The widely observed 'benign overfitting phenomenon' in the neural network literature raises the challenge to the 'bias-variance trade-off' doctrine in the statistical learning theory. Since the generalization ability of the 'lazy trained' over-parametrized neural network can be well approximated by that of the neural tangent kernel regression, the curve of the excess risk (namely, the learning curve) of kernel ridge regression attracts increasing attention recently. However, most recent arguments on the learning curve are heuristic and are based on the 'Gaussian design' assumption. In this paper, under mild and more realistic assumptions, we rigorously provide a full characterization of the learning curve: elaborating the effect and the interplay of the choice of the regularization parameter, the source condition and the noise. In particular, our results suggest that the 'benign overfitting phenomenon' exists in very wide neural networks only when the noise level is small.
Learning and decision-making in domains with naturally high noise-to-signal ratio, such as Finance or Healthcare, is often challenging, while the stakes are very high. In this paper, we study the problem of learning and acting under a general noisy generative process. In this problem, the data distribution has a significant proportion of uninformative samples with high noise in the label, while part of the data contains useful information represented by low label noise. This dichotomy is present during both training and inference, which requires the proper handling of uninformative data during both training and testing. We propose a novel approach to learning under these conditions via a loss inspired by the selective learning theory. By minimizing this loss, the model is guaranteed to make a near-optimal decision by distinguishing informative data from uninformative data and making predictions. We build upon the strength of our theoretical guarantees by describing an iterative algorithm, which jointly optimizes both a predictor and a selector, and evaluates its empirical performance in a variety of settings.
Recent years have seen significant activity on the problem of using data for the purpose of learning properties of quantum systems or of processing classical or quantum data via quantum computing. As in classical learning, quantum learning problems involve settings in which the mechanism generating the data is unknown, and the main goal of a learning algorithm is to ensure satisfactory accuracy levels when only given access to data and, possibly, side information such as expert knowledge. This article reviews the complexity of quantum learning using information-theoretic techniques by focusing on data complexity, copy complexity, and model complexity. Copy complexity arises from the destructive nature of quantum measurements, which irreversibly alter the state to be processed, limiting the information that can be extracted about quantum data. For example, in a quantum system, unlike in classical machine learning, it is generally not possible to evaluate the training loss simultaneously on multiple hypotheses using the same quantum data. To make the paper self-contained and approachable by different research communities, we provide extensive background material on classical results from statistical learning theory, as well as on the distinguishability of quantum states. Throughout, we highlight the differences between quantum and classical learning by addressing both supervised and unsupervised learning, and we provide extensive pointers to the literature.
We address the challenge of solving machine learning tasks using data from privacy-sensitive sellers. Since the data is private, we design a data market that incentivizes sellers to provide their data in exchange for payments. Therefore our objective is to design a mechanism that optimizes a weighted combination of test loss, seller privacy, and payment, striking a balance between building a good privacy-preserving ML model and minimizing payments to the sellers. To achieve this, we first propose an approach to solve logistic regression with known heterogeneous differential privacy guarantees. Building on these results and leveraging standard mechanism design theory, we develop a two-step optimization framework. We further extend this approach to an online algorithm that handles the sequential arrival of sellers.
The key assumption underlying linear Markov Decision Processes (MDPs) is that the learner has access to a known feature map $\phi(x, a)$ that maps state-action pairs to $d$-dimensional vectors, and that the rewards and transitions are linear functions in this representation. But where do these features come from? In the absence of expert domain knowledge, a tempting strategy is to use the ``kitchen sink" approach and hope that the true features are included in a much larger set of potential features. In this paper we revisit linear MDPs from the perspective of feature selection. In a $k$-sparse linear MDP, there is an unknown subset $S \subset [d]$ of size $k$ containing all the relevant features, and the goal is to learn a near-optimal policy in only poly$(k,\log d)$ interactions with the environment. Our main result is the first polynomial-time algorithm for this problem. In contrast, earlier works either made prohibitively strong assumptions that obviated the need for exploration, or required solving computationally intractable optimization problems. Along the way we introduce the notion of an emulator: a succinct approximate representation of the transitions that suffices for computing certain Bellman backups. Since linear MDPs are a non-parametric model, it is not even obvious whether polynomial-sized emulators exist. We show that they do exist and can be computed efficiently via convex programming. As a corollary of our main result, we give an algorithm for learning a near-optimal policy in block MDPs whose decoding function is a low-depth decision tree; the algorithm runs in quasi-polynomial time and takes a polynomial number of samples. This can be seen as a reinforcement learning analogue of classic results in computational learning theory. Furthermore, it gives a natural model where improving the sample complexity via representation learning is computationally feasible.
