The paper proposes a learning-theoretic perspective on representation alignment.
problem Understanding how AI models' representations become aligned as they scale.
method Reviewing and connecting different notions of alignment, focusing on stitching.
result Relating properties of stitching to kernel alignment of representations.
The standard objective in machine learning is to train a single model for all users. However, in many learning scenarios, such as cloud computing and federated learning, it is possible to learn a personalized model per user. In this work, we present a systematic learning-theoretic study of personalization. We propose a…
This paper suggests a learning-theoretic perspective on how synaptic plasticity benefits global brain functioning. We introduce a model, the selectron, that (i) arises as the fast time constant limit of leaky integrate-and-fire neurons equipped with spiking timing dependent plasticity (STDP) and (ii) is amenable to the…
New approach prevents model collapse in language generation.
problem Model collapse risk in large language models.
method Introduces a replay adversary to study language generation.
result Replay limits generation in certain ways but not all.
Develops hypothesis tests for conditional distributions using learning-theoretic bounds.
problem Testing differences in conditional distributions and functionals.
method Transforming learning-theoretic bounds into hypothesis tests for conditional expectations.
result Establishes comprehensive foundation for conditional testing, including theoretical guarantees and practical implementations.
Graph neural networks generalize well under certain conditions, explained by learning theory.
problem Understanding why graph neural networks generalize well in transductive inference.
method Analysis of transductive Rademacher complexity to explain generalization properties of graph convolutional networks.
result Transductive Rademacher complexity can explain the generalization of graph convolutional networks for node classification in stochastic block models.
Despite the huge empirical success of deep learning, theoretical understanding of neural networks learning process is still lacking. This is the reason, why some of its features seem "mysterious". We emphasize two mysteries of deep learning: generalization mystery, and optimization mystery. In this essay we review and …
Theoretical analysis improves understanding of Deep Q-Learning's behavior.
problem Lack of formal guarantees and gaps between theory and practice of Deep Q-Learning.
method Dynamical systems perspective, focusing on realistic assumptions.
result Proves convergence of Deep Q-Learning under specific conditions.
We propose a novel combination of optimization tools with learning theory bounds in order to analyze the sample complexity of optimal kernel sum classifiers. This contrasts the typical learning theoretic results which hold for all (potentially suboptimal) classifiers. Our work also justifies assumptions made in prior w…
The paper tackles extrapolation in extreme regions of regression problems.
problem Extrapolation on the tails of covariates in continuous regression problems.
method Statistical regression on a subsample of furthest observations, focusing on their angular components, using multivariate regular variation theory.
result Quantifies predictive performance on tail regions in terms of excess risk, presenting it as a finite sample risk bound with a bias-variance decomposition.
Common statistical practice has shown that the full power of Bayesian methods is not realized until hierarchical priors are used, as these allow for greater "robustness" and the ability to "share statistical strength." Yet it is an ongoing challenge to provide a learning-theoretically sound formalism of such notions th…
Paper analyzes approximation and learning of MoEs with (P)ReLU activation.
problem Scaling up deep learning models with MoEs and (P)ReLU activation.
method Approximation and learning-theoretic analysis of MoMLPs with (P)ReLU.
result MoMLPs can uniformly approximate Lipschitz functions with ε accuracy using O(ε−1) parameters. Foundation for learning in changing conditions.
problem Learning under varying conditions and states.
method Admissible transport, protected-core preservation, and evaluator-aware learning evolution.
result Established first theorem-supporting layer for regime-varying learning.
This paper investigates domain generalization: How to take knowledge acquired from an arbitrary number of related domains and apply it to previously unseen domains? We propose Domain-Invariant Component Analysis (DICA), a kernel-based optimization algorithm that learns an invariant transformation by minimizing the diss…
New algorithm extracts features from superpositions in machine learning models.
problem Challenges in extracting interpretable features from complex models in superposition.
method An efficient query algorithm that identifies non-degenerate feature directions and reconstructs the function.
result Identifies all feature directions whose responses are non-degenerate and reconstructs the function \( f \) in a general superposition setting.
The paper connects machine learning interpretability with learning theory.
problem Performance and explanation generalization in local machine learning models.
method Theoretical analysis and empirical validation of local approximation explanations.
result Theoretical bounds on test-time accuracy and explanation generalization.
