Benchmark contamination has become a significant concern in the LLM evaluation community. Previous Agents-as-an-Evaluator address this issue by involving agents in the generation of questions. Despite their success, the biases in Agents-as-an-Evaluator methods remain largely unexplored. In this paper, we present a theoretical formulation of evaluation bias, providing valuable insights into designing unbiased evaluation protocols. Furthermore, we identify two type of bias in Agents-as-an-Evaluator through carefully designed probing tasks on a minimal Agents-as-an-Evaluator setup. To address these issues, we propose the Unbiased Evaluator, an evaluation protocol that delivers a more comprehensive, unbiased, and interpretable assessment of LLMs. Extensive experiments reveal significant room for improvement in current LLMs. Additionally, we demonstrate that the Unbiased Evaluator not only offers strong evidence of benchmark contamination but also provides interpretable evaluation results.

We study a novel problem in this paper, that is, if a modern text-to-image diffusion model can tailor any image classifier across domains and categories. Existing domain adaption works exploit both source and target data for domain alignment so as to transfer the knowledge from the labeled source data to the unlabeled target data. However, as the development of text-to-image diffusion models, we wonder if the high-fidelity synthetic data can serve as a surrogate of the source data in real world. In this way, we do not need to collect and annotate the source data for each image classification task in a one-for-one manner. Instead, we utilize only one off-the-shelf text-to-image model to synthesize images with labels derived from text prompts, and then leverage them as a bridge to dig out the knowledge from the task-agnostic text-to-image generator to the task-oriented image classifier via domain adaptation. Such a one-for-all adaptation paradigm allows us to adapt anything in the world using only one text-to-image generator as well as any unlabeled target data. Extensive experiments validate the feasibility of this idea, which even surprisingly surpasses the state-of-the-art domain adaptation works using the source data collected and annotated in real world.

Diversifying training data techniques have achieved tremendous success in Domain Generalization (DG) tasks. The key to diversifying domain data is by increasing the types of domain styles. After investigating this issue from the perspective of the Fourier transform, the domain cue is found to be implicitly encoded in the amplitude component of Fourier features, which is more indicative of domain-specific information than statistics (means and standard deviations). However, Fourier-based methods tend to augment amplitude components via linear interpolation between two samples, which limits the diversity. To break this limitation, we aim to augment novel amplitude components from a perturbation perspective, which is termed Multivariate Fourier Distribution Perturbation. Specially, we design channel-wise and pixel-wise random perturbations for in-sample and cross-sample distribution to expand the distribution scope of probabilistic feature amplitude components.

Domain shift degrades the performance of object detection models in practical applications. To alleviate the influence of domain shift, plenty of previous work try to decouple and learn the domain-invariant (common) features from source domains via domain adversarial learning (DAL). However, inspired by causal mechanisms, we find that previous methods ignore the implicit insignificant non-causal factors hidden in the common features. This is mainly due to the single-view nature of DAL. In this work, we present an idea to remove non-causal factors from common features by multi-view adversarial training on source domains, because we observe that such insignificant non-causal factors may still be significant in other latent spaces (views) due to the multi-mode structure of data. To summarize, we propose a Multi-view Adversarial Discriminator (MAD) based domain generalization model, consisting of a Spurious Correlations Generator (SCG) that increases the diversity of source domain by random augmentation and a Multi-View Domain Classifier (MVDC) that maps features to multiple latent spaces, such that the non-causal factors are removed and the domain-invariant features are purified. Extensive experiments on six benchmarks show our MAD obtains state-of-the-art performance.

In the age of social media, posting attractive mugshots is commonplace, leading to an urgent need for automatic facial beautification techniques. To better meet the esthetic preferences of users, we devise a customized automatic face beautification task that can retouch the face adaptively to match the user-entered target score whilst preserving the ID information as much as possible. To accomplish this task, we propose a Human Esthetics Guided StyleGAN Inversion method to retouch each face in the embedding space using StyleGAN inversion. This process is guided by a pre-trained facial beauty prediction model that measures the difference between the target score and the predicted score of the retouched face. We conduct extensive experiments on various faces with different attributes, where the experimental results show that our method achieves the competitive performance, both in terms of visual effect and the proposed criterion.

