Neural Networks (NNs) have been driving machine learning progress in recent years, but their larger models present challenges in resource-limited environments. Weight pruning reduces the computational demand, often with performance degradation and long training procedures. This work introduces Distilled Gradual Pruning with Pruned Fine-tuning (DG2PF), a comprehensive algorithm that iteratively prunes pre-trained neural networks using knowledge distillation. We employ a magnitude-based unstructured pruning function that selectively removes a specified proportion of unimportant weights from the network. This function also leads to an efficient compression of the model size while minimizing classification accuracy loss. Additionally, we introduce a simulated pruning strategy with the same effects of weight recovery but while maintaining stable convergence. Furthermore, we propose a multi-step self-knowledge distillation strategy to effectively transfer the knowledge of the full, unpruned network to the pruned counterpart. We validate the performance of our algorithm through extensive experimentation on diverse benchmark datasets, including CIFAR-10 and ImageNet, as well as a set of model architectures. The results highlight how our algorithm prunes and optimizes pretrained neural networks without substantially degrading their classification accuracy while delivering significantly faster and more compact models.
Interpretable computer vision models are able to explain their reasoning through comparing the distances between the image patch embeddings and prototypes within a latent space. However, many of these approaches introduce additional complexity, can require multiple training steps and often have a performance cost in comparison to black-box approaches. In this work, we introduce Component Features (ComFe), a novel interpretable-by-design image classification approach that is highly scalable and can obtain better accuracy and robustness in comparison to non-interpretable methods. Inspired by recent developments in computer vision foundation models, ComFe uses a transformer-decoder head and a hierarchical mixture-modelling approach with a foundation model backbone to obtain higher accuracy compared to previous interpretable models across a range of fine-grained vision benchmarks, without the need to individually tune hyper-parameters for each dataset. With only global image labels and no segmentation or part annotations, ComFe can identify consistent component features within an image and determine which of these features are informative in making a prediction.
In this paper, we study a defense against poisoned encoders in SSL called distillation, which is a defense used in supervised learning originally. Distillation aims to distill knowledge from a given model (a.k.a the teacher net) and transfer it to another (a.k.a the student net). Now, we use it to distill benign knowledge from poisoned pre-trained encoders and transfer it to a new encoder, resulting in a clean pre-trained encoder. In particular, we conduct an empirical study on the effectiveness and performance of distillation against poisoned encoders. Using two state-of-the-art backdoor attacks against pre-trained image encoders and four commonly used image classification datasets, our experimental results show that distillation can reduce attack success rate from 80.87% to 27.51% while suffering a 6.35% loss in accuracy. Moreover, we investigate the impact of three core components of distillation on performance: teacher net, student net, and distillation loss. By comparing 4 different teacher nets, 3 student nets, and 6 distillation losses, we find that fine-tuned teacher nets, warm-up-training-based student nets, and attention-based distillation loss perform best, respectively.
Binary neural networks utilize 1-bit quantized weights and activations to reduce both the model's storage demands and computational burden. However, advanced binary architectures still incorporate millions of inefficient and nonhardware-friendly full-precision multiplication operations. A&B BNN is proposed to directly remove part of the multiplication operations in a traditional BNN and replace the rest with an equal number of bit operations, introducing the mask layer and the quantized RPReLU structure based on the normalizer-free network architecture. The mask layer can be removed during inference by leveraging the intrinsic characteristics of BNN with straightforward mathematical transformations to avoid the associated multiplication operations. The quantized RPReLU structure enables more efficient bit operations by constraining its slope to be integer powers of 2. Experimental results achieved 92.30%, 69.35%, and 66.89% on the CIFAR-10, CIFAR-100, and ImageNet datasets, respectively, which are competitive with the state-of-the-art. Ablation studies have verified the efficacy of the quantized RPReLU structure, leading to a 1.14% enhancement on the ImageNet compared to using a fixed slope RLeakyReLU. The proposed add&bit-operation-only BNN offers an innovative approach for hardware-friendly network architecture.
