Open Internet by MindsNet

Neural networks' inability to detect out-of-distribution data

Current neural networks have a fundamental geometry problem, leading to overconfidence and hallucinations when faced with garbage data or out-of-distribution samples. This issue arises from the standard Cross-Entropy loss requiring models to push features far away from the origin, resulting in a jagged latent space. As a result, models lack a mathematically sound place to represent uncertainty or reject unknown inputs.

Computing & Technology, Computer Science, Machine Learning