New 'horizon loss' method improves classifier accuracy over cross-entropy

A new research paper introduces the "horizon loss" as an alternative to cross-entropy for training classifiers, particularly in the context of reinforcement learning and large language models. This method aims to improve accuracy by considering the long-term impact of learning updates, rather than just immediate gains. Experiments on MNIST and ImageNet datasets using various architectures like ResNet and ViT demonstrated improved top-1 accuracy over standard cross-entropy, with gains increasing in the presence of noisy labels. AI

IMPACT Introduces a new training objective that could enhance classifier performance and potentially impact LLM post-training techniques.

RANK_REASON The cluster contains a new academic paper detailing a novel machine learning method. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Classification
• ImageNet
• MNIST database
• Policy Gradient Methods for Reinforcement Learning with Function Approximation
• reinforcement learning
• ResNet-101
• ResNet-50
• ViT-S/16

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →