How humans and machines make sense of current inputs for relation reasoning and question-answering while putting the perceived information into context of our past memories, has been a challenging conundrum in cognitive science and artificial intelligence. Inspired by human brain's memory system and cognitive architectures, we propose a PMI framework that consists of perception, memory and inference components. Notably, the memory module comprises working and long-term memory, with the latter endowed with a higher-order structure to retain extensive and complex relational knowledge and experience. Through a differentiable competitive write access, current perceptions update working memory, which is later merged with long-term memory via outer product associations, reducing information conflicts and averting memory overflow. In the inference module, relevant information is retrieved from two separate memory origins and associatively integrated to attain a more comprehensive and precise interpretation of current perceptions. We exploratively apply our PMI to improve prevailing Transformers and CNN models on question-answering tasks like bAbI-20k and Sort-of-CLEVR datasets, as well as detecting equilateral triangles, language modeling and image classification tasks, and in each case, our PMI enhancements consistently outshine their original counterparts significantly. Visualization analyses reveal that relational memory consolidation, along with the interaction and integration of information from diverse memory sources, substantially contributes to the model effectiveness on inference tasks.
Deep Equilibrium Models (DEQs) and Neural Ordinary Differential Equations (Neural ODEs) are two branches of implicit models that have achieved remarkable success owing to their superior performance and low memory consumption. While both are implicit models, DEQs and Neural ODEs are derived from different mathematical formulations. Inspired by homotopy continuation, we establish a connection between these two models and illustrate that they are actually two sides of the same coin. Homotopy continuation is a classical method of solving nonlinear equations based on a corresponding ODE. Given this connection, we proposed a new implicit model called HomoODE that inherits the property of high accuracy from DEQs and the property of stability from Neural ODEs. Unlike DEQs, which explicitly solve an equilibrium-point-finding problem via Newton's methods in the forward pass, HomoODE solves the equilibrium-point-finding problem implicitly using a modified Neural ODE via homotopy continuation. Further, we developed an acceleration method for HomoODE with a shared learnable initial point. It is worth noting that our model also provides a better understanding of why Augmented Neural ODEs work as long as the augmented part is regarded as the equilibrium point to find. Comprehensive experiments with several image classification tasks demonstrate that HomoODE surpasses existing implicit models in terms of both accuracy and memory consumption.
As a heuristic for improving test accuracy in classification, the "flooding" method proposed by Ishida et al. (2020) sets a threshold for the average surrogate loss at training time; above the threshold, gradient descent is run as usual, but below the threshold, a switch to gradient ascent is made. While setting the threshold is non-trivial and is usually done with validation data, this simple technique has proved remarkably effective in terms of accuracy. On the other hand, what if we are also interested in other metrics such as model complexity or average surrogate loss at test time? As an attempt to achieve better overall performance with less fine-tuning, we propose a softened, pointwise mechanism called SoftAD (soft ascent-descent) that downweights points on the borderline, limits the effects of outliers, and retains the ascent-descent effect of flooding, with no additional computational overhead. We contrast formal stationarity guarantees with those for flooding, and empirically demonstrate how SoftAD can realize classification accuracy competitive with flooding (and the more expensive alternative SAM) while enjoying a much smaller loss generalization gap and model norm.
We propose Re-parameterized Refocusing Convolution (RefConv) as a replacement for regular convolutional layers, which is a plug-and-play module to improve the performance without any inference costs. Specifically, given a pre-trained model, RefConv applies a trainable Refocusing Transformation to the basis kernels inherited from the pre-trained model to establish connections among the parameters. For example, a depth-wise RefConv can relate the parameters of a specific channel of convolution kernel to the parameters of the other kernel, i.e., make them refocus on the other parts of the model they have never attended to, rather than focus on the input features only. From another perspective, RefConv augments the priors of existing model structures by utilizing the representations encoded in the pre-trained parameters as the priors and refocusing on them to learn novel representations, thus further enhancing the representational capacity of the pre-trained model. Experimental results validated that RefConv can improve multiple CNN-based models by a clear margin on image classification (up to 1.47% higher top-1 accuracy on ImageNet), object detection and semantic segmentation without introducing any extra inference costs or altering the original model structure. Further studies demonstrated that RefConv can reduce the redundancy of channels and smooth the loss landscape, which explains its effectiveness.