Survey of robust streaming techniques and their relationships.
problem Challenges in robust streaming and online learning.
method Overview and survey of robust streaming techniques, unifying theorems.
result Proved the relationship between robust streaming techniques.
Study identifies and analyzes three types of errors in learning Fourier operators.
problem Statistical, discretization, and truncation errors in learning Fourier operators.
method Analysis of a Discrete Fourier Transform (DFT) based least squares estimator.
result Established upper and lower bounds on statistical, discretization, and truncation errors.
Discussing AI's difficulty and physics' simplicity, suggesting AI benefits from physics principles.
problem AI difficulty compared to physics simplicity.
method Drawing on physical intuition and theoretical physics to improve AI.
result AI and physics principles are strongly coupled through sparsity.
The paper reviews MARL and its causal challenges, advocating for a 'causality first' approach.
problem Challenges in multi-agent reinforcement learning (MARL) and lack of theoretical guarantees for emergent behavior.
method Discussion of how causal methods can improve safety, interpretability, and robustness in MARL.
result Causal methods can provide strong theoretical guarantees for emergent behavior in MARL.
We derive bounds for a notion of adversarial risk, designed to characterize the robustness of linear and neural network classifiers to adversarial perturbations. Specifically, we introduce a new class of function transformations with the property that the risk of the transformed functions upper-bounds the adversarial r…
New research on limits of transfer learning, proving key selection and dependence requirements.
problem Insufficient theoretical foundation for transfer learning.
method Proved novel results on transfer learning, emphasizing selection of information and dependence between domains.
result Upper bound on improvement possible with transfer learning, highlighting the need for careful selection.
We show through theory and experiment that gradient-based explanations of a model quickly reveal the model itself. Our results speak to a tension between the desire to keep a proprietary model secret and the ability to offer model explanations. On the theoretical side, we give an algorithm that provably learns a two-la…
New model learns from random graph samples to estimate graph parameters.
problem Scalability issues in graph learning methods for large graphs.
method Develops a graph classification model working on randomly sampled subgraphs.
result Validates mini-batch learning on graphs and provides generalization bounds.
Contrastive learning outperforms autoencoders and GANs in feature recovery and downstream tasks.
problem Theoretical understanding of contrastive learning's superiority in feature learning.
method Theoretical analysis of contrastive learning in linear representation settings.
result Contrastive learning outperforms autoencoders and GANs for feature recovery and in-domain downstream tasks.
TAWT improves cross-task learning efficiency and guarantees.
problem Improving cross-task learning efficiency and theoretical understanding.
method Target-Aware Weighted Training (TAWT) minimizing representation-based task distance.
result TAWT achieves non-asymptotic learning-theoretic guarantees and is effective in NLP tasks.
We demonstrate, theoretically and empirically, that adversarial robustness can significantly benefit from semisupervised learning. Theoretically, we revisit the simple Gaussian model of Schmidt et al. that shows a sample complexity gap between standard and robust classification. We prove that unlabeled data bridges thi…
New method estimates Schrödinger bridge potentials via empirical risk minimization.
problem Estimating Schrödinger bridge potentials from samples.
method Rewriting Schrödinger system as a fixed-point equation and estimating the potential via empirical risk minimization.
result Uniform concentration of empirical risk around population counterpart under sub-Gaussian assumptions.
Unified Uncertainty Calibration improves AI predictions by combining different types of uncertainty.
problem AI classifiers struggle with uncertainty, leading to miscalibrated predictions and poor performance.
method Unified Uncertainty Calibration (U2C) combines aleatoric and epistemic uncertainties to improve prediction quality.
result U2C outperforms traditional reject-or-classify methods across various ImageNet benchmarks.
Bayesian approach learns linear operators from noisy data.
problem Learning linear operators from noisy data in infinite-dimensional spaces.
method Bayesian approach with Gaussian priors.
result Establishes posterior contraction rates and generalization error guarantees.
Proposes KMvDA for object recognition from multi-view data.
problem Recognizing objects from different views, even when views are heterogeneous.
method Introduces kernel multi-view discriminant analysis (KMvDA) and uses random Fourier features (RFF) for large-scale learning.
result KMvDA and RFF approximation improve object recognition from multi-view data.
This paper organizes recent deep learning theory advances.
problem Lack of theoretical foundations in deep learning.
method Literature review and categorization into six groups.
result Organized recent advances in deep learning theory.