In standard supervised machine learning, it is necessary to provide a label for every input in the data. While raw data in many application domains is easily obtainable on the Internet, manual labelling of this data is prohibitively expensive. To circumvent this issue, contrastive learning methods produce low-dimensional vector representations (also called features) of high-dimensional inputs on large unlabelled datasets. This is done by training with a contrastive loss function, which enforces that similar inputs have high inner product and dissimilar inputs have low inner product in the feature space. Rather than annotating each input individually, it suffices to define a means of sampling pairs of similar and dissimilar inputs. Contrastive features can then be fed as inputs to supervised learning systems on much smaller labelled datasets to obtain high accuracy on end tasks of interest. The goal of this thesis is to provide an overview of the current theoretical understanding of contrastive learning, specifically as it pertains to the minimizers of contrastive loss functions and their relationship to prior methods for learning features from unlabelled data. We highlight popular contrastive loss functions whose minimizers implicitly approximate a positive semidefinite (PSD) kernel. The latter is a well-studied object in functional analysis and learning theory that formalizes a notion of similarity between elements of a space. PSD kernels provide an implicit definition of features through the theory of reproducing kernel Hilbert spaces.
The protection of Industrial Control Systems (ICS) that are employed in public critical infrastructures is of utmost importance due to catastrophic physical damages cyberattacks may cause. The research community requires testbeds for validation and comparing various intrusion detection algorithms to protect ICS. However, there exist high barriers to entry for research and education in the ICS cybersecurity domain due to expensive hardware, software, and inherent dangers of manipulating real-world systems. To close the gap, built upon recently developed 3D high-fidelity simulators, we further showcase our integrated framework to automatically launch cyberattacks, collect data, train machine learning models, and evaluate for practical chemical and manufacturing processes. On our testbed, we validate our proposed intrusion detection model called Minimal Threshold and Window SVM (MinTWin SVM) that utilizes unsupervised machine learning via a one-class SVM in combination with a sliding window and classification threshold. Results show that MinTWin SVM minimizes false positives and is responsive to physical process anomalies. Furthermore, we incorporate our framework with ICS cybersecurity education by using our dataset in an undergraduate machine learning course where students gain hands-on experience in practicing machine learning theory with a practical ICS dataset. All of our implementations have been open-sourced.
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.
It is becoming increasingly important to understand the vulnerability of machine learning models to adversarial attacks. One of the fundamental problems in adversarial machine learning is to quantify how much training data is needed in the presence of evasion attacks, where data is corrupted at test time. In this thesis, we work with the exact-in-the-ball notion of robustness and study the feasibility of adversarially robust learning from the perspective of learning theory, considering sample complexity. We first explore the setting where the learner has access to random examples only, and show that distributional assumptions are essential. We then focus on learning problems with distributions on the input data that satisfy a Lipschitz condition and show that robustly learning monotone conjunctions has sample complexity at least exponential in the adversary's budget (the maximum number of bits it can perturb on each input). However, if the adversary is restricted to perturbing $O(\log n)$ bits, then one can robustly learn conjunctions and decision lists w.r.t. log-Lipschitz distributions. We then study learning models where the learner is given more power. We first consider local membership queries, where the learner can query the label of points near the training sample. We show that, under the uniform distribution, the exponential dependence on the adversary's budget to robustly learn conjunctions remains inevitable. We then introduce a local equivalence query oracle, which returns whether the hypothesis and target concept agree in a given region around a point in the training sample, and a counterexample if it exists. We show that if the query radius is equal to the adversary's budget, we can develop robust empirical risk minimization algorithms in the distribution-free setting. We give general query complexity upper and lower bounds, as well as for concrete concept classes.
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.
This paper explores the potential impacts of large language models (LLMs) on the Chinese labor market. We analyze occupational exposure to LLM capabilities by incorporating human expertise and LLM classifications, following the methodology of Eloundou et al. (2023). The results indicate a positive correlation between occupational exposure and both wage levels and experience premiums at the occupation level. This suggests that higher-paying and experience-intensive jobs may face greater exposure risks from LLM-powered software. We then aggregate occupational exposure at the industry level to obtain industrial exposure scores. Both occupational and industrial exposure scores align with expert assessments. Our empirical analysis also demonstrates a distinct impact of LLMs, which deviates from the routinization hypothesis. We present a stylized theoretical framework to better understand this deviation from previous digital technologies. By incorporating entropy-based information theory into the task-based framework, we propose an AI learning theory that reveals a different pattern of LLM impacts compared to the routinization hypothesis.
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.
Reliable automatic target segmentation in Synthetic Aperture Radar (SAR) imagery has played an important role in the SAR fields. Different from the traditional methods, Spectral Residual (SR) and CFAR detector, with the recent adavance in machine learning theory, there has emerged a novel method for SAR target segmentation, based on the deep learning networks. In this paper, we proposed a deep deformable residual learning network for target segmentation that attempts to preserve the precise contour of the target. For this, the deformable convolutional layers and residual learning block are applied, which could extract and preserve the geometric information of the targets as much as possible. Based on the Moving and Stationary Target Acquisition and Recognition (MSTAR) data set, experimental results have shown the superiority of the proposed network for the precise targets segmentation.
