Rohit Singh·3d agoBeyond Fine-Tuning: Building High-Performance State-Action Embeddings with a Frozen LLM BackboneIn this article, we train two small projection heads on top of a frozen LLM, using contrastive learning, to pick the best action for the…
Ankith Reddy Subhanpuram·Nov 2From Raw Text to the Unit Hypersphere: Building and Visualizing a Modern Text Embedding Model from…A complete guide to engineering a dense text embedding pipeline in PyTorch — covering BPE tokenization, Unsupervised SimCSE, InfoNCE…
Vahid Naghshin·Oct 26Seeing Without Being Told: How Machines Learn to Understand ImagesA deep dive into the ingredients that make self-supervised visual representation learning work — and why they matter.
InTowards AIbyPop Alexandru·Oct 10The Strange Reason AI Needs Wrong Answers to LearnContrastive learning explained from first principles — how neural networks discover meaningful representations through similarity…
Ioannis Petrousov·Sep 1Fine-tuning models with Supervised Contrastive Learning (SupCon)Speech Emotion Recognition (SER) is a notoriously difficult challenge. When working with small datasets like RAVDESS (roughly 1200 training…
Arun Dobriyal·Aug 29Contrastive Learning, InfoNCE, and Why Two-Tower Systems Train Their Own EmbeddingsEvery search or recommendation system hits the same wall eventually. You’ve got a query, ten million documents, and you need the top twenty…
Kelvin Teo Wei Min·Aug 21An algorithm for contrastive learning which does away with the use of Euclidean distance and cosine…Abstract
ComfyAI·Aug 15CLIP’s Image Embeddings Were Never “Uncalibrated”Over the past couple of years, a story has quietly become common knowledge: image-text contrastive learning like CLIP only pulls images and…
Outermostkt·Jul 3Explaining the Paper ‘LeVLJEPA: End-to-End Vision-Language Pretraining Without Negatives’Paper: LeVLJEPA: End-to-End Vision-Language Pretraining Without Negatives Authors: Lukas Kuhn, Giuseppe Serra, Randall Balestriero, Florian…
