Raman spectroscopy in open world learning settings using the Objectosphere approach

Abstraction: Objectosphere loss reduces false positives for unknown Raman spectra classes

Key points:

• Standard ML classifiers on Raman spectra produce high false-positive rates when encountering substances not in the training set, limiting clinical utility
• The paper applies Entropic Open Set and Objectosphere loss functions to enable the network to identify unknown classes while maintaining high accuracy on known classes
• Database: 40 chemical classes split into 20 biologically relevant (amino acids), 10 irrelevant bio-related chemicals, and 10 unseen classes for open-set evaluation
• Approach dramatically reduces false positives on novel/unknown substances while preserving accuracy on the 20 known biological classes
• Bridges gap between controlled lab experiments and clinical deployment where novel substances will inevitably appear at inference time

Connections: Arxiv · Open Set Recognition · Machine Learning · Deep Learning

Source: https://arxiv.org/abs/2111.06268