Second-order optimization techniques have the potential to achieve faster convergence rates compared to first-order methods through the incorporation of second-order derivatives or statistics. However, their utilization in deep learning is limited due to their computational inefficiency. Various approaches have been proposed to address this issue, primarily centered on minimizing the size of the matrix to be inverted. Nevertheless, the necessity of performing the inverse operation iteratively persists. In this work, we present a fast natural gradient descent (FNGD) method that only requires inversion during the first epoch. Specifically, it is revealed that natural gradient descent (NGD) is essentially a weighted sum of per-sample gradients. Our novel approach further proposes to share these weighted coefficients across epochs without affecting empirical performance. Consequently, FNGD exhibits similarities to the average sum in first-order methods, leading to the computational complexity of FNGD being comparable to that of first-order methods. Extensive experiments on image classification and machine translation tasks demonstrate the efficiency of the proposed FNGD. For training ResNet-18 on CIFAR-100, FNGD can achieve a speedup of 2.07$\times$ compared with KFAC. For training Transformer on Multi30K, FNGD outperforms AdamW by 24 BLEU score while requiring almost the same training time.
Recurrent Spiking Neural Networks (RSNNs) have emerged as a computationally efficient and brain-inspired learning model. The design of sparse RSNNs with fewer neurons and synapses helps reduce the computational complexity of RSNNs. Traditionally, sparse SNNs are obtained by first training a dense and complex SNN for a target task, and, then, pruning neurons with low activity (activity-based pruning) while maintaining task performance. In contrast, this paper presents a task-agnostic methodology for designing sparse RSNNs by pruning a large randomly initialized model. We introduce a novel Lyapunov Noise Pruning (LNP) algorithm that uses graph sparsification methods and utilizes Lyapunov exponents to design a stable sparse RSNN from a randomly initialized RSNN. We show that the LNP can leverage diversity in neuronal timescales to design a sparse Heterogeneous RSNN (HRSNN). Further, we show that the same sparse HRSNN model can be trained for different tasks, such as image classification and temporal prediction. We experimentally show that, in spite of being task-agnostic, LNP increases computational efficiency (fewer neurons and synapses) and prediction performance of RSNNs compared to traditional activity-based pruning of trained dense models.
Since the era of deep learning, convolutional neural networks (CNNs) and vision transformers (ViTs) have been extensively studied and widely used in medical image classification tasks. Unfortunately, CNN's limitations in modeling long-range dependencies result in poor classification performances. In contrast, ViTs are hampered by the quadratic computational complexity of their self-attention mechanism, making them difficult to deploy in real-world settings with limited computational resources. Recent studies have shown that state space models (SSMs) represented by Mamba can effectively model long-range dependencies while maintaining linear computational complexity. Inspired by it, we proposed MedMamba, the first Vision Mamba for generalized medical image classification. Concretely, we introduced a novel hybrid basic block named SS-Conv-SSM, which purely integrates the convolutional layers for extracting local features with the abilities of SSM to capture long-range dependencies, aiming to model medical images from different image modalities efficiently. By employing the grouped convolution strategy and channel-shuffle operation, MedMamba successfully provides fewer model parameters and a lower computational burden for efficient applications without sacrificing accuracy. We thoroughly evaluated MedMamba using 16 datasets containing ten imaging modalities and 411,007 images. Experimental results show that MedMamba demonstrates competitive performance on most tasks compared with the state-of-the-art methods. This work aims to explore the potential of Vision Mamba and establish a new baseline for medical image classification, thereby providing valuable insights for developing more powerful Mamba-based artificial intelligence algorithms and applications in medicine. The source codes and all pre-trained weights of MedMamba are available at https://github.com/YubiaoYue/MedMamba.
The evolutionary paradigm has been successfully applied to neural network search(NAS) in recent years. Due to the vast search complexity of the global space, current research mainly seeks to repeatedly stack partial architectures to build the entire model or to seek the entire model based on manually designed benchmark modules. The above two methods are attempts to reduce the search difficulty by narrowing the search space. To efficiently search network architecture in the global space, this paper proposes another solution, namely a computationally efficient neural architecture evolutionary search framework based on network growth (G-EvoNAS). The complete network is obtained by gradually deepening different Blocks. The process begins from a shallow network, grows and evolves, and gradually deepens into a complete network, reducing the search complexity in the global space. Then, to improve the ranking accuracy of the network, we reduce the weight coupling of each network in the SuperNet by pruning the SuperNet according to elite groups at different growth stages. The G-EvoNAS is tested on three commonly used image classification datasets, CIFAR10, CIFAR100, and ImageNet, and compared with various state-of-the-art algorithms, including hand-designed networks and NAS networks. Experimental results demonstrate that G-EvoNAS can find a neural network architecture comparable to state-of-the-art designs in 0.2 GPU days.
Transformers have become the dominant architecture for sequence modeling tasks such as natural language processing or audio processing, and they are now even considered for tasks that are not naturally sequential such as image classification. Their ability to attend to and to process a set of tokens as context enables them to develop in-context few-shot learning abilities. However, their potential for online continual learning remains relatively unexplored. In online continual learning, a model must adapt to a non-stationary stream of data, minimizing the cumulative nextstep prediction loss. We focus on the supervised online continual learning setting, where we learn a predictor $x_t \rightarrow y_t$ for a sequence of examples $(x_t, y_t)$. Inspired by the in-context learning capabilities of transformers and their connection to meta-learning, we propose a method that leverages these strengths for online continual learning. Our approach explicitly conditions a transformer on recent observations, while at the same time online training it with stochastic gradient descent, following the procedure introduced with Transformer-XL. We incorporate replay to maintain the benefits of multi-epoch training while adhering to the sequential protocol. We hypothesize that this combination enables fast adaptation through in-context learning and sustained longterm improvement via parametric learning. Our method demonstrates significant improvements over previous state-of-the-art results on CLOC, a challenging large-scale real-world benchmark for image geo-localization.
This paper introduces a new stochastic optimization method based on the regularized Fisher information matrix (FIM), named SOFIM, which can efficiently utilize the FIM to approximate the Hessian matrix for finding Newton's gradient update in large-scale stochastic optimization of machine learning models. It can be viewed as a variant of natural gradient descent, where the challenge of storing and calculating the full FIM is addressed through making use of the regularized FIM and directly finding the gradient update direction via Sherman-Morrison matrix inversion. Additionally, like the popular Adam method, SOFIM uses the first moment of the gradient to address the issue of non-stationary objectives across mini-batches due to heterogeneous data. The utilization of the regularized FIM and Sherman-Morrison matrix inversion leads to the improved convergence rate with the same space and time complexities as stochastic gradient descent (SGD) with momentum. The extensive experiments on training deep learning models using several benchmark image classification datasets demonstrate that the proposed SOFIM outperforms SGD with momentum and several state-of-the-art Newton optimization methods in term of the convergence speed for achieving the pre-specified objectives of training and test losses as well as test accuracy.
Due to the effectiveness of second-order algorithms in solving classical optimization problems, designing second-order optimizers to train deep neural networks (DNNs) has attracted much research interest in recent years. However, because of the very high dimension of intermediate features in DNNs, it is difficult to directly compute and store the Hessian matrix for network optimization. Most of the previous second-order methods approximate the Hessian information imprecisely, resulting in unstable performance. In this work, we propose a compound optimizer, which is a combination of a second-order optimizer with a precise partial Hessian matrix for updating channel-wise parameters and the first-order stochastic gradient descent (SGD) optimizer for updating the other parameters. We show that the associated Hessian matrices of channel-wise parameters are diagonal and can be extracted directly and precisely from Hessian-free methods. The proposed method, namely SGD with Partial Hessian (SGD-PH), inherits the advantages of both first-order and second-order optimizers. Compared with first-order optimizers, it adopts a certain amount of information from the Hessian matrix to assist optimization, while compared with the existing second-order optimizers, it keeps the good generalization performance of first-order optimizers. Experiments on image classification tasks demonstrate the effectiveness of our proposed optimizer SGD-PH. The code is publicly available at \url{https://github.com/myingysun/SGDPH}.
Channel attention mechanisms endeavor to recalibrate channel weights to enhance representation abilities of networks. However, mainstream methods often rely solely on global average pooling as the feature squeezer, which significantly limits the overall potential of models. In this paper, we investigate the statistical moments of feature maps within a neural network. Our findings highlight the critical role of high-order moments in enhancing model capacity. Consequently, we introduce a flexible and comprehensive mechanism termed Extensive Moment Aggregation (EMA) to capture the global spatial context. Building upon this mechanism, we propose the Moment Channel Attention (MCA) framework, which efficiently incorporates multiple levels of moment-based information while minimizing additional computation costs through our Cross Moment Convolution (CMC) module. The CMC module via channel-wise convolution layer to capture multiple order moment information as well as cross channel features. The MCA block is designed to be lightweight and easily integrated into a variety of neural network architectures. Experimental results on classical image classification, object detection, and instance segmentation tasks demonstrate that our proposed method achieves state-of-the-art results, outperforming existing channel attention methods.
From content moderation to wildlife conservation, the number of applications that require models to recognize nuanced or subjective visual concepts is growing. Traditionally, developing classifiers for such concepts requires substantial manual effort measured in hours, days, or even months to identify and annotate data needed for training. Even with recently proposed Agile Modeling techniques, which enable rapid bootstrapping of image classifiers, users are still required to spend 30 minutes or more of monotonous, repetitive data labeling just to train a single classifier. Drawing on Fiske's Cognitive Miser theory, we propose a new framework that alleviates manual effort by replacing human labeling with natural language interactions, reducing the total effort required to define a concept by an order of magnitude: from labeling 2,000 images to only 100 plus some natural language interactions. Our framework leverages recent advances in foundation models, both large language models and vision-language models, to carve out the concept space through conversation and by automatically labeling training data points. Most importantly, our framework eliminates the need for crowd-sourced annotations. Moreover, our framework ultimately produces lightweight classification models that are deployable in cost-sensitive scenarios. Across 15 subjective concepts and across 2 public image classification datasets, our trained models outperform traditional Agile Modeling as well as state-of-the-art zero-shot classification models like ALIGN, CLIP, CuPL, and large visual question-answering models like PaLI-X.
Convolutional Neural Networks (CNNs) have demonstrated outstanding performance in computer vision tasks such as image classification, detection, segmentation, and medical image analysis. In general, an arbitrary number of epochs is used to train such neural networks. In a single epoch, the entire training data -- divided by batch size -- are fed to the network. In practice, validation error with training loss is used to estimate the neural network's generalization, which indicates the optimal learning capacity of the network. Current practice is to stop training when the training loss decreases and the gap between training and validation error increases (i.e., the generalization gap) to avoid overfitting. However, this is a trial-and-error-based approach which raises a critical question: Is it possible to estimate when neural networks stop learning based on training data? This research work introduces a hypothesis that analyzes the data variation across all the layers of a CNN variant to anticipate its near-optimal learning capacity. In the training phase, we use our hypothesis to anticipate the near-optimal learning capacity of a CNN variant without using any validation data. Our hypothesis can be deployed as a plug-and-play to any existing CNN variant without introducing additional trainable parameters to the network. We test our hypothesis on six different CNN variants and three different general image datasets (CIFAR10, CIFAR100, and SVHN). The result based on these CNN variants and datasets shows that our hypothesis saves 58.49\% of computational time (on average) in training. We further conduct our hypothesis on ten medical image datasets and compared with the MedMNIST-V2 benchmark. Based on our experimental result, we save $\approx$ 44.1\% of computational time without losing accuracy against the MedMNIST-V2 benchmark.
The attention mechanism is the primary component of the transformer architecture; it has led to significant advancements in deep learning spanning many domains and covering multiple tasks. In computer vision, the attention mechanism was first incorporated in the Vision Transformer ViT, and then its usage has expanded into many tasks in the vision domain, such as classification, segmentation, object detection, and image generation. While the attention mechanism is very expressive and capable, it comes with the disadvantage of being computationally expensive and requiring datasets of considerable size for effective optimization. To address these shortcomings, many designs have been proposed in the literature to reduce the computational burden and alleviate the data size requirements. Examples of such attempts in the vision domain are the MLP-Mixer, the Conv-Mixer, the Perciver-IO, and many more attempts with different sets of advantages and disadvantages. This paper introduces a new computational block as an alternative to the standard ViT block. The newly proposed block reduces the computational requirements by replacing the normal attention layers with a Network in Network structure, therefore enhancing the static approach of the MLP-Mixer with a dynamic learning of element-wise gating function generated by a token mixing process. Extensive experimentation shows that the proposed design provides better performance than the baseline architectures on multiple datasets applied in the image classification task of the vision domain.
