to the RED approach.

Multilayer-perceptrons (MLP) are known to struggle learning functions of high-frequencies, and in particular, instances of wide frequency bands.We present a progressive mapping scheme for input signals of MLP networks, enabling them to better fit a wide range of frequencies without sacrificing training stability or requiring any domain specific preprocessing. We introduce Spatially Adaptive Progressive Encoding (SAPE) layers, which gradually unmask signal components with increasing frequencies as a function of time and space. The progressive exposure of frequencies is monitored by a feedback loop throughout the neural optimization process, allowing changes to propagate at different rates among local spatial portions of the signal space. We demonstrate the advantage of our method on variety of domains and applications: regression of low dimensional signals and images, representation learning of occupancy networks, and a geometric task of mesh transfer between 3D shapes.

Ill-posed linear inverse problems appear in many scientific setups and are typically addressed by solving optimization problems, which are composed of data fidelity and prior terms. Recently, several works have considered a back-projection (BP) based fidelity term as an alternative to the common least squares (LS) and demonstrated excellent results for popular inverse problems. These works have also empirically shown that using the BP term, rather than the LS term, requires fewer iterations of optimization algorithms. In this paper, we examine the convergence rate of the projected gradient descent algorithm for the BP objective. Our analysis allows us to identify an inherent source for its faster convergence compared to using the LS objective, while making only mild assumptions. We also analyze the more general proximal gradient method under a relaxed contraction condition on the proximal mapping of the prior. This analysis further highlights the advantage of BP when the linear measurement operator is badly conditioned. Numerical experiments with both $\ell_1$-norm and GAN based priors corroborate our theoretical results.

We introduce a novel technique for neural point cloud consolidation which learns from only the input point cloud. Unlike other point up-sampling methods which analyze shapes via local patches, in this work, we learn from global subsets. We repeatedly self-sample the input point cloud with global subsets that are used to train a deep neural network. Specifically, we define source and target subsets according to the desired consolidation criteria (e.g., generating sharp points or points in sparse regions). The network learns a mapping from source to target subsets, and implicitly learns to consolidate the point cloud. During inference, the network is fed with random subsets of points from the input, which it displaces to synthesize a consolidated point set. We leverage the inductive bias of neural networks to eliminate noise and outliers, a notoriously difficult problem in point cloud consolidation. The shared weights of the network are optimized over the entire shape, learning non-local statistics and exploiting the recurrence of local-scale geometries. Specifically, the network encodes the distribution of the underlying shape surface within a fixed set of local kernels, which results in the best explanation of the underlying shape surface. We demonstrate the ability to consolidate point sets from a variety of shapes, while eliminating outliers and noise.

Establishing a consistent normal orientation for point clouds is a notoriously difficult problem in geometry processing, requiring attention to both local and global shape characteristics. The normal direction of a point is a function of the local surface neighborhood; yet, point clouds do not disclose the full underlying surface structure. Even assuming known geodesic proximity, calculating a consistent normal orientation requires the global context. In this work, we introduce a novel approach for establishing a globally consistent normal orientation for point clouds. Our solution separates the local and global components into two different sub-problems. In the local phase, we train a neural network to learn a coherent normal direction per patch ( i.e. , consistently oriented normals within a single patch). In the global phase, we propagate the orientation across all coherent patches using a dipole propagation. Our dipole propagation decides to orient each patch using the electric field defined by all previously orientated patches. This gives rise to a global propagation that is stable, as well as being robust to nearby surfaces, holes, sharp features and noise.

In the recent years, there has been a significant improvement in the quality of samples produced by (deep) generative models such as variational auto-encoders and generative adversarial networks. However, the representation capabilities of these methods still do not capture the full distribution for complex classes of images, such as human faces. This deficiency has been clearly observed in previous works that use pre-trained generative models to solve imaging inverse problems. In this paper, we suggest to mitigate the limited representation capabilities of generators by making them image-adaptive and enforcing compliance of the restoration with the observations via back-projections. We empirically demonstrate the advantages of our proposed approach for image super-resolution and compressed sensing.

