Implicit Neural Representations (INRs) have revolutionized signal processing and computer vision by modeling signals as continuous, differentiable functions parameterized by neural networks. However, INRs are prone to the spectral bias problem, limiting their ability to retain high-frequency information, and often struggle with noise robustness. Motivated by recent trends in iterative refinement processes, we propose Iterative Implicit Neural Representations (I-INRs). This novel plug-and-play framework iteratively refines signal reconstructions to restore high-frequency details, improve noise robustness, and enhance generalization, ultimately delivering superior reconstruction quality. I-INRs integrate seamlessly into existing INR architectures with only a 0.5–2% increase in parameters. During reconstruction, the iterative refinement adds just 0.8–1.6% additional FLOPs over the baseline while delivering a substantial performance boost of up to +2.0 PSNR. Extensive experiments demonstrate that I-INRs consistently outperform WIRE, SIREN, and Gauss across various computer vision tasks, including image fitting, image denoising, and object occupancy prediction.

Diffusion-based generative models demonstrate state-of-the-art performance across various image synthesis tasks, yet their tendency to replicate and amplify dataset biases remains poorly understood. Although previous research has viewed bias amplification as an inherent characteristic of diffusion models, this work provides the first analysis of how sampling algorithms and their hyperparameters influence bias amplification. We empirically demonstrate that samplers for diffusion models – commonly optimized for sample quality and speed – have a significant and measurable effect on bias amplification. Through controlled studies with models trained on Biased MNIST, Multi-Color MNIST and BFFHQ, and with Stable Diffusion, we show that sampling hyperparameters can induce both bias reduction and amplification, even when the trained model is fixed. Source code is available at https://github.com/How-I-met-your-bias/how_i_met_your_bias.

Mental states, a complex union of cognitive, emotional, and perceptual conditions, fundamentally shape how individuals perceive and interact with their surroundings. Detecting these states is vital, as it reveals the underlying processes that govern behaviour and enables targeted interventions across diverse fields such as mental health, education, and human-computer interaction. Generalisability across subjects and trials is essential to ensure that these interventions are effective and reliable in varied real-world settings, thereby enhancing their practical applicability. In this paper, we introduce an end-to-end optimised pipeline for classifying mental states from electroencephalography (EEG) signals. Through quantitative studies of data preprocessing and feature enhancement of continuous data collected under less stringent conditions, our pipeline utilises specially designed, cutting-edge, lightweight classifiers and achieves new state-of-the-art performance. Specifically addressing the challenge of generalisability in EEG signal research, our pipeline demonstrates robust performance, achieving a peak accuracy of 79.1% and an average of 71.9% in cross-subject scenarios, and a high of 89.3% with an average of 85.4% in cross-trial evaluations.

Tuning hyperparameters such as learning rate decay and stop conditions is typically done by assessing loss on a held-out validation set. This work introduces neural velocity (NeVe), the rate of change in neuron transfer functions, as a novel indicator of model convergence. We leverage NeVe to create a dynamic training method where learning rates and stop conditions are adjusted based on neural velocity computed from an auxiliary dataset generated by sampling noise. Our experiments across multiple tasks and architectures show that our approach performs on par with traditional validation-based methods, outperforming them in some tasks. Our method does not require withholding data for validation, making it especially advantageous in data-limited scenarios. This work highlights the potential of neural velocity as a key metric for optimizing neural network training.

In deep learning, the conventional transfer learning paradigm involves fine-tuning a model pre-trained on a complex source task to adapt it to a simpler target task, capitalizing on abundant training data. Concurrently, the paradigm of neural network pruning has emerged as a powerful strategy for enhancing model efficiency, reducing complexity, and optimizing resource utilization. This paper focuses on pruned model transferability estimation for resource-constraint scenarios, where the goal is to rank the performance of pruned pre-trained models on a downstream task without fine-tuning. To this end, from a formal analysis of the intra-class mutual information between samples belonging to the same target class, we observe that, as pruning increases, a sweet phase naturally rises, where the model benefits from better features at the encoder’s output. From this, we derive a Transferability Estimation for Pruned Backbones (TEP-ones) that eases the choice of which pruned model (without the need to train the classifier) is the best candidate for transfer learning.