Conventional deep learning (DL) model compression and scaling methods focus on altering the model's components, impacting the results across all samples uniformly. However, since samples vary in difficulty, a dynamic model that adapts computation based on sample complexity offers a novel perspective for compression and scaling. Despite this potential, existing dynamic models are typically monolithic and model-specific, limiting their generalizability as broad compression and scaling methods. Additionally, most deployed DL systems are fixed, unable to adjust their scale once deployed and, therefore, cannot adapt to the varying real-time demands. This paper introduces DyCE, a dynamically configurable system that can adjust the performance-complexity trade-off of a DL model at runtime without requiring re-initialization or redeployment on inference hardware. DyCE achieves this by adding small exit networks to intermediate layers of the original model, allowing computation to terminate early if acceptable results are obtained. DyCE also decouples the design of an efficient dynamic model, facilitating easy adaptation to new base models and potential general use in compression and scaling. We also propose methods for generating optimized configurations and determining the types and positions of exit networks to achieve desired performance and complexity trade-offs. By enabling simple configuration switching, DyCE provides fine-grained performance tuning in real-time. We demonstrate the effectiveness of DyCE through image classification tasks using deep convolutional neural networks (CNNs). DyCE significantly reduces computational complexity by 23.5% for ResNet152 and 25.9% for ConvNextv2-tiny on ImageNet, with accuracy reductions of less than 0.5%.
Adaptive methods are extremely popular in machine learning as they make learning rate tuning less expensive. This paper introduces a novel optimization algorithm named KATE, which presents a scale-invariant adaptation of the well-known AdaGrad algorithm. We prove the scale-invariance of KATE for the case of Generalized Linear Models. Moreover, for general smooth non-convex problems, we establish a convergence rate of $O \left(\frac{\log T}{\sqrt{T}} \right)$ for KATE, matching the best-known ones for AdaGrad and Adam. We also compare KATE to other state-of-the-art adaptive algorithms Adam and AdaGrad in numerical experiments with different problems, including complex machine learning tasks like image classification and text classification on real data. The results indicate that KATE consistently outperforms AdaGrad and matches/surpasses the performance of Adam in all considered scenarios.
Transformers have revolutionized computer vision and natural language processing, but their high computational complexity limits their application in high-resolution image processing and long-context analysis. This paper introduces Vision-RWKV (VRWKV), a model adapted from the RWKV model used in the NLP field with necessary modifications for vision tasks. Similar to the Vision Transformer (ViT), our model is designed to efficiently handle sparse inputs and demonstrate robust global processing capabilities, while also scaling up effectively, accommodating both large-scale parameters and extensive datasets. Its distinctive advantage lies in its reduced spatial aggregation complexity, which renders it exceptionally adept at processing high-resolution images seamlessly, eliminating the necessity for windowing operations. Our evaluations demonstrate that VRWKV surpasses ViT's performance in image classification and has significantly faster speeds and lower memory usage processing high-resolution inputs. In dense prediction tasks, it outperforms window-based models, maintaining comparable speeds. These results highlight VRWKV's potential as a more efficient alternative for visual perception tasks. Code is released at https://github.com/OpenGVLab/Vision-RWKV.
Computer vision models normally witness degraded performance when deployed in real-world scenarios, due to unexpected changes in inputs that were not accounted for during training. Data augmentation is commonly used to address this issue, as it aims to increase data variety and reduce the distribution gap between training and test data. However, common visual augmentations might not guarantee extensive robustness of computer vision models. In this paper, we propose Auxiliary Fourier-basis Augmentation (AFA), a complementary technique targeting augmentation in the frequency domain and filling the augmentation gap left by visual augmentations. We demonstrate the utility of augmentation via Fourier-basis additive noise in a straightforward and efficient adversarial setting. Our results show that AFA benefits the robustness of models against common corruptions, OOD generalization, and consistency of performance of models against increasing perturbations, with negligible deficit to the standard performance of models. It can be seamlessly integrated with other augmentation techniques to further boost performance. Code and models can be found at: https://github.com/nis-research/afa-augment
Benchmark datasets used for image classification tend to have very low levels of label noise. When Bayesian neural networks are trained on these datasets, they often underfit, misrepresenting the aleatoric uncertainty of the data. A common solution is to cool the posterior, which improves fit to the training data but is challenging to interpret from a Bayesian perspective. We explore whether posterior tempering can be replaced by a confidence-inducing prior distribution. First, we introduce a "DirClip" prior that is practical to sample and nearly matches the performance of a cold posterior. Second, we introduce a "confidence prior" that directly approximates a cold likelihood in the limit of decreasing temperature but cannot be easily sampled. Lastly, we provide several general insights into confidence-inducing priors, such as when they might diverge and how fine-tuning can mitigate numerical instability.