Continual Learning (CL) aims to incrementally update a trained model on new tasks without forgetting the acquired knowledge of old ones. Existing CL methods usually reduce forgetting with task priors, \ie using task identity or a subset of previously seen samples for model training. However, these methods would be infeasible when such priors are unknown in real-world applications. To address this fundamental but seldom-studied problem, we propose a Prior-Free Continual Learning (PFCL) method, which learns new tasks without knowing the task identity or any previous data. First, based on a fixed single-head architecture, we eliminate the need for task identity to select the task-specific output head. Second, we employ a regularization-based strategy for consistent predictions between the new and old models, avoiding revisiting previous samples. However, using this strategy alone often performs poorly in class-incremental scenarios, particularly for a long sequence of tasks. By analyzing the effectiveness and limitations of conventional regularization-based methods, we propose enhancing model consistency with an auxiliary unlabeled dataset additionally. Moreover, since some auxiliary data may degrade the performance, we further develop a reliable sample selection strategy to obtain consistent performance improvement. Extensive experiments on multiple image classification benchmark datasets show that our PFCL method significantly mitigates forgetting in all three learning scenarios. Furthermore, when compared to the most recent rehearsal-based methods that replay a limited number of previous samples, PFCL achieves competitive accuracy. Our code is available at: https://github.com/visiontao/pfcl
Synthetic training data has gained prominence in numerous learning tasks and scenarios, offering advantages such as dataset augmentation, generalization evaluation, and privacy preservation. Despite these benefits, the efficiency of synthetic data generated by current methodologies remains inferior when training advanced deep models exclusively, limiting its practical utility. To address this challenge, we analyze the principles underlying training data synthesis for supervised learning and elucidate a principled theoretical framework from the distribution-matching perspective that explicates the mechanisms governing synthesis efficacy. Through extensive experiments, we demonstrate the effectiveness of our synthetic data across diverse image classification tasks, both as a replacement for and augmentation to real datasets, while also benefits such as out-of-distribution generalization, privacy preservation, and scalability. Specifically, we achieve 70.9% top1 classification accuracy on ImageNet1K when training solely with synthetic data equivalent to 1 X the original real data size, which increases to 76.0% when scaling up to 10 X synthetic data.
Gradient-based meta-learning techniques aim to distill useful prior knowledge from a set of training tasks such that new tasks can be learned more efficiently with gradient descent. While these methods have achieved successes in various scenarios, they commonly adapt all parameters of trainable layers when learning new tasks. This neglects potentially more efficient learning strategies for a given task distribution and may be susceptible to overfitting, especially in few-shot learning where tasks must be learned from a limited number of examples. To address these issues, we propose Subspace Adaptation Prior (SAP), a novel gradient-based meta-learning algorithm that jointly learns good initialization parameters (prior knowledge) and layer-wise parameter subspaces in the form of operation subsets that should be adaptable. In this way, SAP can learn which operation subsets to adjust with gradient descent based on the underlying task distribution, simultaneously decreasing the risk of overfitting when learning new tasks. We demonstrate that this ability is helpful as SAP yields superior or competitive performance in few-shot image classification settings (gains between 0.1% and 3.9% in accuracy). Analysis of the learned subspaces demonstrates that low-dimensional operations often yield high activation strengths, indicating that they may be important for achieving good few-shot learning performance. For reproducibility purposes, we publish all our research code publicly.
Vision-Language Models (VLMs) such as CLIP are trained on large amounts of image-text pairs, resulting in remarkable generalization across several data distributions. However, in several cases, their expensive training and data collection/curation costs do not justify the end application. This motivates a vendor-client paradigm, where a vendor trains a large-scale VLM and grants only input-output access to clients on a pay-per-query basis in a black-box setting. The client aims to minimize inference cost by distilling the VLM to a student model using the limited available task-specific data, and further deploying this student model in the downstream application. While naive distillation largely improves the In-Domain (ID) accuracy of the student, it fails to transfer the superior out-of-distribution (OOD) generalization of the VLM teacher using the limited available labeled images. To mitigate this, we propose Vision-Language to Vision - Align, Distill, Predict (VL2V-ADiP), which first aligns the vision and language modalities of the teacher model with the vision modality of a pre-trained student model, and further distills the aligned VLM representations to the student. This maximally retains the pre-trained features of the student, while also incorporating the rich representations of the VLM image encoder and the superior generalization of the text embeddings. The proposed approach achieves state-of-the-art results on the standard Domain Generalization benchmarks in a black-box teacher setting as well as a white-box setting where the weights of the VLM are accessible.