Capsule networks are biologically inspired neural networks that group neurons into vectors called capsules, each explicitly representing an object or one of its parts. The routing mechanism connects capsules in consecutive layers, forming a hierarchical structure between parts and objects, also known as a parse tree. Capsule networks often attempt to model all elements in an image, requiring large network sizes to handle complexities such as intricate backgrounds or irrelevant objects. However, this comprehensive modeling leads to increased parameter counts and computational inefficiencies. Our goal is to enable capsule networks to focus only on the object of interest, reducing the number of parse trees. We accomplish this with REM (Routing Entropy Minimization), a technique that minimizes the entropy of the parse tree-like structure. REM drives the model parameters distribution towards low entropy configurations through a pruning mechanism, significantly reducing the generation of intra-class parse trees. This empowers capsules to learn more stable and succinct representations with fewer parameters and negligible performance loss.

3D Gaussian Splatting (3DGS) has recently emerged as a pioneering approach in explicit scene rendering and computer graphics. Unlike traditional neural radiance field (NeRF) methods, which typically rely on implicit, coordinate-based models to map spatial coordinates to pixel values, 3DGS utilizes millions of learnable 3D Gaussians. Its differentiable rendering technique and inherent capability for explicit scene representation and manipulation positions 3DGS as a potential game-changer for the next generation of 3D reconstruction and representation technologies. This enables 3DGS to deliver real-time rendering speeds while offering unparalleled editability levels. However, despite its advantages, 3DGS suffers from substantial memory and storage requirements, posing challenges for deployment on resource-constrained devices. In this survey, we provide a comprehensive overview focusing on the scalability and compression of 3DGS. We begin with a detailed background overview of 3DGS, followed by a structured taxonomy of existing compression methods. Additionally, we analyze and compare current methods from the topological perspective, evaluating their strengths and limitations in terms of fidelity, compression ratios, and computational efficiency. Furthermore, we explore how advancements in efficient NeRF representations can inspire future developments in 3DGS optimization. Finally, we conclude with current research challenges and highlight key directions for future exploration.

Learned image compression (LIC) has drawn much attention recently as it outperforms standardized codecs in rate-distortion (RD) efficiency. However, an LIC model is typically trained for a specific RD tradeoff, and achieving a different target rate requires retraining the model and storing the weights as a whole, limiting the practical applicability of LIC. In this article, we introduce CALICE, a framework for achieving continuous bitrate control by plugging into a pre-trained LIC model a set of modular adapters. Unlike similar methods that require a distinct set of adapters for each target rate, our method achieves continuous bitrate control by modulating a single set of adapters via a scalar parameter \(\boldsymbol{\alpha}\) , with a total overhead of less than \(\mathbf{0.35}\boldsymbol{\%}\) of the parameters of the LIC model. This design enables efficient support for multiple distortion objectives by learning lightweight, distortion-aware adapters. We also extend our strategy beyond rate control, demonstrating its ability to provide fine-grained adaptation of perceptual quality along the distortion–perception tradeoff. To our knowledge, this is the first method that jointly addresses rate and perceptual control using a unified, low-cost strategy. We publicly released the code at https://github.com/EIDOSLAB/CALICE .

Hypergraphs are powerful mathematical structures that can model complex, high-order relationships in various domains, including social networks, bioinformatics, and recommender systems. However, generating realistic and diverse hypergraphs remains challenging due to their inherent complexity and lack of effective generative models. In this paper, we introduce a diffusion-based Hypergraph Generation (HYGENE) method that addresses these challenges through a progressive local expansion approach. HYGENE works on the bipartite representation of hypergraphs, starting with a single pair of connected nodes and iteratively expanding it to form the target hypergraph. At each step, nodes and hyperedges are added in a localized manner using a denoising diffusion process, which allows for the construction of the global structure before refining local details. Our experiments demonstrated the effectiveness of HYGENE, proving its ability to closely mimic a variety of properties in hypergraphs. To the best of our knowledge, this is the first attempt to employ diffusion models for hypergraph generation.

