TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Show HN: Drop-In Out-of-Distribution Data Detector

4 点作者 barthelomew5 个月前
Out-of-Distribution (OOD) detection is possibly the most important problem for safe and deployable ML, because it provides the first line of defense by preventing silent failures in critical ML systems; bounds AI capabilities by recognition of model knowledge; allows for safe fallback and enables human oversight when needed.<p>Forte takes a novel approach to OOD detection with several key advantages. It utilizes self-supervised representations to capture semantic features while incorporating manifold estimation to account for local topology. Care has been taken to minimize deployment overhead; by eliminating additional model training requirements. Furthermore, it requires no class labels, no exposure to OOD data during training, and no restrictions to architecture of predictive or generative models. We have demonstrated strong domain generalization – having tested it on detecting synthetic data, MRI images etc.

暂无评论

暂无评论