Numerous deep learning algorithms have been inspired by and understood via the notion of information bottleneck, where unnecessary information is (often implicitly) minimized while task-relevant information is maximized. However, a rigorous argument for justifying why it is desirable to control information bottlenecks has been elusive. In this paper, we provide the first rigorous learning theory for justifying the benefit of information bottleneck in deep learning by mathematically relating information bottleneck to generalization errors. Our theory proves that controlling information bottleneck is one way to control generalization errors in deep learning, although it is not the only or necessary way. We investigate the merit of our new mathematical findings with experiments across a range of architectures and learning settings. In many cases, generalization errors are shown to correlate with the degree of information bottleneck: i.e., the amount of the unnecessary information at hidden layers. This paper provides a theoretical foundation for current and future methods through the lens of information bottleneck. Our new generalization bounds scale with the degree of information bottleneck, unlike the previous bounds that scale with the number of parameters, VC dimension, Rademacher complexity, stability or robustness. Our code is publicly available at: https://github.com/xu-ji/information-bottleneck
Deep learning models are known to overfit and memorize spurious features in the training dataset. While numerous empirical studies have aimed at understanding this phenomenon, a rigorous theoretical framework to quantify it is still missing. In this paper, we consider spurious features that are uncorrelated with the learning task, and we provide a precise characterization of how they are memorized via two separate terms: (i) the stability of the model with respect to individual training samples, and (ii) the feature alignment between the spurious feature and the full sample. While the first term is well established in learning theory and it is connected to the generalization error in classical work, the second one is, to the best of our knowledge, novel. Our key technical result gives a precise characterization of the feature alignment for the two prototypical settings of random features (RF) and neural tangent kernel (NTK) regression. We prove that the memorization of spurious features weakens as the generalization capability increases and, through the analysis of the feature alignment, we unveil the role of the model and of its activation function. Numerical experiments show the predictive power of our theory on standard datasets (MNIST, CIFAR-10).
Long-term stability and physical consistency are critical properties for AI-based weather models if they are going to be used for subseasonal-to-seasonal forecasts or beyond, e.g., climate change projection. However, current AI-based weather models can only provide short-term forecasts accurately since they become unstable or physically inconsistent when time-integrated beyond a few weeks or a few months. Either they exhibit numerical blow-up or hallucinate unrealistic dynamics of the atmospheric variables, akin to the current class of autoregressive large language models. The cause of the instabilities is unknown, and the methods that are used to improve their stability horizons are ad-hoc and lack rigorous theory. In this paper, we reveal that the universal causal mechanism for these instabilities in any turbulent flow is due to \textit{spectral bias} wherein, \textit{any} deep learning architecture is biased to learn only the large-scale dynamics and ignores the small scales completely. We further elucidate how turbulence physics and the absence of convergence in deep learning-based time-integrators amplify this bias, leading to unstable error propagation. Finally, using the quasi-geostrophic flow and European Center for Medium-Range Weather Forecasting (ECMWF) Reanalysis data as test cases, we bridge the gap between deep learning theory and numerical analysis to propose one mitigative solution to such unphysical behavior. We develop long-term physically-consistent data-driven models for the climate system and demonstrate accurate short-term forecasts, and hundreds of years of time-integration with accurate mean and variability.
Double descent is a surprising phenomenon in machine learning, in which as the number of model parameters grows relative to the number of data, test error drops as models grow ever larger into the highly overparameterized (data undersampled) regime. This drop in test error flies against classical learning theory on overfitting and has arguably underpinned the success of large models in machine learning. This non-monotonic behavior of test loss depends on the number of data, the dimensionality of the data and the number of model parameters. Here, we briefly describe double descent, then provide an explanation of why double descent occurs in an informal and approachable manner, requiring only familiarity with linear algebra and introductory probability. We provide visual intuition using polynomial regression, then mathematically analyze double descent with ordinary linear regression and identify three interpretable factors that, when simultaneously all present, together create double descent. We demonstrate that double descent occurs on real data when using ordinary linear regression, then demonstrate that double descent does not occur when any of the three factors are ablated. We use this understanding to shed light on recent observations in nonlinear models concerning superposition and double descent. Code is publicly available.
In this work, we advocate for the importance of singular learning theory (SLT) as it pertains to the theory and practice of variational inference in Bayesian neural networks (BNNs). To begin, using SLT, we lay to rest some of the confusion surrounding discrepancies between downstream predictive performance measured via e.g., the test log predictive density, and the variational objective. Next, we use the SLT-corrected asymptotic form for singular posterior distributions to inform the design of the variational family itself. Specifically, we build upon the idealized variational family introduced in \citet{bhattacharya_evidence_2020} which is theoretically appealing but practically intractable. Our proposal takes shape as a normalizing flow where the base distribution is a carefully-initialized generalized gamma. We conduct experiments comparing this to the canonical Gaussian base distribution and show improvements in terms of variational free energy and variational generalization error.