Today, deep neural networks are widely used since they can handle a variety of complex tasks. Their generality makes them very powerful tools in modern technology. However, deep neural networks are often overparameterized. The usage of these large models consumes a lot of computation resources. In this paper, we introduce a method called Till the Layers Collapse (TLC), which compresses deep neural networks through the lenses of batch normalization layers. By reducing the depth of these networks, our method decreases deep neural networks' computational requirements and overall latency. We validate our method on popular models such as Swin-T, MobileNet-V2, and RoBERTa, across both image classification and natural language processing (NLP) tasks.

Although deep neural networks are well-known for their outstanding performance in tackling complex tasks, their hunger for computational resources remains a significant hurdle, posing energy-consumption issues and restricting their deployment on resource-constrained devices, preventing their widespread adoption. In this paper, we present an optimal transport-based method to reduce the depth of over-parametrized deep neural networks, alleviating their computational burden. More specifically, we propose a new regularization strategy based on the Max-Sliced Wasserstein distance to minimize the distance between the intermediate feature distributions in the neural network. We show that minimizing this distance enables the complete removal of intermediate layers in the network, achieving better performance/depth trade-off compared to existing techniques. We assess the effectiveness of our method on traditional image classification setups and extend it to generative image models. Our code is available at https://github.com/VGCQ/LaCoOT.

Recently, Multi-modal Large Language Models (MLLMs) have shown remarkable effectiveness for multi-modal tasks due to their abilities to generate and understand cross-modal data. However, processing long sequences of visual tokens extracted from visual backbones poses a challenge for deployment in real-time applications. To address this issue, we introduce FOLDER, a simple yet effective plug-and-play module designed to reduce the length of the visual token sequence, mitigating both computational and memory demands during training and inference. Through a comprehensive analysis of the token reduction process, we analyze the information loss introduced by different reduction strategies and develop FOLDER to preserve key information while removing visual redundancy. We showcase the effectiveness of FOLDER by integrating it into the visual backbone of several MLLMs, significantly accelerating the inference phase. Furthermore, we evaluate its utility as a training accelerator or even performance booster for MLLMs. In both contexts, FOLDER achieves comparable or even better performance than the original models, while dramatically reducing complexity by removing up to 70% of visual tokens.

Efficient compression of low-bit-rate point clouds is critical for bandwidth-constrained applications. However, existing techniques mainly focus on high-fidelity reconstruction, requiring many bits for compression. This paper proposes a "Denoising Diffusion Probabilistic Model" (DDPM) architecture for point cloud compression (DDPM-PCC) at low bit-rates. A PointNet encoder produces the condition vector for the generation, which is then quantized via a learnable vector quantizer. This configuration allows to achieve a low bitrates while preserving quality. Experiments on ShapeNet and ModelNet40 show improved rate-distortion at low rates compared to standardized and state-of-the-art approaches. We publicly released the code at https://github.com/EIDOSLAB/DDPM-PCC.

On-device learning has emerged as a promising direction for AI development, particularly because of its potential to reduce latency issues and mitigate privacy risks associated with device-server communication, while improving energy efficiency. Despite these advantages, significant memory and computational constraints still represent major challenges for its deployment. Drawing on previous studies on low-rank decomposition methods that address activation memory bottlenecks in backpropagation, we propose a novel shortcut approach as an alternative. Our analysis and experiments demonstrate that our method can reduce activation memory usage, even up to $120.09\times$ compared to vanilla training, while also reducing overall training FLOPs up to $1.86\times$ when evaluated on traditional benchmarks.

Hyperparameter tuning, such as learning rate decay and defining a stopping criterion, often relies on monitoring the validation loss. This paper presents NeVe, a dynamic training approach that adjusts the learning rate and defines the stop criterion based on the novel notion of "neural velocity". The neural velocity measures the rate of change of each neuron’s transfer function and is an indicator of model convergence: sampling neural velocity can be performed even by forwarding noise in the network, reducing the need for a held-out dataset. Our findings show the potential of neural velocity as a key metric for optimizing neural network training efficiently.

The effectiveness of deep learning models in classification tasks is often challenged by the quality and quantity of training data whenever they are affected by strong spurious correlations between specific attributes and target labels. This results in a form of bias affecting training data, which typically leads to unrecoverable weak generalization in prediction. This paper addresses this problem by leveraging bias amplification with generated synthetic data only: we introduce Diffusing DeBias (DDB), a novel approach acting as a plug-in for common methods of unsupervised model debiasing, exploiting the inherent bias-learning tendency of diffusion models in data generation. Specifically, our approach adopts conditional diffusion models to generate synthetic bias-aligned images, which fully replace the original training set for learning an effective bias amplifier model to be subsequently incorporated into an end-to-end and a two-step unsupervised debiasing approach. By tackling the fundamental issue of bias-conflicting training samples’ memorization in learning auxiliary models, typical of this type of technique, our proposed method outperforms the current state-of-the-art in multiple benchmark datasets, demonstrating its potential as a versatile and effective tool for tackling bias in deep learning models. Code is available at https://github.com/Malga-Vision/DiffusingDeBias

3D models have recently been popularized by the potentiality of end-to-end training offered first by Neural Radiance Fields and most recently by 3D Gaussian Splatting models. The latter has the big advantage of naturally providing fast training convergence and high editability. However, as the research around these is still in its infancy, there is still a gap in the literature regarding the model's scalability. In this work, we propose an approach enabling both memory and computation scalability of such models. More specifically, we propose an iterative pruning strategy that removes redundant information encoded in the model. We also enhance compressibility for the model by including a differentiable quantization and entropy coding estimator in the optimization strategy. Our results on popular benchmarks showcase the effectiveness of the proposed approach and open the road to the broad deployability of such a solution even on resource-constrained devices.

Learned progressive image compression is gaining momentum as it allows improved image reconstruction as more bits are decoded at the receiver. We propose a progressive image compression method in which an image is first represented as a pair of base-quality and top-quality latent representations. Next, a residual latent representation is encoded as the element-wise difference between the top and base representations. Our scheme enables progressive image compression with element-wise granularity by introducing a masking system that ranks each element of the residual latent representation from most to least important, dividing it into complementary components, which can be transmitted separately to the decoder in order to obtain different reconstruction quality. The masking system does not add further parameters or complexity. At the receiver, any elements of the top latent representation excluded from the transmitted components can be independently replaced with the mean predicted by the hyperprior architecture, ensuring reliable reconstructions at any intermediate quality level. We also in-troduced Rate Enhancement Modules (REMs), which refine the estimation of entropy parameters using already decoded components. We obtain results competitive with state-of-the-art competitors, while significantly reducing computational complexity, decoding time, and number of parameters.

In recent years, Graph Neural Networks (GNNs) have demonstrated strong adaptability to various real-world challenges, with architectures such as Vision GNN (ViG) achieving state-of-the-art performance in several computer vision tasks. However, their practical applicability is hindered by the computational complexity of constructing the graph, which scales quadratically with the image size. In this paper, we introduce a novel Windowed vision Graph neural Network (WiGNet) model for efficient image processing. WiGNet explores a different strategy from previous works by partitioning the image into windows and constructing a graph within each window. Therefore, our model uses graph convolutions instead of the typical 2D convolution or self-attention mechanism. WiGNet effectively manages computational and memory complexity for large image sizes. We evaluate our method in the ImageNet-1k benchmark dataset and test the adaptability of WiGNet using the CelebA-HQ dataset as a downstream task with higher-resolution images. In both of these scenarios, our method achieves competitive results compared to previous vision GNNs while keeping memory and computational complexity at bay. WiGNet offers a promising solution toward the deployment of vision GNNs in real-world applications. We publicly released the code and pre-trained models at https://github.com/EIDOSLAB/WiGNet.

In end-to-end learned image compression, encoder and decoder are jointly trained to minimize a R + λD cost function, where λ controls the trade-off between rate of the quantized latent representation and image quality. Unfortunately, a distinct encoder-decoder pair with millions of parameters must be trained for each λ, hence the need to switch encoders and to store multiple encoders and decoders on the user device for every target rate. This paper proposes to exploit a differentiable quantizer designed around a parametric sum of hyperbolic tangents, called STanH, that relaxes the step-wise quantization function. STanH is implemented as a differentiable activation layer with learnable quantization parameters that can be plugged into a pre-trained fixed rate model and refined to achieve different target bitrates. Experimental results show that our method enables variable rate coding with comparable efficiency to the state-of-the-art, yet with significant savings in terms of ease of deployment, training time, and storage costs.

