Representation Learning

Novel metric space magnitude measures rigorously quantify the diversity of latent representations across multiple scales, showing superior performance in detecting mode collapse and characterizing emb…

New method efficiently measures how well AI models memorize training data, revealing that open-source models memorize less than expected.

MatrixNet learns efficient group representations for improved deep learning on symmetry groups, achieving higher sample efficiency and generalization than existing methods.

This study marries causal representation learning with dynamical systems to enable parameter identification in real-world scientific data, unlocking downstream causal analysis for various applications…

LMSPS: a novel framework efficiently leverages long-range dependencies in large heterogeneous graphs by dynamically identifying effective meta-paths, mitigating computational costs and over-smoothing.

Recurrent Graph Neural Networks (GNNs) with real and floating-point numbers are precisely characterized by rule-based and infinitary modal logics, respectively, enabling a deeper understanding of thei…

GCFormer, a novel graph Transformer, enhances node representation learning by employing a hybrid token generator and contrastive learning, outperforming existing methods on various datasets.

Deterministically shaping in-distribution feature space solves OOD detection’s distributional assumption challenge, leading to superior performance.

HypStructure boosts representation learning by embedding label hierarchies into hyperbolic space, improving accuracy and interpretability.

This research pioneers the integration of machine learning with the dependently-typed programming language Agda, introducing a novel dataset and neural architecture for faithful program representation…

Learn graph representations efficiently by identifying the minimal data needed to uniquely define a graph’s structure, achieving robust performance with fewer resources.

Neural networks learn place cell-like representations and context-dependent remapping using a novel similarity-based objective function, providing insights into hippocampal encoding.

Learning hidden structures from sparse interactions in data is computationally hard but can be achieved with sufficient samples using gradient-based methods; This is shown by analyzing the gradient dy…

LiNGCREL, a novel algorithm, provably recovers linear causal representations from diverse environments, achieving identifiability despite intrinsic ambiguities, thus advancing causal AI.

FCR, a novel method, reveals causal structure in single-cell perturbation data by learning disentangled cellular representations specific to covariates, treatments, and their interactions, outperformi…

Aligning AI’s world representation with humans enables faster, safer learning of human values, improving both exploration and generalization.

AI learns emergent system features from time-series data using a novel differentiable architecture maximizing causal emergence, outperforming pure mutual information maximization.

CoupleNet dynamically links protein sequences and structures for improved representations, surpassing state-of-the-art methods in function prediction, particularly for uncommon proteins.

NeuRes, a novel neuro-symbolic approach, achieves superior SAT solving accuracy using significantly less training data than existing methods by combining certificate-driven learning with expert iterat…

Latent Functional Maps (LFM) offers a novel spectral framework for comparing, aligning, and transferring neural network representations, boosting downstream task performance and interpretability.