This work revisits the task of training sequence tagging models with limited resources using transfer learning.We investigate several proposed approaches introduced in recent works and suggest a new loss that relies on sentence reconstruction from normalized embeddings.Specifically, our method demonstrates how by adding a decoding layer for sentence reconstruction, we can improve the performance of various baselines.We show improved results on the CoNLL02 NER and UD 1.2 POS datasets and demonstrate the power of the method for transfer learning with low-resources achieving 0.6 F1 score in Dutch using only one sample from it.The code is publicly available at: https://github.com/tperl/Low- Resource-

Automatic methods for Neural ArchitectureSearch (NAS) have been shown to produce state-of-the-art network models, yet, their main drawback is the computational complexity of the search process. As some primal methods optimized over a discrete search space, thousands of days of GPU were required for convergence. A recent approach is based on constructing a differentiable search space that enables gradient-based optimization, thus reducing the search time to a few days. While successful, such methods still include some incontinuous steps, e.g., the pruning of many weak connections at once. In this paper, we propose a differentiable search space that allows the annealing of architecture weights, while gradually pruning inferior operations, thus the search converges to a single output network in a continuous manner. Experiments on several vision datasets demonstrate the effectiveness of our method with respect to the search cost, accuracy and the memory footprint of the achieved model.

Deep neural networks (DNNs) are notorious for their vulnerability to adversarial attacks, which are small perturbations added to their input images to mislead their prediction. Detection of adversarial examples is, therefore, a fundamental requirement for robust classification frameworks. In this work, we present a method for detecting such adversarial attacks, which is suitable for any pre-trained neural network classifier. We use influence functions to measure the impact of every training sample on the validation set data. From the influence scores, we find the most supportive training samples for any given validation example. A k-nearest neighbor (k-NN) model fitted on the DNN's activation layers is employed to search for the ranking of these supporting training samples. We observe that these samples are highly correlated with the nearest neighbors of the normal inputs, while this correlation is much weaker for adversarial inputs. We train an adversarial detector using the k-NN ranks and distances and show that it successfully distinguishes adversarial examples, getting state-of-the-art results on six attack methods with three datasets. Code is available at https://github.com/giladcohen/NNIF_adv_defense.

Point clouds are a popular representation for 3D shapes. However, they encode a particular sampling without accounting for shape priors or non-local information. We advocate for the use of a hierarchical Gaussian mixture model (hGMM), which is a compact, adaptive and lightweight representation that probabilistically defines the underlying 3D surface. We present PointGMM, a neural network that learns to generate hGMMs which are characteristic of the shape class, and also coincide with the input point cloud. PointGMM is trained over a collection of shapes to learn a class-specific prior. The hierarchical representation has two main advantages: (i) coarse-to-fine learning, which avoids converging to poor local-minima; and (ii) (an unsupervised) consistent partitioning of the input shape. We show that as a generative model, PointGMM learns a meaningful latent space which enables generating consistent interpolations between existing shapes, as well as synthesizing novel shapes. We also present a novel framework for rigid registration using PointGMM, that learns to disentangle orientation from structure of an input shape.

The single image super-resolution task is one of the most examined inverse problems in the past decade. In the recent years, Deep Neural Networks (DNNs) have shown superior performance over alternative methods when the acquisition process uses a fixed known downscaling kernel---typically a bicubic kernel. However, several recent works have shown that in practical scenarios, where the test data mismatch the training data (e.g. when the downscaling kernel is not the bicubic kernel or is not available at training), the leading DNN methods suffer from a huge performance drop. Inspired by the literature on generalized sampling, in this work we propose a method for improving the performance of DNNs that have been trained with a fixed kernel on observations acquired by other kernels. For a known kernel, we design a closed-form correction filter that modifies the low-resolution image to match one which is obtained by another kernel (e.g. bicubic), and thus improves the results of existing pre-trained DNNs. For an unknown kernel, we extend this idea and propose an algorithm for blind estimation of the required correction filter. We show that our approach outperforms other super-resolution methods, which are designed for general downscaling kernels.

Recently, there have been several works on unsupervised learning for training deep learning based denoisers without clean images. Approaches based on Stein's unbiased risk estimator (SURE) have shown promising results for training Gaussian deep denoisers. However, their performance is sensitive to hyper-parameter selection in approximating the divergence term in the SURE expression. In this work, we briefly study the computational efficiency of Monte-Carlo (MC) divergence approximation over recently available exact divergence computation using backpropagation. Then, we investigate the relationship between smoothness of nonlinear activation functions in deep denoisers and robust divergence term approximations. Lastly, we propose a new divergence term that does not contain hyper-parameters. Both unsupervised training methods yield comparable performance to supervised training methods with ground truth for denoising on various datasets. While the former method still requires roughly tuned hyper parameter selection, the latter method removes the necessity of choosing one.