Neoteric works have shown that modern deep learning models can exhibit a sparse double descent phenomenon. Indeed, as the sparsity of the model increases, the test performance first worsens since the model is overfitting the training data; then, the overfitting reduces, leading to an improvement in performance, and finally, the model begins to forget critical information, resulting in underfitting. Such a behavior prevents using traditional early stop criteria. In this work, we have three key contributions. First, we propose a learning framework that avoids such a phenomenon and improves generalization. Second, we introduce an entropy measure providing more insights into the insurgence of this phenomenon and enabling the use of traditional stop criteria. Third, we provide a comprehensive quantitative analysis of contingent factors such as re-initialization methods, model width and depth, and dataset noise. The contributions are supported by empirical evidence in typical setups. Our code is available at https://github.com/VGCQ/DSD2.

Deep neural networks are characterized by multiple symmetrical, equi-loss solutions that are redundant. Thus, the order of neurons in a layer and feature maps can be given arbitrary permutations, without affecting (or minimally affecting) their output. If we shuffle these neurons, or if we apply to them some perturbations (like fine-tuning) can we put them back in the original order i.e. re-synchronize? Is there a possible corruption threat? Answering these questions is important for applications like neural network white-box watermarking for ownership tracking and integrity verification. We advance a method to re-synchronize the order of permuted neurons. Our method is also effective if neurons are further altered by parameter pruning, quantization, and fine-tuning, showing robustness to integrity attacks. Additionally, we provide theoretical and practical evidence for the usual means to corrupt the integrity of the model, resulting in a solution to counter it. We test our approach on popular computer vision datasets and models, and we illustrate the threat and our countermeasure on a popular white-box watermarking method.

In Learned Image Compression (LIC), a model is trained at encoding and decoding images sampled from a source domain, often outperforming traditional codecs on natural images; yet its performance may be far from optimal on images sampled from different domains. In this work, we tackle the problem of adapting a pre-trained model to multiple target domains by plugging into the decoder an adapter module for each of them, including the source one. Each adapter improves the decoder performance on a specific domain, without the model forgetting about the images seen at training time. A gate network computes the weights to optimally blend the contributions from the adapters when the bitstream is decoded. We experimentally validate our method over two state-of-the-art pre-trained models, observing improved rate-distortion efficiency on the target domains without penalties on the source domain. Furthermore, the gate’s ability to find similarities with the learned target domains enables better encoding efficiency also for images outside them.

While standardized codecs like JPEG and HEVC-intra represent the industry standard in image compression, neural Learned Image Compression (LIC) codecs represent a promising alternative. In detail, integrating attention mechanisms from Vision Transformers into LIC models has shown improved compression efficiency. However, extra efficiency often comes at the cost of aggregating redundant features. This work proposes a Graph-based Attention Block for Image Compression (GABIC), a method to reduce feature redundancy based on a k-Nearest Neighbors enhanced attention mechanism. Our experiments show that GABIC outperforms comparable methods, particularly at high bit rates, enhancing compression performance.

Internet of Things and Deep Learning are synergetically and exponentially growing industrial fields with a massive call for their unification into a common framework called Edge AI. While on-device inference is a well-explored topic in recent research, backpropagation remains an open challenge due to its prohibitive computational and memory costs compared to the extreme resource constraints of embedded devices. Drawing on tensor decomposition research, we tackle the main bottleneck of backpropagation, namely the memory footprint of activation map storage. We investigate and compare the effects of activation compression using Singular Value Decomposition and its tensor variant, High-Order Singular Value Decomposition. The application of low-order decomposition results in considerable memory savings while preserving the features essential for learning, and also offers theoretical guarantees to convergence. Experimental results obtained on main-stream architectures and tasks demonstrate Pareto-superiority over other state-of-the-art solutions, in terms of the trade-off between generalization and memory footprint.

When training a Learned Image Compression model, the loss function is minimized such that the encoder and the decoder attain a target Rate-Distorsion trade-off. Therefore, a distinct model shall be trained and stored at the transmitter and receiver for each target rate, fostering the quest for efficient variable bitrate compression schemes. This paper proposes plugging Low-Rank Adapters into a transformer-based pre-trained LIC model and training them to meet different target rates. With our method, encoding an image at a variable rate is as simple as training the corresponding adapters and plugging them into the frozen pre-trained model. Our experiments show performance comparable with state-of-the-art fixed-rate LIC models at a fraction of the training and deployment cost. We publicly released the code at https://github.com/EIDOSLAB/ALICE.

