Recent weakly supervised semantic segmentation (WSSS) methods strive to incorporate contextual knowledge to improve the completeness of class activation maps (CAM). In this work, we argue that the knowledge bias between instances and contexts affects the capability of the prototype to sufficiently understand instance semantics. Inspired by prototype learning theory, we propose leveraging prototype awareness to capture diverse and fine-grained feature attributes of instances. The hypothesis is that contextual prototypes might erroneously activate similar and frequently co-occurring object categories due to this knowledge bias. Therefore, we propose to enhance the prototype representation ability by mitigating the bias to better capture spatial coverage in semantic object regions. With this goal, we present a Context Prototype-Aware Learning (CPAL) strategy, which leverages semantic context to enrich instance comprehension. The core of this method is to accurately capture intra-class variations in object features through context-aware prototypes, facilitating the adaptation to the semantic attributes of various instances. We design feature distribution alignment to optimize prototype awareness, aligning instance feature distributions with dense features. In addition, a unified training framework is proposed to combine label-guided classification supervision and prototypes-guided self-supervision. Experimental results on PASCAL VOC 2012 and MS COCO 2014 show that CPAL significantly improves off-the-shelf methods and achieves state-of-the-art performance. The project is available at https://github.com/Barrett-python/CPAL.
While deep neural networks have demonstrated groundbreaking performance in various settings, these models often suffer from \emph{catastrophic forgetting} when trained on new tasks in sequence. Several works have empirically demonstrated that increasing the width of a neural network leads to a decrease in catastrophic forgetting but have yet to characterize the exact relationship between width and continual learning. We design one of the first frameworks to analyze Continual Learning Theory and prove that width is directly related to forgetting in Feed-Forward Networks (FFN). Specifically, we demonstrate that increasing network widths to reduce forgetting yields diminishing returns. We empirically verify our claims at widths hitherto unexplored in prior studies where the diminishing returns are clearly observed as predicted by our theory.
In statistical learning theory, a generalization bound usually involves a complexity measure imposed by the considered theoretical framework. This limits the scope of such bounds, as other forms of capacity measures or regularizations are used in algorithms. In this paper, we leverage the framework of disintegrated PAC-Bayes bounds to derive a general generalization bound instantiable with arbitrary complexity measures. One trick to prove such a result involves considering a commonly used family of distributions: the Gibbs distributions. Our bound stands in probability jointly over the hypothesis and the learning sample, which allows the complexity to be adapted to the generalization gap as it can be customized to fit both the hypothesis class and the task.
We investigate the test risk of continuous-time stochastic gradient flow dynamics in learning theory. Using a path integral formulation we provide, in the regime of a small learning rate, a general formula for computing the difference between test risk curves of pure gradient and stochastic gradient flows. We apply the general theory to a simple model of weak features, which displays the double descent phenomenon, and explicitly compute the corrections brought about by the added stochastic term in the dynamics, as a function of time and model parameters. The analytical results are compared to simulations of discrete-time stochastic gradient descent and show good agreement.
In today's data-driven world, the proliferation of publicly available information raises security concerns due to the information leakage (IL) problem. IL involves unintentionally exposing sensitive information to unauthorized parties via observable system information. Conventional statistical approaches rely on estimating mutual information (MI) between observable and secret information for detecting ILs, face challenges of the curse of dimensionality, convergence, computational complexity, and MI misestimation. Though effective, emerging supervised machine learning based approaches to detect ILs are limited to binary system sensitive information and lack a comprehensive framework. To address these limitations, we establish a theoretical framework using statistical learning theory and information theory to quantify and detect IL accurately. Using automated machine learning, we demonstrate that MI can be accurately estimated by approximating the typically unknown Bayes predictor's log-loss and accuracy. Based on this, we show how MI can effectively be estimated to detect ILs. Our method performs superior to state-of-the-art baselines in an empirical study considering synthetic and real-world OpenSSL TLS server datasets.
In recent years, various powerful policy gradient algorithms have been proposed in deep reinforcement learning. While all these algorithms build on the Policy Gradient Theorem, the specific design choices differ significantly across algorithms. We provide a holistic overview of on-policy policy gradient algorithms to facilitate the understanding of both their theoretical foundations and their practical implementations. In this overview, we include a detailed proof of the continuous version of the Policy Gradient Theorem, convergence results and a comprehensive discussion of practical algorithms. We compare the most prominent algorithms on continuous control environments and provide insights on the benefits of regularization. All code is available at https://github.com/Matt00n/PolicyGradientsJax.