Grounded language-image pre-trained models have shown strong zero-shot generalization to various downstream object detection tasks. Despite their promising performance, the models rely heavily on the laborious prompt engineering. Existing works typically address this problem by tuning text prompts using downstream training data in a few-shot or fully supervised manner. However, a rarely studied problem is to optimize text prompts without using any annotations. In this paper, we delve into this problem and propose an Unsupervised Prompt Tuning framework for text-driven object detection, which is composed of two novel mean teaching mechanisms. In conventional mean teaching, the quality of pseudo boxes is expected to optimize better as the training goes on, but there is still a risk of overfitting noisy pseudo boxes. To mitigate this problem, 1) we propose Nested Mean Teaching, which adopts nested-annotation to supervise teacher-student mutual learning in a bi-level optimization manner; 2) we propose Dual Complementary Teaching, which employs an offline pre-trained teacher and an online mean teacher via data-augmentation-based complementary labeling so as to ensure learning without accumulating confirmation bias. By integrating these two mechanisms, the proposed unsupervised prompt tuning framework achieves significant performance improvement on extensive object detection datasets.

Semi-supervised domain generalization (SSDG) focuses on training a model on a single labeled source domain and several unlabeled source domains simultaneously for the purpose of generalizing to out-of-distribution domain. To prevent model overfitting on the unlabeled source domains and enhance model generalization, we propose an effective two-stage framework that disentangle the SSDG task into a burn-in stage and a mutual-training stage. In the burn-in stage, we train a domain adaptation model to reduce domain gap and produce high-accuracy pseudo labels for the unlabeled data. In the second stage, we devise two peer networks equipped with style randomization modules to take advantage of both the style and class information from the pseudo-labeled data. The peer networks are mutually teaching each other, which allows them to avoid overfitting to noisy data and ultimately improves generalization ability. Extensive experiments show that our method achieves state-of-the-art performance on various benchmarks, demonstrating the effectiveness of our method on SSDG.

Agnostic domain shift is the main reason of model degradation on the unknown target domains, which brings an urgent need to develop Domain generalization (DG). Recent advances at DG use dynamic networks to achieve training-free adaptation on the unknown target domains, termed Dynamic Domain Generalization (DDG), which compensates for the lack of self-adaptability in static models with fixed weights. The parameters of dynamic networks can be decoupled into a static and a dynamic component, which are designed to learn domain-invariant and domain-specific features, respectively. Based on the existing arts, in this work, we try to push the limits of DDG by disentangling the static and dynamic components more thoroughly from an optimization perspective. Our main consideration is that we can enable the static component to learn domain-invariant features more comprehensively by augmenting the domain-specific information. As a result, the more comprehensive domain-invariant features learned by the static component can then enforce the dynamic component to focus more on learning adaptive domain-specific features. To this end, we propose a simple yet effective Parameter Exchange (PE) method to perturb the combination between the static and dynamic components. We optimize the model using the gradients from both the perturbed and non-perturbed feed-forward jointly to implicitly achieve the aforementioned disentanglement. In this way, the two components can be optimized in a mutually-beneficial manner, which can resist the agnostic domain shifts and improve the self-adaptability on the unknown target domain. Extensive experiments show that PE can be easily plugged into existing dynamic networks to improve their generalization ability without bells and whistles.

Predicting individual aesthetic preferences holds significant practical applications and academic implications for human society. However, existing studies mainly focus on learning and predicting the commonality of facial attractiveness, with little attention given to Personalized Facial Beauty Prediction (PFBP). PFBP aims to develop a machine that can adapt to individual aesthetic preferences with only a few images rated by each user. In this paper, we formulate this task from a meta-learning perspective that each user corresponds to a meta-task. To address such PFBP task, we draw inspiration from the human aesthetic mechanism that visual aesthetics in society follows a Gaussian distribution, which motivates us to disentangle user preferences into a commonality and an individuality part. To this end, we propose a novel MetaFBP framework, in which we devise a universal feature extractor to capture the aesthetic commonality and then optimize to adapt the aesthetic individuality by shifting the decision boundary of the predictor via a meta-learning mechanism. Unlike conventional meta-learning methods that may struggle with slow adaptation or overfitting to tiny support sets, we propose a novel approach that optimizes a high-order predictor for fast adaptation. In order to validate the performance of the proposed method, we build several PFBP benchmarks by using existing facial beauty prediction datasets rated by numerous users. Extensive experiments on these benchmarks demonstrate the effectiveness of the proposed MetaFBP method.