Vision Transformers (ViTs) have become one of the dominant architectures in computer vision, and pre-trained ViT models are commonly adapted to new tasks via finetuning. Recent works proposed several parameter-efficient transfer learning methods, such as adapters, to avoid the prohibitive training and storage cost of finetuning.In this work, we observe that adapters perform poorly when the dimension of adapters is small, and we propose MiMi, a training framework that addresses this issue. We start with large adapters which can reach high performance, and iteratively reduce their size. To enable automatic estimation of the hidden dimension of every adapter, we also introduce a new scoring function, specifically designed for adapters, that compares the neuron importance across layers. Our method outperforms existing methods in finding the best trade-off between accuracy and trained parameters across the three dataset benchmarks DomainNet, VTAB, and Multi-task, for a total of 29 datasets.1

Despite significant research efforts, deep neural networks remain vulnerable to biases: this raises concerns about their fairness and limits their generalization. In this paper, we propose a bias-agnostic approach to mitigate the impact of biases in deep neural networks. Unlike traditional debiasing approaches, we rely on a metric to quantify "bias alignment/misalignment" on target classes and use this information to discourage the propagation of bias-target alignment information through the network. We conduct experiments on several commonly used datasets for debiasing and compare our method with supervised and bias-specific approaches. Our results indicate that the proposed method achieves comparable performance to state-of-the-art supervised approaches, despite being bias-agnostic, even in the presence of multiple biases in the same sample.

Finding the optimal size of deep learning models is very actual and of broad impact, especially in energy-saving schemes. Very recently, an unexpected phenomenon, the "double descent", has caught the attention of the deep learning community. As the model’s size grows, the performance gets first worse and then goes back to improving. It raises serious questions about the optimal model’s size to maintain high generalization: the model needs to be sufficiently over-parametrized, but adding too many parameters wastes training resources. Is it possible to find, in an efficient way, the best trade-off?Our work shows that the double descent phenomenon is potentially avoidable with proper conditioning of the learning problem, but a final answer is yet to be found. We empirically observe that there is hope to dodge the double descent in complex scenarios with proper regularization, as a simple ℓ2regularization is already positively contributing to such a perspective.

In today’s digital age, multimedia content is omnipresent, and the demand for efficient compression techniques is ever-increasing. In particular, the successful delivery of services based on video transmission largely depends on achieving the lowest latency values. One solution has been to use extrapolation for latency compensation in video transmission that allows to reduce the latency by an arbitrary amount. Nevertheless, this latency reduction comes at the cost of an increased distortion of the displayed images, since they are based on temporal extrapolation. Latency can also be traded with coding rate. This paper introduces ELR-PSNR and EPR-Latency as unified metrics to assess the three-way trade-off between rate, distortion, and latency simultaneously.

NeRFs have revolutionized the world of per-scene radiance field reconstruction because of their intrinsic compactness. One of the main limitations of NeRFs is their slow rendering speed, both at training and inference time. Recent research focuses on the optimization of an explicit voxel grid (EVG) that represents the scene, which can be paired with neural networks to learn radiance fields. This approach significantly enhances the speed both at train and inference time, but at the cost of large memory occupation. In this work we propose Re:NeRF, an approach that specifically targets EVG-NeRFs compressibility, aiming to reduce memory storage of NeRF models while maintaining comparable performance. We benchmark our approach with three different EVG-NeRF architectures on four popular benchmarks, showing Re:NeRF’s broad usability and effectiveness.

Deep learning models are nowadays broadly deployed to solve an incredibly large variety of tasks. However, little attention has been devoted to connected legal aspects. In 2016, the European Union approved the General Data Protection Regulation which entered into force in 2018. Its main rationale was to protect the privacy and data protection of its citizens by the way of operating the so-called “Data Economy”. As data is the fuel of modern Artificial Intelligence, it is argued that the GDPR can be partly applicable to a series of algorithmic decision-making tasks before a more structured AI Regulation enters into force. In the meantime, AI should not allow undesired information leakage deviating from the purpose for which is created. In this work, we propose DisP, an approach for deep learning models disentangling the information related to some classes we desire to keep private, from the data processed by AI. In particular, DisP is a regularization strategy de-correlating the features belonging to the same private class at training time, hiding the information about private class membership. Our experiments on state-of-the-art deep learning models show the effectiveness of DisP, minimizing the risk of extraction for the classes we desire to keep private.