Blind deconvolution and demixing is the problem of reconstructing convolved signals and kernels from the sum of their convolutions. This problem arises in many applications, such as blind MIMO. In this work, we present a separable approach to blind deconvolution and demixing via convex optimization. Unlike previous works, our formulation allows separation into smaller optimization problems, allowing significantly improved complexity. We demonstrate the near optimal performance of our method, in accordance with the theoretical guarantees of the original, non-separable problem, under several normalization constraints.

Face verification is a fast-growing authentication tool for everyday systems, such as smartphones. While current 2D face recognition methods are very accurate, it has been suggested recently that one may wish to add a 3D sensor to such solutions to make them more reliable and robust to spoofing, e.g., using a 2D print of a person's face. Yet, this requires an additional relatively expensive depth sensor. To mitigate this, we propose a novel authentication system, based on slim grayscale coded light field imaging. We provide a reconstruction free fast anti-spoofing mechanism, working directly on the coded image. It is followed by a multi-view, multi-modal face verification network that given grayscale data together with a low-res depth map achieves competitive results to the RGB case. We demonstrate the effectiveness of our solution on a simulated 3D (RGBD) version of LFW, which will be made public, and a set of real faces acquired by a light field computational camera.

We treat the problem of color enhancement as an image translation task, which we tackle using both supervised and unsupervised learning. Unlike traditional image to image generators, our translation is performed using a global parameterized color transformation instead of learning to directly map image information. In the supervised case, every training image is paired with a desired target image and a convolutional neural network (CNN) learns from the expert retouched images the parameters of the transformation. In the unpaired case, we employ two-way generative adversarial networks (GANs) to learn these parameters and apply a circularity constraint. We achieve state-of-the-art results compared to both supervised (paired data) and unsupervised (unpaired data) image enhancement methods on the MIT-Adobe FiveK benchmark. Moreover, we show the generalization capability of our method, by applying it on photos from the early 20th century and to dark video frames.

Ill-posed linear inverse problems appear in many image processing applications, such as deblurring, superresolution and compressed sensing. Many restoration strategies involve minimizing a cost function, which is composed of fidelity and prior terms, balanced by a regularization parameter. While a vast amount of research has been focused on different prior models, the fidelity term is almost always chosen to be the least squares (LS) objective, that encourages fitting the linearly transformed optimization variable to the observations. In this paper, we examine a different fidelity term, which has been implicitly used by the recently proposed iterative denoising and backward projections (IDBP) framework. This term encourages agreement between the projection of the optimization variable onto the row space of the linear operator and the pseudoinverse of the linear operator ("back-projection") applied on the observations. We analytically examine the difference between the two fidelity terms for Tikhonov regularization and identify cases (such as a badly conditioned linear operator) where the new term has an advantage over the standard LS one. Moreover, we demonstrate empirically that the behavior of the two induced cost functions for sophisticated convex and non-convex priors, such as total-variation, BM3D, and deep generative models, correlates with the obtained theoretical analysis.

In this paper, we introduce Point2Mesh , a technique for reconstructing a surface mesh from an input point cloud. Instead of explicitly specifying a prior that encodes the expected shape properties, the prior is defined automatically using the input point cloud, which we refer to as a self-prior. The self-prior encapsulates reoccurring geometric repetitions from a single shape within the weights of a deep neural network. We optimize the network weights to deform an initial mesh to shrink-wrap a single input point cloud. This explicitly considers the entire reconstructed shape, since shared local kernels are calculated to fit the overall object. The convolutional kernels are optimized globally across the entire shape, which inherently encourages local-scale geometric self-similarity across the shape surface. We show that shrink-wrapping a point cloud with a self-prior converges to a desirable solution; compared to a prescribed smoothness prior, which often becomes trapped in undesirable local minima. While the performance of traditional reconstruction approaches degrades in non-ideal conditions that are often present in real world scanning, i.e. , unoriented normals, noise and missing (low density) parts, Point2Mesh is robust to non-ideal conditions. We demonstrate the performance of Point2Mesh on a large variety of shapes with varying complexity.