Data-Free Knowledge Distillation (DFKD) has shown great potential in creating a compact student model while alleviating the dependency on real training data by synthesizing surrogate data. However, prior arts are seldom discussed under distribution shifts, which may be vulnerable in real-world applications. Recent Vision-Language Foundation Models, e.g., CLIP, have demonstrated remarkable performance in zero-shot out-of-distribution generalization, yet consuming heavy computation resources. In this paper, we discuss the extension of DFKD to Vision-Language Foundation Models without access to the billion-level image-text datasets. The objective is to customize a student model for distribution-agnostic downstream tasks with given category concepts, inheriting the out-of-distribution generalization capability from the pre-trained foundation models. In order to avoid generalization degradation, the primary challenge of this task lies in synthesizing diverse surrogate images driven by text prompts. Since not only category concepts but also style information are encoded in text prompts, we propose three novel Prompt Diversification methods to encourage image synthesis with diverse styles, namely Mix-Prompt, Random-Prompt, and Contrastive-Prompt. Experiments on out-of-distribution generalization datasets demonstrate the effectiveness of the proposed methods, with Contrastive-Prompt performing the best.

Recent researches have made a great progress in domain adaptive object detectors. These detectors aim to learn explicit domain-invariant features by adversarially mitigating domain divergence and simultaneously optimizing source risks. However, an inherent problem is that they ignore the informative knowledge implied in domain-specific features, which is recognized as implicit domain-invariant feature. This is mainly caused by the multimode structure underlying target distribution, characterized by various scales and categories of objects in target images. To solve that, we propose the Implicit Domain-invariant Faster R-CNN (IDF) by using non-adversarial domain discriminator, dual attention mechanism and selective feature perception. This idea is implemented on the Faster R-CNN backbone, but with an improved architecture of two branches, i.e. domain-invariant branch and domain-specific branch. The former can clearly learn explicit domain adaptive features w.r.t. easy samples, while the latter aims to learn implicit domain-invariant features w.r.t. hard samples. Experiments on numerous benchmark datasets, including the Cityscapes, Foggy Cityscapes, KITTI and SIM10K, show the superiority of our IDF over other state-of-the-art domain adaptive object detectors. The demo code is released inhttps://github.com/sea123321/IDF.

Generalization of the person re-identification (ReID) model plays an important role in practical application, and we discuss a simple yet effective regularizer to improve it inspired by Adversarial Training (AT). AT has been indicated as an advanced regularizer due to its adversarial mechanism, ability to mine hard samples, and nature of data augmentation. However, serving as an augmentation-based regularizer, AT shows low diversity of the perturbation, excessive computational cost, and the optimization dilemma between adversarial robustness and accuracy for ReID task, and is thus suboptimal. To tackle these limitations and get a more effective regularizer for ReID, we rethink the nature of AT and unveil that the adversarial data augmentation is essentially reflected by gradients. Based on this, a novel implicit regularizer, named Stochastic Gradient Perturbation (SGP), is proposed, which naturally brings three merits: 1) Better diversity of the perturbation due to the proposed non-directional stochastic perturbations rather than directional adversarial perturbations. 2) Lower computational cost due to the proposed implicit gradient augmentation rather than explicitly additional data. 3) The optimization dilemma of the adversarial robustness and generalization is naturally overcome since SGP contains the adversarial gradient perturbation. Further, we put forward a perspective that the generalization and adversarial robustness may have an inter unity. Experiments on the baseline and SOTA models demonstrate powerful performances of the plugged-played SGP, and both generalization and adversarial robustness can be guaranteed.