Providing generalization guarantees for modern neural networks has been a crucial task in statistical learning. Recently, several studies have attempted to analyze the generalization error in such settings by using tools from fractal geometry. While these works have successfully introduced new mathematical tools to apprehend generalization, they heavily rely on a Lipschitz continuity assumption, which in general does not hold for neural networks and might make the bounds vacuous. In this work, we address this issue and prove fractal geometry-based generalization bounds without requiring any Lipschitz assumption. To achieve this goal, we build up on a classical covering argument in learning theory and introduce a data-dependent fractal dimension. Despite introducing a significant amount of technical complications, this new notion lets us control the generalization error (over either fixed or random hypothesis spaces) along with certain mutual information (MI) terms. To provide a clearer interpretation to the newly introduced MI terms, as a next step, we introduce a notion of "geometric stability" and link our bounds to the prior art. Finally, we make a rigorous connection between the proposed data-dependent dimension and topological data analysis tools, which then enables us to compute the dimension in a numerically efficient way. We support our theory with experiments conducted on various settings.
To cope with real-world dynamics, an intelligent system needs to incrementally acquire, update, accumulate, and exploit knowledge throughout its lifetime. This ability, known as continual learning, provides a foundation for AI systems to develop themselves adaptively. In a general sense, continual learning is explicitly limited by catastrophic forgetting, where learning a new task usually results in a dramatic performance degradation of the old tasks. Beyond this, increasingly numerous advances have emerged in recent years that largely extend the understanding and application of continual learning. The growing and widespread interest in this direction demonstrates its realistic significance as well as complexity. In this work, we present a comprehensive survey of continual learning, seeking to bridge the basic settings, theoretical foundations, representative methods, and practical applications. Based on existing theoretical and empirical results, we summarize the general objectives of continual learning as ensuring a proper stability-plasticity trade-off and an adequate intra/inter-task generalizability in the context of resource efficiency. Then we provide a state-of-the-art and elaborated taxonomy, extensively analyzing how representative methods address continual learning, and how they are adapted to particular challenges in realistic applications. Through an in-depth discussion of promising directions, we believe that such a holistic perspective can greatly facilitate subsequent exploration in this field and beyond.
We analyze the Nystr\"om approximation of a positive definite kernel associated with a probability measure. We first prove an improved error bound for the conventional Nystr\"om approximation with i.i.d. sampling and singular-value decomposition in the continuous regime; the proof techniques are borrowed from statistical learning theory. We further introduce a refined selection of subspaces in Nystr\"om approximation with theoretical guarantees that is applicable to non-i.i.d. landmark points. Finally, we discuss their application to convex kernel quadrature and give novel theoretical guarantees as well as numerical observations.
Graph neural networks (GNNs) excel in modeling relational data such as biological, social, and transportation networks, but the underpinnings of their success are not well understood. Traditional complexity measures from statistical learning theory fail to account for observed phenomena like the double descent or the impact of relational semantics on generalization error. Motivated by experimental observations of ``transductive'' double descent in key networks and datasets, we use analytical tools from statistical physics and random matrix theory to precisely characterize generalization in simple graph convolution networks on the contextual stochastic block model. Our results illuminate the nuances of learning on homophilic versus heterophilic data and predict double descent whose existence in GNNs has been questioned by recent work. We show how risk is shaped by the interplay between the graph noise, feature noise, and the number of training labels. Our findings apply beyond stylized models, capturing qualitative trends in real-world GNNs and datasets. As a case in point, we use our analytic insights to improve performance of state-of-the-art graph convolution networks on heterophilic datasets.
Grammatical inference is a classical problem in computational learning theory and a topic of wider influence in natural language processing. We treat grammars as a model of computation and propose a novel neural approach to induction of regular grammars from positive and negative examples. Our model is fully explainable, its intermediate results are directly interpretable as partial parses, and it can be used to learn arbitrary regular grammars when provided with sufficient data. We find that our method consistently attains high recall and precision scores across a range of tests of varying complexity.