Recently, deep generative adversarial networks for image generation have advanced rapidly; yet, only a small amount of research has focused on generative models for irregular structures, particularly meshes. Nonetheless, mesh generation and synthesis remains a fundamental topic in computer graphics. In this work, we propose a novel framework for synthesizing geometric textures. It learns geometric texture statistics from local neighborhoods ( i.e. , local triangular patches) of a single reference 3D model. It learns deep features on the faces of the input triangulation, which is used to subdivide and generate offsets across multiple scales, without parameterization of the reference or target mesh. Our network displaces mesh vertices in any direction ( i.e. , in the normal and tangential direction), enabling synthesis of geometric textures, which cannot be expressed by a simple 2D displacement map. Learning and synthesizing on local geometric patches enables a genus-oblivious framework, facilitating texture transfer between shapes of different genus.

Example synthesis is one of the leading methods to tackle the problem of few-shot learning, where only a small number of samples per class are available. However, current synthesis approaches only address the scenario of a single category label per image. In this work, we propose a novel technique for synthesizing samples with multiple labels for the (yet unhandled) multi-label few-shot classification scenario. We propose to combine pairs of given examples in feature space, so that the resulting synthesized feature vectors will correspond to examples whose label sets are obtained through certain set operations on the label sets of the corresponding input pairs. Thus, our method is capable of producing a sample containing the intersection, union or set-difference of labels present in two input samples. As we show, these set operations generalize to labels unseen during training. This enables performing augmentation on examples of novel categories, thus, facilitating multi-label few-shot classifier learning. We conduct numerous experiments showing promising results for the label-set manipulation capabilities of the proposed approach, both directly (using the classification and retrieval metrics), and in the context of performing data augmentation for multi-label few-shot learning. We propose a benchmark for this new and challenging task and show that our method compares favorably to all the common baselines.

Many images shared over the web include overlaid objects, or visual motifs, such as text, symbols or drawings, which add a description or decoration to the image. For example, decorative text that specifies where the image was taken, repeatedly appears across a variety of different images. Often, the reoccurring visual motif, is semantically similar, yet, differs in location, style and content (e.g. text placement, font and letters). This work proposes a deep learning based technique for blind removal of such objects. In the blind setting, the location and exact geometry of the motif are unknown. Our approach simultaneously estimates which pixels contain the visual motif, and synthesizes the underlying latent image. It is applied to a single input image, without any user assistance in specifying the location of the motif, achieving state-of-the-art results for blind removal of both opaque and semi-transparent visual motifs.

Distance metric learning (DML) has been successfully applied to object classification, both in the standard regime of rich training data and in the few-shot scenario, where each category is represented by only a few examples. In this work, we propose a new method for DML that simultaneously learns the backbone network parameters, the embedding space, and the multi-modal distribution of each of the training categories in that space, in a single end-to-end training process. Our approach outperforms state-of-the-art methods for DML-based object classification on a variety of standard fine-grained datasets. Furthermore, we demonstrate the effectiveness of our approach on the problem of few-shot object detection, by incorporating the proposed DML architecture as a classification head into a standard object detection model. We achieve the best results on the ImageNet-LOC dataset compared to strong baselines, when only a few training examples are available. We also offer the community a new episodic benchmark based on the ImageNet dataset for the few-shot object detection task.

Sparse coding techniques for image processing traditionally rely on the processing of small overlapping patches separately, followed by averaging. This has the disadvantage that the reconstructed image no longer obeys the sparsity prior used in the processing. For this purpose convolutional sparse coding has been introduced, where a shift-invariant dictionary is used and the sparsity of the recovered image is maintained. Most such strategies target the $\ell_0$ “norm” or the $\ell_1$ norm of the whole image, which may create an imbalanced sparsity across various regions in the image. In order to face this challenge, the $\ell_{0,\infty}$ “norm” has been proposed as an alternative that “operates locally while thinking globally." The approaches taken for tackling the nonconvexity of these optimization problems have been using either a convex relaxation or local pursuit algorithms. In this paper, we present an efficient greedy method for sparse coding and dictionary learning, which is specifically tailored to $\ell_{0,\infty}$ and is based on matching pursuit. We demonstrate the usage of our approach in salt-and-pepper noise removal and image inpainting.