GENIE balances domain-invariant feature learning and gradient alignment for improved DG performance.
problem Domain Generalization (DG) overfitting to domain-specific features
method GENIE (Generalization-ENhancing Iterative Equalizer) optimizer
result Prevents a small subset of parameters from dominating optimization, promoting domain-invariant feature learning
This thesis explores robust machine learning against adversarial examples.
problem How to create machine learning systems robust to adversarial examples.
method Theoretical exploration and development of new learning algorithms with robustness guarantees.
result Developed new learning algorithms with provable robustness guarantees.
Regularized kernel methods such as, e.g., support vector machines and least-squares support vector regression constitute an important class of standard learning algorithms in machine learning. Theoretical investigations concerning asymptotic properties have manly focused on rates of convergence during the last years bu…
New approach for algorithms that learn predictors to improve performance.
problem Improving algorithm performance by learning better predictions.
method Identify functional dependence and apply online learning techniques.
result Improved results in various online settings with simpler analysis.
Efficient algorithms recover two sparse models from a mix of linear queries.
problem Recovering two sparse models from a mix of linear queries.
method Efficient algorithms for query complexity problem.
result Improved query complexity for model recovery.
SLIM model tackles graph classification by resolving part-interaction dilemmas.
problem Difficulty in modeling graph parts and their interactions in graph classification.
method SLIM model, which solves resolution dilemmas and leverages explicit interactions.
result SLIM offers improved interpretability, accuracy, and new insights in graph representation learning.
Unified theory explains how data augmentation improves deep learning models.
problem Understanding why data augmentation improves model generalization.
method Unified theoretical framework explaining two key effects: partial semantic feature removal and feature mixing.
result Data augmentation enhances generalization through partial semantic feature removal and feature mixing.
The paper explores stability and generalization of deep GCNs.
problem Understanding the stability and generalization of deep GCNs from a theoretical perspective.
method Theoretical analysis of stability and generalization properties of deep GCNs.
result The stability and generalization of deep GCNs are influenced by the maximum absolute eigenvalue of the graph filter operators and the depth of the network.
This paper generalizes regularized regression problems in a hyper-reproducing kernel Hilbert space (hyper-RKHS), illustrates its utility for kernel learning and out-of-sample extensions, and proves asymptotic convergence results for the introduced regression models in an approximation theory view. Algorithmically, we c…
New algorithms detect and react to multiple change points in online learning.
problem Learning under multiple change points in environments with unknown and frequent shifts.
method Proposed Anytime Tracking CUSUM (ATC) algorithms that balance detection of significant shifts.
result Properly tuned ATC algorithms achieve nearly minimax-optimal performance.
This work compares human feedback methods for reward learning in bandits.
problem Understanding how human feedback affects the performance of reward learning methods.
method Theoretical comparison of human feedback approaches in offline contextual bandits.
result Human bias and uncertainty in feedback modeling impact the theoretical guarantees of reward learning methods.
Gradient descent struggles to learn equivariant neural networks, even with symmetries.
problem Learning equivariant neural networks via gradient descent is hard.
method Lower bounds for various equivariant neural network classes.
result Gradient descent struggles to learn equivariant neural networks, even with symmetries.
A new method decomposes subjective risk into epistemic and aleatoric uncertainties.
problem Uncertainty quantification in modeling decisions.
method Subjective risk decomposition using strictly proper loss.
result Recovery of classic uncertainty measures and new learning-theoretic connections.
High-dimensional models become unstable when sample size falls below a critical level, leading to a phase transition.
problem Instability in high-dimensional learning models when sample size is insufficient.
method Proved the necessity of a Fisher eigenvalue threshold for stability, introduced Fisher floor for verification.
result A sharp phase transition between reliable concentration and inevitable failure in high-dimensional learning.
We consider distributed statistical optimization in one-shot setting, where there are m machines each observing n i.i.d. samples. Based on its observed samples, each machine sends a B-bit-long message to a server. The server then collects messages from all machines, and estimates a parameter that minimizes an exp…
The paper explores fundamental limits of learning non-hallucinating generative models.
problem Hallucinations in generative models producing invalid outputs.
method Developed a theoretical framework to analyze learnability from a learning-theoretic perspective, incorporating inductive biases.
result Non-hallucinating learning is statistically impossible without additional inductive biases.