Proof-of-Learning (PoL) proposes that a model owner logs training checkpoints to establish a proof of having expended the computation necessary for training. The authors of PoL forego cryptographic approaches and trade rigorous security guarantees for scalability to deep learning. They empirically argued the benefit of this approach by showing how spoofing--computing a proof for a stolen model--is as expensive as obtaining the proof honestly by training the model. However, recent work has provided a counter-example and thus has invalidated this observation. In this work we demonstrate, first, that while it is true that current PoL verification is not robust to adversaries, recent work has largely underestimated this lack of robustness. This is because existing spoofing strategies are either unreproducible or target weakened instantiations of PoL--meaning they are easily thwarted by changing hyperparameters of the verification. Instead, we introduce the first spoofing strategies that can be reproduced across different configurations of the PoL verification and can be done for a fraction of the cost of previous spoofing strategies. This is possible because we identify key vulnerabilities of PoL and systematically analyze the underlying assumptions needed for robust verification of a proof. On the theoretical side, we show how realizing these assumptions reduces to open problems in learning theory.We conclude that one cannot develop a provably robust PoL verification mechanism without further understanding of optimization in deep learning.
Investigating cooperativity of interlocutors is central in studying pragmatics of dialogue. Models of conversation that only assume cooperative agents fail to explain the dynamics of strategic conversations. Thus, we investigate the ability of agents to identify non-cooperative interlocutors while completing a concurrent visual-dialogue task. Within this novel setting, we study the optimality of communication strategies for achieving this multi-task objective. We use the tools of learning theory to develop a theoretical model for identifying non-cooperative interlocutors and apply this theory to analyze different communication strategies. We also introduce a corpus of non-cooperative conversations about images in the GuessWhat?! dataset proposed by De Vries et al. (2017). We use reinforcement learning to implement multiple communication strategies in this context and find empirical results validate our theory.
The integrity of training data, even when annotated by experts, is far from guaranteed, especially for non-IID datasets comprising both in- and out-of-distribution samples. In an ideal scenario, the majority of samples would be in-distribution, while samples that deviate semantically would be identified as out-of-distribution and excluded during the annotation process. However, experts may erroneously classify these out-of-distribution samples as in-distribution, assigning them labels that are inherently unreliable. This mixture of unreliable labels and varied data types makes the task of learning robust neural networks notably challenging. We observe that both in- and out-of-distribution samples can almost invariably be ruled out from belonging to certain classes, aside from those corresponding to unreliable ground-truth labels. This opens the possibility of utilizing reliable complementary labels that indicate the classes to which a sample does not belong. Guided by this insight, we introduce a novel approach, termed \textit{Gray Learning} (GL), which leverages both ground-truth and complementary labels. Crucially, GL adaptively adjusts the loss weights for these two label types based on prediction confidence levels. By grounding our approach in statistical learning theory, we derive bounds for the generalization error, demonstrating that GL achieves tight constraints even in non-IID settings. Extensive experimental evaluations reveal that our method significantly outperforms alternative approaches grounded in robust statistics.
Federated Learning is an emerging learning paradigm that allows training models from samples distributed across a large network of clients while respecting privacy and communication restrictions. Despite its success, federated learning faces several challenges related to its decentralized nature. In this work, we develop a novel algorithmic procedure with theoretical speedup guarantees that simultaneously handles two of these hurdles, namely (i) data heterogeneity, i.e., data distributions can vary substantially across clients, and (ii) system heterogeneity, i.e., the computational power of the clients could differ significantly. Our method relies on ideas from representation learning theory to find a global common representation using all clients' data and learn a user-specific set of parameters leading to a personalized solution for each client. Furthermore, our method mitigates the effects of stragglers by adaptively selecting clients based on their computational characteristics and statistical significance, thus achieving, for the first time, near optimal sample complexity and provable logarithmic speedup. Experimental results support our theoretical findings showing the superiority of our method over alternative personalized federated schemes in system and data heterogeneous environments.
To unveil how the brain learns, ongoing work seeks biologically-plausible approximations of gradient descent algorithms for training recurrent neural networks (RNNs). Yet, beyond task accuracy, it is unclear if such learning rules converge to solutions that exhibit different levels of generalization than their nonbiologically-plausible counterparts. Leveraging results from deep learning theory based on loss landscape curvature, we ask: how do biologically-plausible gradient approximations affect generalization? We first demonstrate that state-of-the-art biologically-plausible learning rules for training RNNs exhibit worse and more variable generalization performance compared to their machine learning counterparts that follow the true gradient more closely. Next, we verify that such generalization performance is correlated significantly with loss landscape curvature, and we show that biologically-plausible learning rules tend to approach high-curvature regions in synaptic weight space. Using tools from dynamical systems, we derive theoretical arguments and present a theorem explaining this phenomenon. This predicts our numerical results, and explains why biologically-plausible rules lead to worse and more variable generalization properties. Finally, we suggest potential remedies that could be used by the brain to mitigate this effect. To our knowledge, our analysis is the first to identify the reason for this generalization gap between artificial and biologically-plausible learning rules, which can help guide future investigations into how the brain learns solutions that generalize.