Recently, sharpness-aware minimization (SAM) has attracted much attention because of its surprising effectiveness in improving generalization performance. However, compared to stochastic gradient descent (SGD), it is more prone to getting stuck at the saddle points, which as a result may lead to performance degradation. To address this issue, we propose a simple renormalization strategy, dubbed Stable SAM (SSAM), so that the gradient norm of the descent step maintains the same as that of the ascent step. Our strategy is easy to implement and flexible enough to integrate with SAM and its variants, almost at no computational cost. With elementary tools from convex optimization and learning theory, we also conduct a theoretical analysis of sharpness-aware training, revealing that compared to SGD, the effectiveness of SAM is only assured in a limited regime of learning rate. In contrast, we show how SSAM extends this regime of learning rate and then it can consistently perform better than SAM with the minor modification. Finally, we demonstrate the improved performance of SSAM on several representative data sets and tasks.
Open-world classification systems should discern out-of-distribution (OOD) data whose labels deviate from those of in-distribution (ID) cases, motivating recent studies in OOD detection. Advanced works, despite their promising progress, may still fail in the open world, owing to the lack of knowledge about unseen OOD data in advance. Although one can access auxiliary OOD data (distinct from unseen ones) for model training, it remains to analyze how such auxiliary data will work in the open world. To this end, we delve into such a problem from a learning theory perspective, finding that the distribution discrepancy between the auxiliary and the unseen real OOD data is the key to affecting the open-world detection performance. Accordingly, we propose Distributional-Augmented OOD Learning (DAL), alleviating the OOD distribution discrepancy by crafting an OOD distribution set that contains all distributions in a Wasserstein ball centered on the auxiliary OOD distribution. We justify that the predictor trained over the worst OOD data in the ball can shrink the OOD distribution discrepancy, thus improving the open-world detection performance given only the auxiliary OOD data. We conduct extensive evaluations across representative OOD detection setups, demonstrating the superiority of our DAL over its advanced counterparts.
In this work, we formally prove that, under certain conditions, if a neural network is invariant to a finite group then its weights recover the Fourier transform on that group. This provides a mathematical explanation for the emergence of Fourier features -- a ubiquitous phenomenon in both biological and artificial learning systems. The results hold even for non-commutative groups, in which case the Fourier transform encodes all the irreducible unitary group representations. Our findings have consequences for the problem of symmetry discovery. Specifically, we demonstrate that the algebraic structure of an unknown group can be recovered from the weights of a network that is at least approximately invariant within certain bounds. Overall, this work contributes to a foundation for an algebraic learning theory of invariant neural network representations.
In this work we revisit the most fundamental building block in deep learning, the multi-layer perceptron (MLP), and study the limits of its performance on vision tasks. Empirical insights into MLPs are important for multiple reasons. (1) Given the recent narrative "less inductive bias is better", popularized due to transformers eclipsing convolutional models, it is natural to explore the limits of this hypothesis. To that end, MLPs offer an ideal test bed, as they lack any vision-specific inductive bias. (2) MLPs have almost exclusively been the main protagonist in the deep learning theory literature due to their mathematical simplicity, serving as a proxy to explain empirical phenomena observed for more complex architectures. Surprisingly, experimental datapoints for MLPs are very difficult to find in the literature, especially when coupled with large pre-training protocols. This discrepancy between practice and theory is worrying: Do MLPs reflect the empirical advances exhibited by practical models? Or do theorists need to rethink the role of MLPs as a proxy? We provide insights into both these aspects. We show that the performance of MLPs drastically improves with scale (95% on CIFAR10, 82% on CIFAR100, 58% on ImageNet ReaL), highlighting that lack of inductive bias can indeed be compensated. We observe that MLPs mimic the behaviour of their modern counterparts faithfully, with some components in the learning setting however exhibiting stronger or unexpected behaviours. Due to their inherent computational efficiency, large pre-training experiments become more accessible for academic researchers. All of our experiments were run on a single GPU.