Recent advances in deep learning optimization showed that just a subset of parameters are really necessary to successfully train a model. Potentially, such a discovery has broad impact from the theory to application; however, it is known that finding these trainable sub-network is a typically costly process. This inhibits practical applications: can the learned sub-graph structures in deep learning models be found at training timeƒ In this work we explore such a possibility, observing and motivating why common approaches typically fail in the extreme scenarios of interest, and proposing an approach which potentially enables training with reduced computational effort. The experiments on either challenging architectures and datasets suggest the algorithmic accessibility over such a computational gain, and in particular a trade-off between accuracy achieved and training complexity deployed emerges.

Recent advances in deep learning optimization showed that, with some a-posteriori information on fully-trained models, it is possible to match the same performance by simply training a subset of their parameters. Such a discovery has a broad impact from theory to applications, driving the research towards methods to identify the minimum subset of parameters to train without look-ahead information exploitation. However, the methods proposed do not match the state-of-the-art performance, and rely on unstructured sparsely connected models.In this work we shift our focus from the single parameters to the behavior of the whole neuron, exploiting the concept of neuronal equilibrium (NEq). When a neuron is in a configuration at equilibrium (meaning that it has learned a specific input-output relationship), we can halt its update; on the contrary, when a neuron is at non-equilibrium, we let its state evolve towards an equilibrium state, updating its parameters. The proposed approach has been tested on different state-of-the-art learning strategies and tasks, validating NEq and observing that the neuronal equilibrium depends on the specific learning setup.

Deep neural networks include millions of learnable parameters, making their deployment over resource-constrained devices problematic. Sensitivity-based regularization of neurons (SeReNe) is a method for learning sparse topologies with a structure, exploiting neural sensitivity as a regularizer. We define the sensitivity of a neuron as the variation of the network output with respect to the variation of the activity of the neuron. The lower the sensitivity of a neuron, the less the network output is perturbed if the neuron output changes. By including the neuron sensitivity in the cost function as a regularization term, we are able to prune neurons with low sensitivity. As entire neurons are pruned rather than single parameters, practical network footprint reduction becomes possible. Our experimental results on multiple network architectures and datasets yield competitive compression ratios with respect to state-of-the-art references.

Artificial neural networks perform state-of-the-art in an ever-growing number of tasks, and nowadays they are used to solve an incredibly large variety of tasks. There are problems, like the presence of biases in the training data, which question the generalization capability of these models. In this work we propose EnD, a regularization strategy whose aim is to prevent deep models from learning unwanted biases. In particular, we insert an "information bottleneck" at a certain point of the deep neural network, where we disentangle the information about the bias, still letting the useful information for the training task forward-propagating in the rest of the model. One big advantage of EnD is that we do not require additional training complexity (like decoders or extra layers in the model), since it is a regularizer directly applied on the trained model. Our experiments show that EnD effectively improves the generalization on unbiased test sets, and it can be effectively applied on real-case scenarios, like removing hidden biases in the COVID-19 detection from radiographic images.

Histopathological characterization of colorectal polyps allows to tailor patients’ management and follow up with the ultimate aim of avoiding or promptly detecting an invasive carcinoma. Colorectal polyps characterization relies on the histological analysis of tissue samples to determine the polyps malignancy and dysplasia grade. Deep neural networks achieve outstanding accuracy in medical patterns recognition, however they require large sets of annotated training images. We introduce UniToPatho, an annotated dataset of 9536 hematoxylin and eosin (H&E) stained patches extracted from 292 whole-slide images, meant for training deep neural networks for colorectal polyps classification and adenomas grading. We present our dataset and provide insights on how to tackle the problem of automatic colorectal polyps characterization by suggesting a multi-resolution deep learning approach.