While deep neural networks exhibit state-of-the-art results in the task of image super-resolution (SR) with a fixed known acquisition process (e.g., a bicubic downscaling kernel), they experience a huge performance loss when the real observation model mismatches the one used in training. Recently, two different techniques suggested to mitigate this deficiency, i.e., enjoy the advantages of deep learning without being restricted by the training phase. The first one follows the plug-and-play (P&P) approach that solves general inverse problems (e.g., SR) by using Gaussian denoisers for handling the prior term in model-based optimization schemes. The second builds on internal recurrence of information inside a single image, and trains a super-resolver network at test time on examples synthesized from the low-resolution image. Our letter incorporates these two independent strategies, enjoying the impressive generalization capabilities of deep learning, captured by the first, and further improving it through internal learning at test time. First, we apply a recent P&P strategy to SR. Then, we show how it may become image-adaptive in test time. This technique outperforms the above two strategies on popular datasets and gives better results than other state-of-the-art methods in practical cases where the observation model is inexact or unknown in advance.

We present DeepISP, a full end-to-end deep neural model of the camera image signal processing (ISP) pipeline. Our model learns a mapping from the raw low-light mosaiced image to the final visually compelling image and encompasses low-level tasks such as demosaicing and denoising as well as higher-level tasks such as color correction and image adjustment. The training and evaluation of the pipeline were performed on a dedicated dataset containing pairs of low-light and well-lit images captured by a Samsung S7 smartphone camera in both raw and processed JPEG formats. The proposed solution achieves state-of-the-art performance in objective evaluation of PSNR on the subtask of joint denoising and demosaicing. For the full end-to-end pipeline, it achieves better visual quality compared to the manufacturer ISP, in both a subjective human assessment and when rated by a deep model trained for assessing image quality.

Inverse problems appear in many applications, such as image deblurring and inpainting. The common approach to address them is to design a specific algorithm for each problem. The Plug-and-Play (P&P) framework, which has been recently introduced, allows solving general inverse problems by leveraging the impressive capabilities of existing denoising algorithms. While this fresh strategy has found many applications, a burdensome parameter tuning is often required in order to obtain high-quality results. In this paper, we propose an alternative method for solving inverse problems using off-the-shelf denoisers, which requires less parameter tuning. First, we transform a typical cost function, composed of fidelity and prior terms, into a closely related, novel optimization problem. Then, we propose an efficient minimization scheme with a P&P property, i.e., the prior term is handled solely by a denoising operation. Finally, we present an automatic tuning mechanism to set the method's parameters. We provide a theoretical analysis of the method and empirically demonstrate its competitiveness with task-specific techniques and the P&P approach for image inpainting and deblurring.

We present a novel method for neural network quantization. Our method, named UNIQ , emulates a non-uniform k -quantile quantizer and adapts the model to perform well with quantized weights by injecting noise to the weights at training time. As a by-product of injecting noise to weights, we find that activations can also be quantized to as low as 8-bit with only a minor accuracy degradation. Our non-uniform quantization approach provides a novel alternative to the existing uniform quantization techniques for neural networks. We further propose a novel complexity metric of number of bit operations performed (BOPs), and we show that this metric has a linear relation with logic utilization and power. We suggest evaluating the trade-off of accuracy vs. complexity (BOPs). The proposed method, when evaluated on ResNet18/34/50 and MobileNet on ImageNet, outperforms the prior state of the art both in the low-complexity regime and the high accuracy regime. We demonstrate the practical applicability of this approach, by implementing our non-uniformly quantized CNN on FPGA.

