🏢 Harvard University

Recommender systems, when influenced by both users and creators, inevitably polarize; however, prioritizing efficiency through methods like top-k truncation can surprisingly enhance diversity.

Unrolled neural networks, trained via gradient descent, provably achieve optimal Bayesian inference for compressed sensing, surpassing prior-aware counterparts.

UniTS: one model to rule them all! This unified multi-task time series model excels in forecasting, classification, anomaly detection, and imputation, outperforming specialized models across 38 divers…

Stable deep learning on graphs achieved using novel unitary group convolutions, preventing over-smoothing and enhancing model robustness.

Researchers achieve a balance between accuracy and efficiency in multiclass classification by introducing partially consistent surrogate losses and novel methods.

Transformers learn to perform in-context learning of Markov chains hierarchically, progressing from simpler unigram strategies to more complex bigram solutions, with the presence of simpler solutions …

This paper presents nearly-linear time algorithms for testing model calibration, improving upon existing methods and providing theoretical lower bounds for various calibration measures.

Weighted reservoir sampling stabilizes online learning algorithms by creating a robust ensemble of intermediate solutions, significantly improving accuracy and mitigating sensitivity to outliers.

SocialGPT cleverly leverages Vision Foundation Models and Large Language Models for zero-shot social relation reasoning, achieving competitive results and offering interpretable outputs via prompt opt…

SkipPredict optimizes scheduling by prioritizing cheap predictions and using expensive ones only when necessary, achieving cost-effective performance.

DASH: a novel pruning framework leverages domain knowledge to improve the interpretability and sparsity of neural network models for gene regulatory dynamics, outperforming existing methods.

Partially observing neural circuits during experiments can create misleading models, even if single neuron activity matches; researchers need better validation methods.

Set-Based Prompting guarantees order-independent LLM outputs by modifying input representations, eliminating unwanted inconsistencies without fine-tuning.

Optimal ablation (OA) improves model interpretability by precisely measuring component importance, outperforming existing methods. OA-based importance shines in circuit discovery, factual recall, and …

A novel diffusion model reconstructs multimodal shape distributions from shading, mirroring human multistable perception.

Multi-group Proportional Representation (MPR) tackles skewed search results by measuring representation across intersectional groups, improving fairness in image retrieval.

SpLiCE unlocks CLIP’s potential by transforming its dense, opaque representations into sparse, human-interpretable concept embeddings.

Researchers reveal how the training dynamics of transformer models behave at infinite width, depth, and head count, providing key insights for scaling up these models.

Online fair division algorithm achieves Õ(T²/³) regret while guaranteeing envy-freeness or proportionality in expectation, a result proven tight.

Neural network learnability under the manifold hypothesis is hard except for efficiently sampleable manifolds.