Data augmentation is a dominant method for reducing model overfitting and improving generalization. Most existing data augmentation methods tend to find a compromise in augmenting the data, \textit{i.e.}, increasing the amplitude of augmentation carefully to avoid degrading some data too much and doing harm to the model performance. We delve into the relationship between data augmentation and model performance, revealing that the performance drop with heavy augmentation comes from the presence of out-of-distribution (OOD) data. Nonetheless, as the same data transformation has different effects for different training samples, even for heavy augmentation, there remains part of in-distribution data which is beneficial to model training. Based on the observation, we propose a novel data augmentation method, named \textbf{DualAug}, to keep the augmentation in distribution as much as possible at a reasonable time and computational cost. We design a data mixing strategy to fuse augmented data from both the basic- and the heavy-augmentation branches. Extensive experiments on supervised image classification benchmarks show that DualAug improve various automated data augmentation method. Moreover, the experiments on semi-supervised learning and contrastive self-supervised learning demonstrate that our DualAug can also improve related method. Code is available at \href{https://github.com/shuguang99/DualAug}{https://github.com/shuguang99/DualAug}.
Visual prompting (VP) is an emerging parameter-efficient fine-tuning approach to adapting pre-trained vision models to solve various downstream image-classification tasks. However, there has hitherto been little systematic study of the design space of VP and no clear benchmark for evaluating its performance. To bridge this gap, we propose AutoVP, an end-to-end expandable framework for automating VP design choices, along with 12 downstream image-classification tasks that can serve as a holistic VP-performance benchmark. Our design space covers 1) the joint optimization of the prompts; 2) the selection of pre-trained models, including image classifiers and text-image encoders; and 3) model output mapping strategies, including nonparametric and trainable label mapping. Our extensive experimental results show that AutoVP outperforms the best-known current VP methods by a substantial margin, having up to 6.7% improvement in accuracy; and attains a maximum performance increase of 27.5% compared to linear-probing (LP) baseline. AutoVP thus makes a two-fold contribution: serving both as an efficient tool for hyperparameter tuning on VP design choices, and as a comprehensive benchmark that can reasonably be expected to accelerate VP's development. The source code is available at https://github.com/IBM/AutoVP.
Despite deep learning (DL) has achieved remarkable progress in various domains, the DL models are still prone to making mistakes. This issue necessitates effective debugging tools for DL practitioners to interpret the decision-making process within the networks. However, existing debugging methods often demand extra data or adjustments to the decision process, limiting their applicability. To tackle this problem, we present NeuroInspect, an interpretable neuron-based debugging framework with three key stages: counterfactual explanations, feature visualizations, and false correlation mitigation. Our debugging framework first pinpoints neurons responsible for mistakes in the network and then visualizes features embedded in the neurons to be human-interpretable. To provide these explanations, we introduce CLIP-Illusion, a novel feature visualization method that generates images representing features conditioned on classes to examine the connection between neurons and the decision layer. We alleviate convoluted explanations of the conventional visualization approach by employing class information, thereby isolating mixed properties. This process offers more human-interpretable explanations for model errors without altering the trained network or requiring additional data. Furthermore, our framework mitigates false correlations learned from a dataset under a stochastic perspective, modifying decisions for the neurons considered as the main causes. We validate the effectiveness of our framework by addressing false correlations and improving inferences for classes with the worst performance in real-world settings. Moreover, we demonstrate that NeuroInspect helps debug the mistakes of DL models through evaluation for human understanding. The code is openly available at https://github.com/yeongjoonJu/NeuroInspect.
Medical Vision-Language Pre-training (VLP) learns representations jointly from medical images and paired radiology reports. It typically requires large-scale paired image-text datasets to achieve effective pre-training for both the image encoder and text encoder. The advent of text-guided generative models raises a compelling question: Can VLP be implemented solely with synthetic images generated from genuine radiology reports, thereby mitigating the need for extensively pairing and curating image-text datasets? In this work, we scrutinize this very question by examining the feasibility and effectiveness of employing synthetic images for medical VLP. We replace real medical images with their synthetic equivalents, generated from authentic medical reports. Utilizing three state-of-the-art VLP algorithms, we exclusively train on these synthetic samples. Our empirical evaluation across three subsequent tasks, namely image classification, semantic segmentation and object detection, reveals that the performance achieved through synthetic data is on par with or even exceeds that obtained with real images. As a pioneering contribution to this domain, we introduce a large-scale synthetic medical image dataset, paired with anonymized real radiology reports. This alleviates the need of sharing medical images, which are not easy to curate and share in practice. The code and the dataset can be found in \href{https://github.com/cheliu-computation/MedSyn-RepLearn/tree/main}{https://github.com/cheliu-computation/MedSyn-RepLearn/tree/main}.
Discriminative learning effectively predicts true object class for image classification. However, it often results in false positives for outliers, posing critical concerns in applications like autonomous driving and video surveillance systems. Previous attempts to address this challenge involved training image classifiers through contrastive learning using actual outlier data or synthesizing outliers for self-supervised learning. Furthermore, unsupervised generative modeling of inliers in pixel space has shown limited success for outlier detection. In this work, we introduce a quantile-based maximum likelihood objective for learning the inlier distribution to improve the outlier separation during inference. Our approach fits a normalizing flow to pre-trained discriminative features and detects the outliers according to the evaluated log-likelihood. The experimental evaluation demonstrates the effectiveness of our method as it surpasses the performance of the state-of-the-art unsupervised methods for outlier detection. The results are also competitive compared with a recent self-supervised approach for outlier detection. Our work allows to reduce dependency on well-sampled negative training data, which is especially important for domains like medical diagnostics or remote sensing.
Owing to advancements in deep learning technology, Vision Transformers (ViTs) have demonstrated impressive performance in various computer vision tasks. Nonetheless, ViTs still face some challenges, such as high computational complexity and the absence of desirable inductive biases. To alleviate these issues, {the potential advantages of combining eagle vision with ViTs are explored. We summarize a Bi-Fovea Visual Interaction (BFVI) structure inspired by the unique physiological and visual characteristics of eagle eyes. A novel Bi-Fovea Self-Attention (BFSA) mechanism and Bi-Fovea Feedforward Network (BFFN) are proposed based on this structural design approach, which can be used to mimic the hierarchical and parallel information processing scheme of the biological visual cortex, enabling networks to learn feature representations of targets in a coarse-to-fine manner. Furthermore, a Bionic Eagle Vision (BEV) block is designed as the basic building unit based on the BFSA mechanism and BFFN. By stacking BEV blocks, a unified and efficient family of pyramid backbone networks called Eagle Vision Transformers (EViTs) is developed. Experimental results show that EViTs exhibit highly competitive performance in various computer vision tasks, such as image classification, object detection and semantic segmentation. Compared with other approaches, EViTs have significant advantages, especially in terms of performance and computational efficiency. Code is available at https://github.com/nkusyl/EViT
Spiking Neural Networks (SNNs) have emerged as a promising alternative to conventional Artificial Neural Networks (ANNs), demonstrating comparable performance in both visual and linguistic tasks while offering the advantage of improved energy efficiency. Despite these advancements, the integration of linguistic and visual features into a unified representation through spike trains poses a significant challenge, and the application of SNNs to multimodal scenarios remains largely unexplored. This paper presents SpikeCLIP, a novel framework designed to bridge the modality gap in spike-based computation. Our approach employs a two-step recipe: an ``alignment pre-training'' to align features across modalities, followed by a ``dual-loss fine-tuning'' to refine the model's performance. Extensive experiments reveal that SNNs achieve results on par with ANNs while substantially reducing energy consumption across various datasets commonly used for multimodal model evaluation. Furthermore, SpikeCLIP maintains robust image classification capabilities, even when dealing with classes that fall outside predefined categories. This study marks a significant advancement in the development of energy-efficient and biologically plausible multimodal learning systems. Our code is available at https://github.com/Lvchangze/SpikeCLIP.
The Vision Transformer (ViT) architecture has emerged as the backbone of choice for state-of-the-art deep models for computer vision applications. However, ViTs are ill-suited for private inference using secure multi-party computation (MPC) protocols, due to the large number of non-polynomial operations (self-attention, feed-forward rectifiers, layer normalization). We propose PriViT, a gradient based algorithm to selectively "Taylorize" nonlinearities in ViTs while maintaining their prediction accuracy. Our algorithm is conceptually simple, easy to implement, and achieves improved performance over existing approaches for designing MPC-friendly transformer architectures in terms of achieving the Pareto frontier in latency-accuracy. We confirm these improvements via experiments on several standard image classification tasks. Public code is available at https://github.com/NYU-DICE-Lab/privit.
