Research from Renmin University: Caution Advised in Data Augmentation for Contrastive Learning

The latest research from Renmin University warns that caution is needed in data augmentation for contrastive learning. Strongly aligned positive samples may harm the generalization ability of contrastive learning. While stronger data augmentation can enhance downstream task performance, alignment performance decreases. The study reveals the mechanisms by which data augmentation affects contrastive learning and proposes seeking better data augmentation strategies from an information-theoretic and spectral perspective.

Tencent Releases New Patent on Training Large Language Models to Enhance Generalization and Accuracy

Recently, Tencent Technology (Shenzhen) Co., Ltd. published a patent regarding a training method and related equipment for large language models on the Tianyancha app. The patent is titled 'Training Method, Device, Computer Equipment, and Storage Medium for Large Language Models' and aims to enhance the learning capacity and accuracy of large language models through innovative training methods. In the training process of large language models, traditional methods often rely on a single text summary, which may lead to model overfitting and negatively impact the accuracy and diversity of generated content. However, Tencent's new...

Microsoft Launches LLM2CLIP: New AI Technology Supports Image Understanding with Language Models

In today's technology landscape, CLIP (Contrastive Language-Image Pre-training) is an important multimodal foundational model. It combines visual signals and text signals into a shared feature space using contrastive learning loss on a large-scale dataset of image-text pairs. As a retriever, CLIP supports various tasks such as zero-shot classification, detection, segmentation, and image-text retrieval. Meanwhile, as a feature extractor, it performs well in nearly all...

Small yet Beautiful! HKU's Latest Recommendation System EasyRec Insights User Voices through Text

EasyRec is a recommendation system based on language models, developed by a team from the University of Hong Kong. Its uniqueness lies in analyzing emotional and detailed user behavior stories through a text behavior alignment framework to predict user preferences without requiring large amounts of user data. The system combines contrastive learning and collaborative language models, enabling accurate predictions of preferences for new users and new products, particularly excelling in zero-shot recommendation scenarios. EasyRec's plug-and-play features make it easy to integrate into existing recommendation systems, enhancing performance. The paper showcases EasyRec's performance across multiple...

Behind the Battle of Trillion-Level Computing Power: OpenAI Astra Has Burned Hundreds of Billions of Dollars, Where Are the Gaps Between Domestic Large Models and American Companies?

Recently, Anthropic released Fable5.1, and OpenAI launched Astra, further intensifying the global competition in large models. Although Astra has sparked discussions about its performance and cost-effectiveness, the massive financial and hardware consumption behind it highlights the gap in computing infrastructure between China and the United States. Reports indicate that OpenAI has not disclosed specific parameters, but it has used the StarGate program to train models utilizing over 100,000 NVIDIA GPUs.

AI Short Film Production Prices Drop Sharply: Per-Minute Quotes Fall from 5,000 to Hundreds, Industry Shifts to Content Competition

This year, AI short films have exploded in popularity. The reduction in technical barriers has led to a surge in practitioners and expanded production capacity, causing a sharp drop in production prices: per-minute quotes have fallen from 5,000 yuan to hundreds of yuan, approaching the cost line; custom short films still maintain per-minute prices of 10,000 to 20,000 yuan. Industry competition is shifting from focusing on output and price to focusing on content and market, with content creativity and operational capabilities becoming new barriers.