Even as Agentic RAG evolves, the quality of the underlying retrieval embedding model remains crucial, and this article compares three training losses—pairwise cosine embedding, triplet margin, and InfoNCE—detailing their inputs, formulas, and practical trade‑offs.
This article explains CLIP’s dual‑encoder architecture—using a Vision Transformer for images and a Transformer for text—how both encoders map inputs into a shared embedding space, the role of cosine similarity, and the InfoNCE contrastive loss that drives joint visual‑language learning.
This article systematically introduces the most common contrastive learning loss functions—including Contrastive Loss, Triplet Loss, N‑pair Loss, InfoNCE, and Cross‑Entropy—explaining their mathematical formulations, advantages, challenges, and typical applications in visual, textual, and multimodal representation learning.
This article reviews cutting‑edge research on integrating large language models with graph‑based recommendation systems, detailing four key strategies—LLM node embeddings, deep graph‑LLM fusion, model‑driven graph data training, and text‑modal enhancements—while analyzing representation learning, InfoNCE optimization, explainable recommendations, and extensive experimental validation.
This article reviews the Kuaishou paper on adapting large language models for recommendation, detailing the LEARN framework's dual‑tower architecture, embedding generation, loss functions, and experimental results that address cold‑start and long‑tail challenges in modern recommender systems.
The article outlines a comprehensive knowledge‑base retrieval matching solution—combining PageRank‑enhanced DSL rewriting, keyword and dual‑tower vector recall, contrastive fine‑ranking, and optimized vector‑based ranking—implemented via offline DP training and Sunfish online inference on Milvus, with applications in enterprise search and recommendations and future plans for graph‑neural embeddings.
By leveraging contrastive learning and large‑scale click‑through data, the article details a dual‑tower model that encodes product titles and queries, explains loss functions, batch‑negative sampling, distributed training tricks, and demonstrates how this approach outperforms traditional NER for product term and category prediction.
This talk explains contrastive learning fundamentals, typical image‑domain models such as SimCLR, MoCo and SwAV, and shows how their principles—positive/negative sample construction, encoder design, loss functions, alignment and uniformity—can be applied to improve dual‑tower retrieval and ranking models, embedding normalization, temperature scaling, and graph‑based recommender systems.