This works explores the benefits of structured parameter pruning in the framework of the MPEG standardization efforts for neural network compression. First less relevant parameters are pruned from the network, then remaining parameters are quantized and finally quantized parameters are entropy coded. We consider an unstructured pruning strategy that maximizes the number of pruned parameters at the price of randomly sparse tensors and a structured strategy that prunes fewer parameters yet yields regularly sparse tensors. We show that structured pruning enables better end-to-end compression despite lower pruning ratio because it boosts the efficiency of the arithmetic coder. As a bonus, once decompressed, the network memory footprint is lower as well as its inference time.

Recently, a race towards the simplification of deep networks has begun, showing that it is effectively possible to reduce the size of these models with minimal or no performance loss. However, there is a general lack in understanding why these pruning strategies are effective. In this work, we are going to compare and analyze pruned solutions with two different pruning approaches, one-shot and gradual, showing the higher effectiveness of the latter. In particular, we find that gradual pruning allows access to narrow, well-generalizing minima, which are typically ignored when using one-shot approaches. In this work we also propose PSP-entropy, a measure to understand how a given neuron correlates to some specific learned classes. Interestingly, we observe that the features extracted by iteratively-pruned models are less correlated to specific classes, potentially making these models a better fit in transfer learning approaches.

In the last decade the use of artificial neural networks (ANNs) in many fields like image processing or speech recognition has become a common practice because of their effectiveness to solve complex tasks. However, in such a rush, very little attention has been paid to security aspects. In this work we explore the possibility to embed a watermark into the ANN parameters. We exploit model redundancy and adaptation capacity to lock a subset of its parameters to carry the watermark sequence. The watermark can be extracted in a simple way to claim copyright on models but can be very easily attacked with model fine-tuning. To tackle this culprit we devise a novel watermark aware training strategy. We aim at delving into the loss landscape to find an optimal configuration of the parameters such that we are robust to fine-tuning attacks towards the watermarked parameters. Our experimental results on classical ANN models trained on well-known MNIST and CIFAR-10 datasets show that the proposed approach makes the embedded watermark robust to fine-tuning and compression attacks.

Artificial neural networks perform state-of-the-art in an ever-growing number of tasks, nowadays they are used to solve an incredibly large variety of tasks. However, typical training strategies do not take into account lawful, ethical and discriminatory potential issues the trained ANN models could incur in. In this work we propose NDR, a non-discriminatory regularization strategy to prevent the ANN model to solve the target task using some discriminatory features like, for example, the ethnicity in an image classification task for human faces. In particular, a part of the ANN model is trained to hide the discriminatory information such that the rest of the network focuses in learning the given learning task. Our experiments show that NDR can be exploited to achieve non-discriminatory models with both minimal computational overhead and performance loss.

The ever-increasing number of parameters in deep neural networks poses challenges for memory-limited applications. Regularize-and-prune methods aim at meeting these challenges by sparsifying the network weights. In this context we quantify the output sensitivity to the parameters (i.e. their relevance to the network output) and introduce a regularization term that gradually lowers the absolute value of parameters with low sensitivity. Thus, a very large fraction of the parameters approach zero and are eventually set to zero by simple thresholding. Our method surpasses most of the recent techniques both in terms of sparsity and error rates. In some cases, the method reaches twice the sparsity obtained by other techniques at equal error rates.

Current nanoscale designs are highly interconnect dominated, taking about 70% of the chip area. Interconnects also consume significant dynamic power, and about 60% of signal delays. It is thus important to be able to synthesize much lower interconnect-complexity designs than are possible with current high-level synthesis (HLS) tools and algorithms. Towards that end, we have developed the new paradigms of: a) flexibly-structured buslets that connect a few “neighborhood” functional units (FUs) instead of dedicated interconnect between pairs of FUs, thereby sharing interconnects among a number of FU pairs that need to communicate; b) communication scheduling (followed by standard operation scheduling) in which communication between FUs are scheduled at appropriate times to minimize the number of buslets needed, subject to buslet cardinality constraints (for the purpose of upper bounding signal delay). Using a force-directed technique for communication and operation scheduling, and a chronological algorithm that simultaneously performs communication-to-buslet, FU-connections-to-buslets and operation-to-FU binding, we obtain significant wirelength (WL) reduction in the range of 35–71% in our designs compared to conventional FDS-based designs with dedicated-interconnects between communicating FU pairs.