The process of aligning a pair of shapes is a fundamental operation in computer graphics. Traditional approaches rely heavily on matching corresponding points or features to guide the alignment, a paradigm that falters when significant shape portions are missing. These techniques generally do not incorporate prior knowledge about expected shape characteristics, which can help compensate for any misleading cues left by inaccuracies exhibited in the input shapes. We present an approach based on a deep neural network, leveraging shape datasets to learn a shape-aware prior for source-to-target alignment that is robust to shape incompleteness. In the absence of ground truth alignments for supervision, we train a network on the task of shape alignment using incomplete shapes generated from full shapes for self-supervision. Our network, called ALIGNet , is trained to warp complete source shapes to incomplete targets, as if the target shapes were complete, thus essentially rendering the alignment partial-shape agnostic . We aim for the network to develop specialized expertise over the common characteristics of the shapes in each dataset, thereby achieving a higher-level understanding of the expected shape space to which a local approach would be oblivious. We constrain ALIGNet through an anisotropic total variation identity regularization to promote piecewise smooth deformation fields, facilitating both partial-shape agnosticism and post-deformation applications. We demonstrate that ALIGNet learns to align geometrically distinct shapes and is able to infer plausible mappings even when the target shape is significantly incomplete. We show that our network learns the common expected characteristics of shape collections without over-fitting or memorization, enabling it to produce plausible deformations on unseen data during test time.

Polygonal meshes provide an efficient representation for 3D shapes. They explicitly captureboth shape surface and topology, and leverage non-uniformity to represent large flat regions as well as sharp, intricate features. This non-uniformity and irregularity, however, inhibits mesh analysis efforts using neural networks that combine convolution and pooling operations. In this paper, we utilize the unique properties of the mesh for a direct analysis of 3D shapes using MeshCNN , a convolutional neural network designed specifically for triangular meshes. Analogous to classic CNNs, MeshCNN combines specialized convolution and pooling layers that operate on the mesh edges, by leveraging their intrinsic geodesic connections. Convolutions are applied on edges and the four edges of their incident triangles, and pooling is applied via an edge collapse operation that retains surface topology, thereby, generating new mesh connectivity for the subsequent convolutions. MeshCNN learns which edges to collapse, thus forming a task-driven process where the network exposes and expands the important features while discarding the redundant ones. We demonstrate the effectiveness of MeshCNN on various learning tasks applied to 3D meshes.

Accelerating iterative algorithms for solving inverse problems using neural networks have become a very popular strategy in the recent years. In this work, we propose a theoretical analysis that may provide an explanation for its success. Our theory relies on the usage of inexact projections with the projected gradient descent (PGD) method. It is demonstrated in various problems including image super-resolution.

Convolutional sparse coding using the ℓ0,∞norm has been described as “a problem that operates locally while thinking globally”. In this paper, we present a matching pursuit based greedy algorithm specifically tailored to the ℓ0,∞norm. We also propose a corresponding dictionary learning algorithm, which trains a local dictionary on a set of global images. Our approach is based on the convolutional relationship between the local dictionary and the global image. It operates locally while taking into account the global nature of the images. We demonstrate the usage of our proposed strategy for the task of image inpainting.

We propose a convolutional recurrent sparse auto-encoder model. The model consists of a sparse encoder, which is a convolutional extension of the learned ISTA (LISTA) method, and a linear convolutional decoder. Our strategy offers a simple strategy for learning a task-driven sparse convolutional dictionary (CD), and producing an approximate convolutional sparse code (CSC) over the learned dictionary. We trained the model to minimize reconstruction loss via gradient decent with back-propagation and have achieved competitve results to KSVD image denoising and to leading CSC methods in image inpainting requiring only a small fraction of their runtime.

Light field photography has been studied thoroughly in recent years. One of its drawbacks is the need for multi-lens in the imaging. To compensate that, compressed light field photography has been proposed to tackle the trade-offs between the spatial and angular resolutions. It obtains by only one lens, a compressed version of the regular multi-lens system. The acquisition system consists of a dedicated hardware followed by a decompression algorithm, which usually suffers from high computational time. In this work, we propose a computationally efficient neural network that recovers a high-quality color light field from a single coded image. Unlike previous works, we compress the color channels as well, removing the need for a CFA in the imaging system. Our approach outperforms existing solutions in terms of recovery quality and computational complexity. We propose also a neural network for depth map extraction based on the decompressed light field, which is trained in an unsupervised manner without the ground truth depth map.

In blind deblurring, the goal is to recover a latent sharp image from its blurred version when the blur kernel is unknown. In this case, natural image priors often lead to intractable algorithms or failures if used with maximum a posteriori (MAP) estimation. Therefore, the ruling approach is to start with estimating only the kernel, and then use it to recover the latent image via non-blind deblurring. While many blind deblurring works focus on the kernel estimation, we consider the second phase, where we build on the recently proposed Iterative Denoising and Backward Projections (IDBP) strategy. The proposed method uses an automatic parameters tuning mechanism, which can tune the parameters differently for each kernel and image, contrary to other deblurring algorithms that are restricted to a uniform tuning in the blind-deblurring setting. We demonstrate the advantages of our method over widely used deblurring algorithms.