Statistical learning theory provides bounds on the necessary number of training samples needed to reach a prescribed accuracy in a learning problem formulated over a given target class. This accuracy is typically measured in terms of a generalization error, that is, an expected value of a given loss function. However, for several applications -- for example in a security-critical context or for problems in the computational sciences -- accuracy in this sense is not sufficient. In such cases, one would like to have guarantees for high accuracy on every input value, that is, with respect to the uniform norm. In this paper we precisely quantify the number of training samples needed for any conceivable training algorithm to guarantee a given uniform accuracy on any learning problem formulated over target classes containing (or consisting of) ReLU neural networks of a prescribed architecture. We prove that, under very general assumptions, the minimal number of training samples for this task scales exponentially both in the depth and the input dimension of the network architecture.
The reliability of a learning model is key to the successful deployment of machine learning in various industries. Creating a robust model, particularly one unaffected by adversarial attacks, requires a comprehensive understanding of the adversarial examples phenomenon. However, it is difficult to describe the phenomenon due to the complicated nature of the problems in machine learning. Consequently, many studies investigate the phenomenon by proposing a simplified model of how adversarial examples occur and validate it by predicting some aspect of the phenomenon. While these studies cover many different characteristics of the adversarial examples, they have not reached a holistic approach to the geometric and analytic modeling of the phenomenon. This paper propose a formal framework to study the phenomenon in learning theory and make use of complex analysis and holomorphicity to offer a robust learning rule for artificial neural networks. With the help of complex analysis, we can effortlessly move between geometric and analytic perspectives of the phenomenon and offer further insights on the phenomenon by revealing its connection with harmonic functions. Using our model, we can explain some of the most intriguing characteristics of adversarial examples, including transferability of adversarial examples, and pave the way for novel approaches to mitigate the effects of the phenomenon.
In learning theory, the performance of an online policy is commonly measured in terms of the static regret metric, which compares the cumulative loss of an online policy to that of an optimal benchmark in hindsight. In the definition of static regret, the action of the benchmark policy remains fixed throughout the time horizon. Naturally, the resulting regret bounds become loose in non-stationary settings where fixed actions often suffer from poor performance. In this paper, we investigate a stronger notion of regret minimization in the context of online caching. In particular, we allow the action of the benchmark at any round to be decided by a finite state machine containing any number of states. Popular caching policies, such as LRU and FIFO, belong to this class. Using ideas from the universal prediction literature in information theory, we propose an efficient online caching policy with a sub-linear regret bound. To the best of our knowledge, this is the first data-dependent regret bound known for the caching problem in the universal setting. We establish this result by combining a recently-proposed online caching policy with an incremental parsing algorithm, namely Lempel-Ziv '78. Our methods also yield a simpler learning-theoretic proof of the improved regret bound as opposed to the involved problem-specific combinatorial arguments used in the earlier works.
Weight norm $\|w\|$ and margin $\gamma$ participate in learning theory via the normalized margin $\gamma/\|w\|$. Since standard neural net optimizers do not control normalized margin, it is hard to test whether this quantity causally relates to generalization. This paper designs a series of experimental studies that explicitly control normalized margin and thereby tackle two central questions. First: does normalized margin always have a causal effect on generalization? The paper finds that no -- networks can be produced where normalized margin has seemingly no relationship with generalization, counter to the theory of Bartlett et al. (2017). Second: does normalized margin ever have a causal effect on generalization? The paper finds that yes -- in a standard training setup, test performance closely tracks normalized margin. The paper suggests a Gaussian process model as a promising explanation for this behavior.
Learning from Label Proportions (LLP) is a weakly supervised learning method that aims to perform instance classification from training data consisting of pairs of bags containing multiple instances and the class label proportions within the bags. Previous studies on multiclass LLP can be divided into two categories according to the learning task: per-instance label classification and per-bag label proportion estimation. However, these methods often results in high variance estimates of the risk when applied to complex models, or lack statistical learning theory arguments. To address this issue, we propose new learning methods based on statistical learning theory for both per-instance and per-bag policies. We demonstrate that the proposed methods are respectively risk-consistent and classifier-consistent in an instance-wise manner, and analyze the estimation error bounds. Additionally, we present a heuristic approximation method that utilizes an existing method for regressing label proportions to reduce the computational complexity of the proposed methods. Through benchmark experiments, we demonstrated the effectiveness of the proposed methods.