We propose a Meta Learning on Randomized Transformations (MLRT) to learn domain invariant object detectors. Domain generalization is a problem about learning an invariant model from multiple source domains which can generalize well on unseen target domains. This problem is overlooked in object detection field, which is formally named as domain generalizable object detection (DGOD). Moreover, existing domain generalization methods have the problem of domain bias so that they can easily overfit to some specific domain (e.g., source domain). In order to alleviate the domain bias, in MLRT model, a novel randomized spectrum transformation (RST) module is proposed to increase the diversity of source domains. Specifically, RST randomizes the domain specific information of images in frequency-space, which can transform single or multiple source domains into various new domains. Besides, we observe a prior that the gradient imbalance degree among domains can also reflect the domain bias. Therefore, we further propose to alleviate the domain bias from the perspective of gradient balancing, and a novel gradient weighting (GW) module is proposed to balance the gradients over all domains via a hand-crafted weight. Finally we embed our RST and GW into a general meta learning framework and the proposed MLRT model is formalized for DGOD task. Extensive experiments are conducted on six benchmarks, and our method achieves the SOTA performance.

Semi-supervised object detection has made significant progress with the development of mean teacher driven self-training. Despite the promising results, the label mismatch problem is not yet fully explored in the previous works, leading to severe confirmation bias during self-training. In this paper, we delve into this problem and propose a simple yet effective LabelMatch framework from two different yet complementary perspectives, i.e., distribution-level and instance-level. For the former one, it is reasonable to approximate the class distribution of the unlabeled data from that of the labeled data according to Monte Carlo Sampling. Guided by this weakly supervision cue, we introduce a re-distribution mean teacher, which leverages adaptive label-distribution-aware confidence thresholds to generate unbiased pseudo labels to drive student learning. For the latter one, there exists an overlooked label assignment ambiguity problem across teacher-student models. To remedy this issue, we present a novel label assignment mechanism for self-training framework, namely proposal self-assignment, which injects the proposals from student into teacher and generates accurate pseudo labels to match each proposal in the student model accordingly. Experiments on both MS-COCO and PASCAL-VOC datasets demonstrate the considerable superiority of our proposed framework to other state-of-the-arts. Code will be available at https://github.com/HIK-LAB/SSOD.

Vanilla unsupervised domain adaptation methods tend to optimize the model with fixed neural architecture, which is not very practical in real-world scenarios since the target data is usually processed by different resource-limited devices. It is therefore of great necessity to facilitate architecture adaptation across various devices. In this paper, we introduce a simple framework, Slimmable Domain Adaptation, to improve cross-domain generalization with a weight-sharing model bank, from which models of different capacities can be sampled to accommodate different accuracy-efficiency trade-offs. The main challenge in this frame-work lies in simultaneously boosting the adaptation performance of numerous models in the model bank. To tackle this problem, we develop a Stochastic EnsEmble Distillation method to fully exploit the complementary knowledge in the model bank for inter-model interaction. Nevertheless, considering the optimization conflict between inter-model interaction and intra-model adaptation, we augment the existing bi-classifier domain confusion architecture into an Optimization-Separated Tri-Classifier counterpart. After optimizing the model bank, architecture adaptation is leveraged via our proposed Unsupervised Performance Evaluation Metric. Under various resource constraints, our framework surpasses other competing approaches by a very large margin on multiple benchmarks. It is also worth emphasizing that our framework can preserve the performance improvement against the source-only model even when the computing complexity is reduced to 1/64. Code will be available at https://github.com/HIK-LAB/SlimDA.

Inspired by the remarkable zero-shot generalization capacity of vision-language pre-trained model, we seek to leverage the supervision from CLIP model to alleviate the burden of data labeling. However, such supervision inevitably contains the label noise, which significantly degrades the discriminative power of the classification model. In this work, we propose Transductive CLIP, a novel framework for learning a classification network with noisy labels from scratch. Firstly, a class-conditional contrastive learning mechanism is proposed to mitigate the reliance on pseudo labels and boost the tolerance to noisy labels. Secondly, ensemble labels is adopted as a pseudo label updating strategy to stabilize the training of deep neural networks with noisy labels. This framework can reduce the impact of noisy labels from CLIP model effectively by combining both techniques. Experiments on multiple benchmark datasets demonstrate the substantial improvements over other state-of-the-art methods.