Learning to classify new categories based on just one or a few examples is a long-standing challenge in modern computer vision. In this work, we propose a simple yet effective method for few-shot (and one-shot) object recognition. Our approach is based on a modified auto-encoder, denoted delta-encoder, that learns to synthesize new samples for an unseen category just by seeing few examples from it. The synthesized samples are then used to train a classifier. The proposed approach learns to both extract transferable intra-class deformations, or "deltas", between same-class pairs of training examples, and to apply those deltas to the few provided examples of a novel class (unseen during training) in order to efficiently synthesize samples from that new class. The proposed method improves the state-of-the-art of one-shot object-recognition and performs comparably in the few-shot case.

We propose a fully-convolutional neural-network architecture for image denoising which is simple yet powerful. Its structure allows to exploit the gradual nature of the denoising process, in which shallow layers handle local noise statistics, while deeper layers recover edges and enhance textures. Our method advances the state-of-the-art when trained for different noise levels and distributions (both Gaussian and Poisson). In addition, we show that making the denoiser class-aware by exploiting semantic class information boosts performance, enhances textures and reduces artifacts.

This paper studies the generalization error of invariant classifiers. In particular, we consider the common scenario where the classification task is invariant to certain transformations of the input, and that the classifier is constructed (or learned) to be invariant to these transformations. Our approach relies on factoring the input space into a product of a base space and a set of transformations. We show that whereas the generalization error of a non-invariant classifier is proportional to the complexity of the input space, the generalization error of an invariant classifier is proportional to the complexity of the base space. We also derive a set of sufficient conditions on the geometry of the base space and the set of transformations that ensure that the complexity of the base space is much smaller than the complexity of the input space. Our analysis applies to general classifiers such as convolutional neural networks. We demonstrate the implications of the developed theory for such classifiers with experiments on the MNIST and CIFAR-10 datasets.

The increasing demand for high image quality in mobile devices brings forth the need for better computational enhancement techniques, and image denoising in particular. To this end, we propose a new fully convolutional deep neural network architecture which is simple yet powerful and achieves state-of-the-art performance for additive Gaussian noise removal. Furthermore, we claim that the personal photo-collections can usually be categorized into a small set of semantic classes. However simple, this observation has not been exploited in image denoising until now. We show that a significant boost in performance of up to 0.4dB PSNR can be achieved by making our network class-aware, namely, by fine-tuning it for images belonging to a specific semantic class. Relying on the hugely successful existing image classifiers, this research advocates for using a class-aware approach in all image enhancement tasks.

In this paper, we propose a novel postprocessing technique for compression-artifact reduction. Our approach is based on posing this task as an inverse problem, with a regularization that leverages on existing state-of-the-art image denoising algorithms. We rely on the recently proposed Plug-and-Play Prior framework, suggesting the solution of general inverse problems via alternating direction method of multipliers, leading to a sequence of Gaussian denoising steps. A key feature in our scheme is a linearization of the compression-decompression process, so as to get a formulation that can be optimized. In addition, we supply a thorough analysis of this linear approximation for several basic compression procedures. The proposed method is suitable for diverse compression techniques that rely on transform coding. In particular, we demonstrate impressive gains in image quality for several leading compression methods-JPEG, JPEG2000, and HEVC.

Two complementary approaches have been extensively used in signal and image processing leading to novel results, the sparse representation methodology and the variational strategy. Recently, a new sparsity based model has been proposed, the cosparse analysis framework, which may potentially help in bridging sparse approximation based methods to the traditional total-variation minimization. Based on this, we introduce a sparsity based framework for solving overparameterized variational problems. The latter has been used to improve the estimation of optical flow and also for general denoising of signals and images. However, the recovery of the space varying parameters involved was not adequately addressed by traditional variational methods. We first demonstrate the efficiency of the new framework for one dimensional signals in recovering a piecewise linear and polynomial function. Then, we illustrate how the new technique can be used for denoising and segmentation of images.