Unsupervised domain adaptation aims to transfer rich knowledge from the annotated source domain to the unlabeled target domain with the same label space. One prevalent solution is the bi-discriminator domain adversarial network, which strives to identify target domain samples outside the support of the source domain distribution and enforces their classification to be consistent on both discriminators. Despite being effective, agnostic accuracy and overconfident estimation for out-of-distribution samples hinder its further performance improvement. To address the above challenges, we propose a novel bi-discriminator domain adversarial neural network with class-level gradient alignment, i.e. BACG. BACG resorts to gradient signals and second-order probability estimation for better alignment of domain distributions. Specifically, for accuracy-awareness, we first design an optimizable nearest neighbor algorithm to obtain pseudo-labels of samples in the target domain, and then enforce the backward gradient approximation of the two discriminators at the class level. Furthermore, following evidential learning theory, we transform the traditional softmax-based optimization method into a Multinomial Dirichlet hierarchical model to infer the class probability distribution as well as samples uncertainty, thereby alleviating misestimation of out-of-distribution samples and guaranteeing high-quality classes alignment. In addition, inspired by contrastive learning, we develop a memory bank-based variant, i.e. Fast-BACG, which can greatly shorten the training process at the cost of a minor decrease in accuracy. Extensive experiments and detailed theoretical analysis on four benchmark data sets validate the effectiveness and robustness of our algorithm.
It is known that sparsity can improve interpretability for deep neural networks. However, existing methods in the area either require networks that are pre-trained with sparsity constraints, or impose sparsity after the fact, altering the network's general behavior. In this paper, we demonstrate, for the first time, that sparsity can instead be incorporated into the interpretation process itself, as a sample-specific preprocessing step. Unlike previous work, this approach, which we call SPADE, does not place constraints on the trained model and does not affect its behavior during inference on the sample. Given a trained model and a target sample, SPADE uses sample-targeted pruning to provide a "trace" of the network's execution on the sample, reducing the network to the most important connections prior to computing an interpretation. We demonstrate that preprocessing with SPADE significantly increases the accuracy of image saliency maps across several interpretability methods. Additionally, SPADE improves the usefulness of neuron visualizations, aiding humans in reasoning about network behavior. Our code is available at https://github.com/IST-DASLab/SPADE.
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
The distribution regression problem encompasses many important statistics and machine learning tasks, and arises in a large range of applications. Among various existing approaches to tackle this problem, kernel methods have become a method of choice. Indeed, kernel distribution regression is both computationally favorable, and supported by a recent learning theory. This theory also tackles the two-stage sampling setting, where only samples from the input distributions are available. In this paper, we improve the learning theory of kernel distribution regression. We address kernels based on Hilbertian embeddings, that encompass most, if not all, of the existing approaches. We introduce the novel near-unbiased condition on the Hilbertian embeddings, that enables us to provide new error bounds on the effect of the two-stage sampling, thanks to a new analysis. We show that this near-unbiased condition holds for three important classes of kernels, based on optimal transport and mean embedding. As a consequence, we strictly improve the existing convergence rates for these kernels. Our setting and results are illustrated by numerical experiments.
The Local Learning Coefficient (LLC) is introduced as a novel complexity measure for deep neural networks (DNNs). Recognizing the limitations of traditional complexity measures, the LLC leverages Singular Learning Theory (SLT), which has long recognized the significance of singularities in the loss landscape geometry. This paper provides an extensive exploration of the LLC's theoretical underpinnings, offering both a clear definition and intuitive insights into its application. Moreover, we propose a new scalable estimator for the LLC, which is then effectively applied across diverse architectures including deep linear networks up to 100M parameters, ResNet image models, and transformer language models. Empirical evidence suggests that the LLC provides valuable insights into how training heuristics might influence the effective complexity of DNNs. Ultimately, the LLC emerges as a crucial tool for reconciling the apparent contradiction between deep learning's complexity and the principle of parsimony.
Developing effective Multi-Agent Systems (MAS) is critical for many applications requiring collaboration and coordination with humans. Despite the rapid advance of Multi-Agent Deep Reinforcement Learning (MADRL) in cooperative MAS, one major challenge is the simultaneous learning and interaction of independent agents in dynamic environments in the presence of stochastic rewards. State-of-the-art MADRL models struggle to perform well in Coordinated Multi-agent Object Transportation Problems (CMOTPs), wherein agents must coordinate with each other and learn from stochastic rewards. In contrast, humans often learn rapidly to adapt to nonstationary environments that require coordination among people. In this paper, motivated by the demonstrated ability of cognitive models based on Instance-Based Learning Theory (IBLT) to capture human decisions in many dynamic decision making tasks, we propose three variants of Multi-Agent IBL models (MAIBL). The idea of these MAIBL algorithms is to combine the cognitive mechanisms of IBLT and the techniques of MADRL models to deal with coordination MAS in stochastic environments from the perspective of independent learners. We demonstrate that the MAIBL models exhibit faster learning and achieve better coordination in a dynamic CMOTP task with various settings of stochastic rewards compared to current MADRL models. We discuss the benefits of integrating cognitive insights into MADRL models.