Recent researches show that data auto-augmentation strategies can enhance the performance of object detection models. However, the existing works mainly focus on in-domain generalization. There is still a blank in out-of-domain generalization. In this paper, for the first time, we propose an auto-augmentation problem under unsupervised domain adaptation scenarios. To solve this problem, we propose a simple yet effective target-aware auto-augmentation technique to search for an optimal data augmentation strategy on labeled source data, so as to boost the detection ability on the given unlabeled target data. Our method can be easily plugged into the existing domain adaptation methods. Extensive experiments have been carried out to verify the effectiveness.

Vanilla unsupervised domain adaptive (UDA) object detection typically requires the labeled source data for joint-training with the unlabeled target data, which is usually unavailable in real-world scenarios due to data privacy, leading to source data-free UDA object detection. Herein, we first analyze the phenomenon of cross-domain detection degradation varying from easy to hard samples (e.g. the objects with different scales or occlusion degrees), termed as domain generalization differentiation. In detail, the ability to detect easy samples is well transferred while the one to detect hard samples is dramatically degraded. To this end, we then revisit the existing self-training method, which is of great challenge to deal with the abundant false negatives (hard samples). Assumed that true positives (easy samples) labeled by the source model can be exploited as supervision cues. UDA is finally modeled into an unsupervised false negatives mining problem. Thus, we propose a Simulation-and-Mining (S&M) framework, which simulates false negatives by augmenting true positives and mines back false negatives alternatively and iteratively. Experimental results show the effectiveness.

Self-training for unsupervised domain adaptive object detection is a challenging task, of which the performance depends heavily on the quality of pseudo boxes. Despite the promising results, prior works have largely overlooked the uncertainty of pseudo boxes during self-training. In this paper, we present a simple yet effective framework, termed as Probabilistic Teacher (PT), which aims to capture the uncertainty of unlabeled target data from a gradually evolving teacher and guides the learning of a student in a mutually beneficial manner. Specifically, we propose to leverage the uncertainty-guided consistency training to promote classification adaptation and localization adaptation, rather than filtering pseudo boxes via an elaborate confidence threshold. In addition, we conduct anchor adaptation in parallel with localization adaptation, since anchor can be regarded as a learnable parameter. Together with this framework, we also present a novel Entropy Focal Loss (EFL) to further facilitate the uncertainty-guided self-training. Equipped with EFL, PT outperforms all previous baselines by a large margin and achieve new state-of-the-arts.

Domain generalization (DG) is a fundamental yet very challenging research topic in machine learning. The existing arts mainly focus on learning domain-invariant features with limited source domains in a static model. Unfortunately, there is a lack of training-free mechanism to adjust the model when generalized to the agnostic target domains. To tackle this problem, we develop a brand-new DG variant, namely Dynamic Domain Generalization (DDG), in which the model learns to twist the network parameters to adapt to the data from different domains. Specifically, we leverage a meta-adjuster to twist the network parameters based on the static model with respect to different data from different domains. In this way, the static model is optimized to learn domain-shared features, while the meta-adjuster is designed to learn domain-specific features. To enable this process, DomainMix is exploited to simulate data from diverse domains during teaching the meta-adjuster to adapt to the agnostic target domains. This learning mechanism urges the model to generalize to different agnostic target domains via adjusting the model without training. Extensive experiments demonstrate the effectiveness of our proposed method. Code is available: https://github.com/MetaVisionLab/DDG

Domain adaptation is an important property in robot vision, which enables the neural networks pre-trained on source domains to adapt target domains automatically without any annotation efforts. During this process, source data is not always accessible due to the constraints of expensive storage overhead and data privacy protection. Therefore, the source domain pre-trained model is expected to optimize with only unlabeled target data, termed as source-free unsupervised domain adaptation. In this paper, we view this problem as a special case of noisy label learning, since the given pre-trained model can generate noisy labels for unlabeled target data via network inference. The potential semantic cues for unsupervised domain adaptation exactly lie on these noisy labels. Inspired by this problem modeling, we propose a simple yet effective Self-Supervised Noisy Label Learning method, which injects self-supervised learning to impose the intrinsic data structure and facilitate label-denoising. Extensive experiments have been conducted on diverse benchmarks to validate the effectiveness. Our method achieves state-of-the-art performance.

