This book aims to demystify large deep networks and generative models, often perceived as "black boxes," by exploring their internal mechanisms through the lens of representation learning. It outlines how modern neural network architectures are designed, utilizing optimization and information theory.
This TMLR paper introduces operational variational semantics to joint-embedding architectures for non-contrastive representation learning. It does so by factorizing embedding likelihood, anchoring posterior uncertainty to likelihood scale, and employing a heavy-tailed Student-t likelihood for empirical benefits.
This research focuses on Subject-Aware Contrastive Learning, a novel AI technique developed for the effective processing and understanding of biosignals. It aims to improve the representation learning of complex biological data, offering advancements in the analysis of physiological measurements.
Autoencoders are neural networks that compress data into a lower-dimensional space and reconstruct the original input, learning non-linear structures unlike linear PCA. Their two-stage design features an encoder that projects input data into a latent space to extract informative features.
The author explores federated learning to overcome latency challenges in voluminous sensor data from multi-robotic autonomous vehicles, optimizing processing in low-bandwidth environments. This approach seeks a distributed alternative to centralized data synchronization through distributed model updates.
This research paper demonstrates that embedded numeric anchors on images systematically bias Vision-Language Model quality judgments across multiple VLMs. Layer-wise probing reveals that optimal layers for quality prediction are deeper than where anchor classification saturates, establishing a causal account of visual anchoring bias.
This content explores methods for anticipating visual representations from unlabeled video. The research investigates models' ability to learn visual features without explicit supervision, enhancing contextual understanding in video sequences.
This work defines Bayes-sufficient representations for supervised learning, focusing on information relevant for prediction based on a fixed decision problem and loss function. It introduces the concept of a Bayes quotient and connects the framework to property elicitation, showing how different loss functions require specific Bayes-optimal actions.
This work introduces SetFlow, a generative architecture that models entire Multiple Instance Learning (MIL) bags directly in the representation space. It leverages the flow matching paradigm and a Set Transformer-inspired design to capture intra-bag dependencies and generate coherent, semantically consistent representations.
This content delves into the concept of universal representations in AI, proposing them as the missing link to connect and process extremely diverse data types. The goal is to establish a unified framework capable of understanding everything from human faces and text to planktons and cat breeds.
This paper introduces Soft Silhouette Loss, a novel differentiable objective for deep learning, inspired by the classical silhouette coefficient. It aims to learn discriminative representations by enforcing intra-class compactness and inter-class separation more efficiently than existing metric learning approaches.
This manuscript introduces Data Driven Variational Basis Learning (DVBL), a novel non-neural framework for learning data-adaptive basis functions directly from high-dimensional data. It provides an explicit, interpretable, and mathematically transparent alternative to neural networks for representation learning, addressing their limitations in control and transparency.
This paper introduces a distillation framework to make large genomic foundation models for mRNA representation learning more efficient, reducing model size by 200-fold. By using embedding-level distillation, the smaller model achieves state-of-the-art performance on mRNA-related tasks, demonstrating an effective strategy for scalable biological AI.
O conteúdo descreve o DFR-Gemma, um novo framework que permite que LLMs raciocinem diretamente sobre embeddings geoespaciais densos. Ele alinha embeddings de alta dimensão com o espaço latente de um LLM através de um projetor leve, injetando-os como tokens semânticos.
This paper introduces H-probes, linear probes designed to extract hierarchical structure, specifically depth and pairwise distance, from the latent representations of large language models. The research shows these probes robustly find low-dimensional subspaces crucial for performance in synthetic tree traversal tasks, generalizing well both within and out-of-domain.
This research investigates the 'grokking' phenomenon in transformers, finding that the long delay to generalization in arithmetic models stems from a decoder bottleneck. The encoder acquires relevant structural knowledge early, but the decoder struggles to access it, a hypothesis supported by causal interventions like transplanting encoders.
Contrastive learning aims to preserve relational structure in sample representations by reflecting a similarity graph. This paper interprets weighted InfoNCE objectives as Distance Geometry Problems, providing a unified geometric framework and exact characterizations of optimal embeddings, revealing how class imbalance affects inter-class similarities in SupCon.
This research explores unsupervised categorization of transformations between input pairs using algebraic constraints, aiming for a principled understanding of good representations. It introduces parameter division to refine prior Galois-theoretic methods, addressing their reliance on auxiliary assumptions and improving the decomposition of groups.
This research explores how world models learn semantic representations from physical exploration without linguistic input. It finds that the latent space develops spatial semantic structures mirroring physical geometry, with semantic alignment improving alongside prediction performance.
This paper presents a survey addressing multimodal learning challenges with the Mixture-of-Experts (MoE) architecture. The study explores how MoE functions as an efficient engine and a representation learner for integrating diverse data modalities. It fills a gap in the literature by offering a comprehensive and systematic review on the topic.