With the recent advances of deep learning, automatic target recognition (ATR) of synthetic aperture radar (SAR) has achieved superior performance. By not being limited to the target category, the SAR ATR system could benefit from the simultaneous extraction of multifarious target attributes. In this paper, we propose a new multi-task learning approach for SAR ATR, which could obtain the accurate category and precise shape of the targets simultaneously. By introducing deep learning theory into multi-task learning, we first propose a novel multi-task deep learning framework with two main structures: encoder and decoder. The encoder is constructed to extract sufficient image features in different scales for the decoder, while the decoder is a tasks-specific structure which employs these extracted features adaptively and optimally to meet the different feature demands of the recognition and segmentation. Therefore, the proposed framework has the ability to achieve superior recognition and segmentation performance. Based on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset, experimental results show the superiority of the proposed framework in terms of recognition and segmentation.
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.
In silico screening uses predictive models to select a batch of compounds with favorable properties from a library for experimental validation. Unlike conventional learning paradigms, success in this context is measured by the performance of the predictive model on the selected subset of compounds rather than the entire set of predictions. By extending learning theory, we show that the selectivity of the selection policy can significantly impact generalization, with a higher risk of errors occurring when exclusively selecting predicted positives and when targeting rare properties. Our analysis suggests a way to mitigate these challenges. We show that generalization can be markedly enhanced when considering a model's ability to predict the fraction of desired outcomes in a batch. This is promising, as the primary aim of screening is not necessarily to pinpoint the label of each compound individually, but rather to assemble a batch enriched for desirable compounds. Our theoretical insights are empirically validated across diverse tasks, architectures, and screening scenarios, underscoring their applicability.
Support Vector Machines (SVM) have gathered significant acclaim as classifiers due to their successful implementation of Statistical Learning Theory. However, in the context of multiclass and multilabel settings, the reliance on vector-based formulations in existing SVM-based models poses limitations regarding flexibility and ease of incorporating additional terms to handle specific challenges. To overcome these limitations, our research paper focuses on introducing a matrix formulation for SVM that effectively addresses these constraints. By employing the Accelerated Gradient Descent method in the dual, we notably enhance the efficiency of solving the Matrix-SVM problem. Experimental evaluations on multilabel and multiclass datasets demonstrate that Matrix SVM achieves superior time efficacy while delivering similar results to Binary Relevance SVM. Moreover, our matrix formulation unveils crucial insights and advantages that may not be readily apparent in traditional vector-based notations. We emphasize that numerous multilabel models can be viewed as extensions of SVM, with customised modifications to meet specific requirements. The matrix formulation presented in this paper establishes a solid foundation for developing more sophisticated models capable of effectively addressing the distinctive challenges encountered in multilabel learning.
Temporal credit assignment is crucial for learning and skill development in natural and artificial intelligence. While computational methods like the TD approach in reinforcement learning have been proposed, it's unclear if they accurately represent how humans handle feedback delays. Cognitive models intend to represent the mental steps by which humans solve problems and perform a number of tasks, but limited research in cognitive science has addressed the credit assignment problem in humans and cognitive models. Our research uses a cognitive model based on a theory of decisions from experience, Instance-Based Learning Theory (IBLT), to test different credit assignment mechanisms in a goal-seeking navigation task with varying levels of decision complexity. Instance-Based Learning (IBL) models simulate the process of making sequential choices with different credit assignment mechanisms, including a new IBL-TD model that combines the IBL decision mechanism with the TD approach. We found that (1) An IBL model that gives equal credit assignment to all decisions is able to match human performance better than other models, including IBL-TD and Q-learning; (2) IBL-TD and Q-learning models underperform compared to humans initially, but eventually, they outperform humans; (3) humans are influenced by decision complexity, while models are not. Our study provides insights into the challenges of capturing human behavior and the potential opportunities to use these models in future AI systems to support human activities.
We prove that it is NP-hard to properly PAC learn decision trees with queries, resolving a longstanding open problem in learning theory (Bshouty 1993; Guijarro-Lavin-Raghavan 1999; Mehta-Raghavan 2002; Feldman 2016). While there has been a long line of work, dating back to (Pitt-Valiant 1988), establishing the hardness of properly learning decision trees from random examples, the more challenging setting of query learners necessitates different techniques and there were no previous lower bounds. En route to our main result, we simplify and strengthen the best known lower bounds for a different problem of Decision Tree Minimization (Zantema-Bodlaender 2000; Sieling 2003). On a technical level, we introduce the notion of hardness distillation, which we study for decision tree complexity but can be considered for any complexity measure: for a function that requires large decision trees, we give a general method for identifying a small set of inputs that is responsible for its complexity. Our technique even rules out query learners that are allowed constant error. This contrasts with existing lower bounds for the setting of random examples which only hold for inverse-polynomial error. Our result, taken together with a recent almost-polynomial time query algorithm for properly learning decision trees under the uniform distribution (Blanc-Lange-Qiao-Tan 2022), demonstrates the dramatic impact of distributional assumptions on the problem.