Universal domain adaptive object detection (UniDAOD) is more challenging than domain adaptive object detection (DAOD) since the label space of the source domain may not be the same as that of the target and the scale of objects in the universal scenarios can vary dramatically (i.e, category shift and scale shift). To this end, we propose US-DAF, namely Universal Scale-Aware Domain Adaptive Faster RCNN with Multi-Label Learning, to reduce the negative transfer effect during training while maximizing transferability as well as discriminability in both domains under a variety of scales. Specifically, our method is implemented by two modules: 1) We facilitate the feature alignment of common classes and suppress the interference of private classes by designing a Filter Mechanism module to overcome the negative transfer caused by category shift. 2) We fill the blank of scale-aware adaptation in object detection by introducing a new Multi-Label Scale-Aware Adapter to perform individual alignment between corresponding scale for two domains. Experiments show that US-DAF achieves state-of-the-art results on three scenarios (\emphi.e, Open-Set, Partial-Set, and Closed-Set) and yields 7.1% and 5.9% relative improvement on benchmark datasets Clipart1k and Watercolor in particular.

Unsupervised domain adaptation (UDA) assumes that source and target domain data are freely available and usually trained together to reduce the domain gap. However, considering the data privacy and the inefficiency of data transmission, it is impractical in real scenarios. Hence, it draws our eyes to optimize the network in the target domain without accessing labeled source data. To explore this direction in object detection, for the first time, we propose a source data-free domain adaptive object detection (SFOD) framework via modeling it into a problem of learning with noisy labels. Generally, a straightforward method is to leverage the pre-trained network from the source domain to generate the pseudo labels for target domain optimization. However, it is difficult to evaluate the quality of pseudo labels since no labels are available in target domain. In this paper, self-entropy descent (SED) is a metric proposed to search an appropriate confidence threshold for reliable pseudo label generation without using any handcrafted labels. Nonetheless, completely clean labels are still unattainable. After a thorough experimental analysis, false negatives are found to dominate in the generated noisy labels. Undoubtedly, false negatives mining is helpful for performance improvement, and we ease it to false negatives simulation through data augmentation like Mosaic. Extensive experiments conducted in four representative adaptation tasks have demonstrated that the proposed framework can easily achieve state-of-the-art performance. From another view, it also reminds the UDA community that the labeled source data are not fully exploited in the existing methods.

Object detection is one of the hottest research directions in computer vision, has already made impressive progress in academia, and has many valuable applications in the industry. However, the mainstream detection methods still have two shortcomings: (1) even a model that is well trained using large amounts of data still cannot generally be used across different kinds of scenes; (2) once a model is deployed, it cannot autonomously evolve along with the accumulated unlabeled scene data. To address these problems, and inspired by visual knowledge theory, we propose a novel scene-adaptive evolution unsupervised video object detection algorithm that can decrease the impact of scene changes through the concept of object groups. We first extract a large number of object proposals from unlabeled data through a pre-trained detection model. Second, we build the visual knowledge dictionary of object concepts by clustering the proposals, in which each cluster center represents an object prototype. Third, we look into the relations between different clusters and the object information of different groups, and propose a graph-based group information propagation strategy to determine the category of an object concept, which can effectively distinguish positive and negative proposals. With these pseudo labels, we can easily fine-tune the pre-trained model. The effectiveness of the proposed method is verified by performing different experiments, and the significant improvements are achieved.