IoT devices are increasingly the source of data for machine learning (ML) applications running on edge servers. Data transmissions from devices to servers are often over local wireless networks whose bandwidth is not just limited but, more importantly, variable. Furthermore, in cyber-physical systems interacting with the physical environment, image offloading is also commonly subject to timing constraints. It is, therefore, important to develop an adaptive approach that maximizes the inference performance of ML applications under timing constraints and the resource constraints of IoT devices. In this paper, we use image classification as our target application and propose progressive neural compression (PNC) as an efficient solution to this problem. Although neural compression has been used to compress images for different ML applications, existing solutions often produce fixed-size outputs that are unsuitable for timing-constrained offloading over variable bandwidth. To address this limitation, we train a multi-objective rateless autoencoder that optimizes for multiple compression rates via stochastic taildrop to create a compression solution that produces features ordered according to their importance to inference performance. Features are then transmitted in that order based on available bandwidth, with classification ultimately performed using the (sub)set of features received by the deadline. We demonstrate the benefits of PNC over state-of-the-art neural compression approaches and traditional compression methods on a testbed comprising an IoT device and an edge server connected over a wireless network with varying bandwidth.
Fusion is a technique for merging multiple independently-trained neural networks in order to combine their capabilities. Past attempts have been restricted to the case of fully-connected, convolutional, and residual networks. This paper presents a systematic approach for fusing two or more transformer-based networks exploiting Optimal Transport to (soft-)align the various architectural components. We flesh out an abstraction for layer alignment, that can generalize to arbitrary architectures - in principle - and we apply this to the key ingredients of Transformers such as multi-head self-attention, layer-normalization, and residual connections, and we discuss how to handle them via various ablation studies. Furthermore, our method allows the fusion of models of different sizes (heterogeneous fusion), providing a new and efficient way to compress Transformers. The proposed approach is evaluated on both image classification tasks via Vision Transformer and natural language modeling tasks using BERT. Our approach consistently outperforms vanilla fusion, and, after a surprisingly short finetuning, also outperforms the individual converged parent models. In our analysis, we uncover intriguing insights about the significant role of soft alignment in the case of Transformers. Our results showcase the potential of fusing multiple Transformers, thus compounding their expertise, in the budding paradigm of model fusion and recombination. Code is available at https://github.com/graldij/transformer-fusion.
While models derived from Vision Transformers (ViTs) have been phonemically surging, pre-trained models cannot seamlessly adapt to arbitrary resolution images without altering the architecture and configuration, such as sampling the positional encoding, limiting their flexibility for various vision tasks. For instance, the Segment Anything Model (SAM) based on ViT-Huge requires all input images to be resized to 1024$\times$1024. To overcome this limitation, we propose the Multi-Head Self-Attention Convolution (MSA-Conv) that incorporates Self-Attention within generalized convolutions, including standard, dilated, and depthwise ones. Enabling transformers to handle images of varying sizes without retraining or rescaling, the use of MSA-Conv further reduces computational costs compared to global attention in ViT, which grows costly as image size increases. Later, we present the Vision Transformer in Convolution (TiC) as a proof of concept for image classification with MSA-Conv, where two capacity enhancing strategies, namely Multi-Directional Cyclic Shifted Mechanism and Inter-Pooling Mechanism, have been proposed, through establishing long-distance connections between tokens and enlarging the effective receptive field. Extensive experiments have been carried out to validate the overall effectiveness of TiC. Additionally, ablation studies confirm the performance improvement made by MSA-Conv and the two capacity enhancing strategies separately. Note that our proposal aims at studying an alternative to the global attention used in ViT, while MSA-Conv meets our goal by making TiC comparable to state-of-the-art on ImageNet-1K. Code will be released at https://github.com/zs670980918/MSA-Conv.
Weight decay is a broadly used technique for training state-of-the-art deep networks from image classification to large language models. Despite its widespread usage and being extensively studied in the classical literature, its role remains poorly understood for deep learning. In this work, we highlight that the role of weight decay in modern deep learning is different from its regularization effect studied in classical learning theory. For deep networks on vision tasks trained with multipass SGD, we show how weight decay modifies the optimization dynamics enhancing the ever-present implicit regularization of SGD via the loss stabilization mechanism. In contrast, for large language models trained with nearly one-epoch training, we describe how weight decay balances the bias-variance tradeoff in stochastic optimization leading to lower training loss and improved training stability. Overall, we present a unifying perspective from ResNets on vision tasks to LLMs: weight decay is never useful as an explicit regularizer but instead changes the training dynamics in a desirable way. The code is available at https://github.com/tml-epfl/why-weight-decay