Complex interactions between two opposing agents frequently occur in domains of machine learning, game theory, and other application domains. Quantitatively analyzing the strategies involved can provide an objective basis for decision-making. One such critical scenario is shot-taking in football, where decisions, such as whether the attacker should shoot or pass the ball and whether the defender should attempt to block the shot, play a crucial role in the outcome of the game. However, there are currently no effective data-driven and/or theory-based approaches to analyzing such situations. To address this issue, we proposed a novel framework to analyze such scenarios based on game theory, where we estimate the expected payoff with machine learning (ML) models, and additional features for ML models were extracted with a theory-based shot block model. Conventionally, successes or failures (1 or 0) are used as payoffs, while a success shot (goal) is extremely rare in football. Therefore, we proposed the Expected Probability of Shot On Target (xSOT) metric to evaluate players' actions even if the shot results in no goal; this allows for effective differentiation and comparison between different shots and even enables counterfactual shot situation analysis. In our experiments, we have validated the framework by comparing it with baseline and ablated models. Furthermore, we have observed a high correlation between the xSOT and existing metrics. This alignment of information suggests that xSOT provides valuable insights. Lastly, as an illustration, we studied optimal strategies in the World Cup 2022 and analyzed a shot situation in EURO 2020.
We consider the general class of time-homogeneous stochastic dynamical systems, both discrete and continuous, and study the problem of learning a representation of the state that faithfully captures its dynamics. This is instrumental to learning the transfer operator or the generator of the system, which in turn can be used for numerous tasks, such as forecasting and interpreting the system dynamics. We show that the search for a good representation can be cast as an optimization problem over neural networks. Our approach is supported by recent results in statistical learning theory, highlighting the role of approximation error and metric distortion in the learning problem. The objective function we propose is associated with projection operators from the representation space to the data space, overcomes metric distortion, and can be empirically estimated from data. In the discrete-time setting, we further derive a relaxed objective function that is differentiable and numerically well-conditioned. We compare our method against state-of-the-art approaches on different datasets, showing better performance across the board.
Optical fibres aim to image in-vivo biological processes. In this context, high spatial resolution and stability to fibre movements are key to enable decision-making processes (e.g., for microendoscopy). Recently, a single-pixel imaging technique based on a multicore fibre photonic lantern has been designed, named computational optical imaging using a lantern (COIL). A proximal algorithm based on a sparsity prior, dubbed SARA-COIL, has been further proposed to solve the associated inverse problem, to enable image reconstructions for high resolution COIL microendoscopy. In this work, we develop a data-driven approach for COIL. We replace the sparsity prior in the proximal algorithm by a learned denoiser, leading to a plug-and-play (PnP) algorithm. The resulting PnP method, based on a proximal primal-dual algorithm, enables to solve the Morozov formulation of the inverse problem. We use recent results in learning theory to train a network with desirable Lipschitz properties, and we show that the resulting primal-dual PnP algorithm converges to a solution to a monotone inclusion problem. Our simulations highlight that the proposed data-driven approach improves the reconstruction quality over variational SARA-COIL method on both simulated and real data.
We develop the theory of Energy Conserving Descent (ECD) and introduce ECDSep, a gradient-based optimization algorithm able to tackle convex and non-convex optimization problems. The method is based on the novel ECD framework of optimization as physical evolution of a suitable chaotic energy-conserving dynamical system, enabling analytic control of the distribution of results - dominated at low loss - even for generic high-dimensional problems with no symmetries. Compared to previous realizations of this idea, we exploit the theoretical control to improve both the dynamics and chaos-inducing elements, enhancing performance while simplifying the hyper-parameter tuning of the optimization algorithm targeted to different classes of problems. We empirically compare with popular optimization methods such as SGD, Adam and AdamW on a wide range of machine learning problems, finding competitive or improved performance compared to the best among them on each task. We identify limitations in our analysis pointing to possibilities for additional improvements.