Layer assignment is seldom picked out as an independent research topic in neural architecture search. In this paper, for the first time, we systematically investigate the impact of different layer assignments to the network performance by building an architecture dataset of layer assignment on CIFAR-100. Through analyzing this dataset, we discover a neural inheritance relation among the networks with different layer assignments, that is, the optimal layer assignments for deeper networks always inherit from those for shallow networks. Inspired by this neural inheritance relation, we propose an efficient one-shot layer assignment search approach via inherited sampling. Specifically, the optimal layer assignment searched in the shallow network can be provided as a strong sampling priori to train and search the deeper ones in supernet, which extremely reduces the network search space. Comprehensive experiments carried out on CIFAR-100 illustrate the efficiency of our proposed method. Our search results are strongly consistent with the optimal ones directly selected from the architecture dataset. To further confirm the generalization of our proposed method, we also conduct experiments on Tiny-ImageNet and ImageNet. Our searched results are remarkably superior to the handcrafted ones under the unchanged computational budgets. The neural inheritance relation discovered in this paper can provide insights to the universal neural architecture search.

Neuron pruning is an efficient method to compress the network into a slimmer one for reducing the computational cost and storage overhead. Most of state-of-the-art results are obtained in a layer-by-layer optimization mode. It discards the unimportant input neurons and uses the survived ones to reconstruct the output neurons approaching to the original ones in a layer-by-layer manner. However, an unnoticed problem arises that the information loss is accumulated as layer increases since the survived neurons still do not encode the entire information as before. A better alternative is to propagate the entire useful information to reconstruct the pruned layer instead of directly discarding the less important neurons. To this end, we propose a novel Layer DecompositionRecomposition Framework (LDRF) for neuron pruning, by which each layer’s output information is recovered in an embedding space and then propagated to reconstruct the following pruned layers with useful information preserved. We mainly conduct our experiments on ILSVRC-12 benchmark with VGG-16 and ResNet-50. What should be emphasized is that our results before end-to-end fine-tuning are significantly superior owing to the information-preserving property of our proposed framework. With end-to-end fine-tuning, we achieve state-of-the-art results of 5.13× and 3× speed-up with only 0.5% and 0.65% top-5 accuracy drop respectively, which outperform the existing neuron pruning methods.

Shift operation is an efficient alternative over depthwise separable convolution. However, it is still bottlenecked by its implementation manner, namely memory movement. To put this direction forward, a new and novel basic component named Sparse Shift Layer (SSL) is introduced in this paper to construct efficient convolutional neural networks. In this family of architectures, the basic block is only composed by 1x1 convolutional layers with only a few shift operations applied to the intermediate feature maps. To make this idea feasible, we introduce shift operation penalty during optimization and further propose a quantization-aware shift learning method to impose the learned displacement more friendly for inference. Extensive ablation studies indicate that only a few shift operations are sufficient to provide spatial information communication. Furthermore, to maximize the role of SSL, we redesign an improved network architecture to Fully Exploit the limited capacity of neural Network (FE-Net). Equipped with SSL, this network can achieve 75.0% top-1 accuracy on ImageNet with only 563M M-Adds. It surpasses other counterparts constructed by depthwise separable convolution and the networks searched by NAS in terms of accuracy and practical speed.

Facial beauty prediction (FBP) aims to develop a machine that automatically makes facial attractiveness assessment. To a large extent, the perception of facial beauty for a human is involved with the attributes of facial appearance, which provides some significant visual cues for FBP. Deep convolution neural networks (CNNs) have shown its power for FBP, but convolution filters with fixed parameters cannot take full advantage of the facial attributes for FBP. To address this problem, we propose an Attribute-aware Convolutional Neural Network (AaNet) that modulates the filters of the main network, adaptively, using parameter generators that take beauty-related attributes as extra inputs. The parameter generators update the filters in the main network in two different manners: filter tuning or filter rebirth. However, AaNet takes attributes information as prior knowledge, that is ill-suited to those datasets merely with task-oriented labels. Therefore, imitating the design of AaNet, we further propose a Pseudo Attribute-aware Convolutional Neural Network (P-AaNet) that modulates filters conditioned on global context embeddings (pseudo attributes) of input faces learnt by a lightweight pseudo attribute distiller. Extensive ablation studies show that the AaNet and P-AaNet improve the performance of FBP when compared to conventional convolution and attention scheme, which validates the effectiveness of our method.