Many applications have benefited remarkably from low-dimensional models in the recent decade. The fact that many signals, though high dimensional, are intrinsically low dimensional has given the possibility to recover them stably from a relatively small number of their measurements. For example, in compressed sensing with the standard (synthesis) sparsity prior and in matrix completion, the number of measurements needed is proportional (up to a logarithmic factor) to the signal's manifold dimension. Recently, a new natural low-dimensional signal model has been proposed: the cosparse analysis prior. In the noiseless case, it is possible to recover signals from this model, using a combinatorial search, from a number of measurements proportional to the signal's manifold dimension. However, if we ask for stability to noise or an efficient (polynomial complexity) solver, all the existing results demand a number of measurements, which is far removed from the manifold dimension, sometimes far greater. Thus, it is natural to ask whether this gap is a deficiency of the theory and the solvers, or if there exists a real barrier in recovering the cosparse signals by relying only on their manifold dimension. Is there an algorithm which, in the presence of noise, can accurately recover a cosparse signal from a number of measurements proportional to the manifold dimension? In this paper, we prove that there is no such algorithm. Furthermore, we show through the numerical simulations that even in the noiseless case convex relaxations fail when the number of measurements is comparable with the manifold dimension. This gives a practical counterexample to the growing literature on the compressed acquisition of signals based on manifold dimension.

Poisson noise appears in various imaging applications, such as low-light photography, medical imaging and space imaging. In many cases we may have occlusions in the received image in addition to the noise. Thus, the problem of Poisson denoising turns to be a Poisson inpainting one in which we need both to remove the noise and recover the values in the occluded locations. In this work we extend a recent novel Poisson denoising method for the task of image inpainting. To the best of our knowledge this is the first work that deals with the problem of Poisson inpainting.

The problem of Poisson denoising appears in various imaging applications, such as low-light photography, medical imaging, and microscopy. In cases of high SNR, several transformations exist so as to convert the Poisson noise into an additive-independent identically distributed. Gaussian noise, for which many effective algorithms are available. However, in a low-SNR regime, these transformations are significantly less accurate, and a strategy that relies directly on the true noise statistics is required. Salmon et al took this route, proposing a patch-based exponential image representation model based on Gaussian mixture model, leading to state-of-the-art results. In this paper, we propose to harness sparse-representation modeling to the image patches, adopting the same exponential idea. Our scheme uses a greedy pursuit with boot-strapping-based stopping condition and dictionary learning within the denoising process. The reconstruction performance of the proposed scheme is competitive with leading methods in high SNR and achieving state-of-the-art results in cases of low SNR.

Signal recovery from a given set of linear measurements using a sparsity prior has been a major subject of research in recent years. In this model, the signal is assumed to have a sparse representation under a given dictionary. Most of the work dealing with this subject has focused on the reconstruction of the signal's representation as the means for recovering the signal itself. This approach forced the dictionary to be of low coherence and with no linear dependencies between its columns. Recently, a series of contributions that focus on signal recovery using the analysis model find that linear dependencies in the analysis dictionary are in fact permitted and beneficial. In this paper we show theoretically that the same holds also for signal recovery in the synthesis case for the ℓ0-synthesis minimization problem. In addition, we demonstrate empirically the relevance of our conclusions for recovering the signal using an ℓ1-relaxation.

A simple method to generate a 2-D binary grid pattern, which allows for absolute and accurate self-location in a finite planar region, is proposed. The pattern encodes position information in a local way so that reading a small number of its black or white pixels at any place provides sufficient data from which the location can be decoded both efficiently and robustly.

A K*-sparse vector x* ∈ RNproduces measurements via linear dimensionality reduction as u = Φx* +n, where Φ ∈ RM×N(MMconsists of independent and identically distributed, zero mean Gaussian entries with variance σ2. An algorithm, after its execution, determines a vector x̃ that has K-nonzero entries, and satisfies ||u - Φx̃|| ≤ ϵ. How far can x̃ be from x*? When the measurement matrix Φ provides stable embedding to 2K-sparse signals (the so-called restricted isometry property), they must be very close. This paper therefore establishes worst-case bounds to characterize the distance ||x̃- x*|| based on the online meta information. These bounds improve the pre-run algorithmic recovery guarantees, and are quite useful in exploring various data error and solution sparsity trade-offs. We also evaluate the performance of some sparse recovery algorithms in the context of our bound.