Measuring Human Perception to Improve Open Set Recognition

TL;DR

This paper leverages human reaction time data from psychophysical experiments to develop a new loss function, significantly enhancing open set recognition performance in deep neural networks, especially with limited labeled data.

Contribution

Introduces a psychophysical loss function based on human reaction times, improving open set recognition in deep networks with limited training data.

Findings

Achieved a 6.02% increase in top-1 validation accuracy.

Improved top-1 test accuracy on known samples by 9.81%.

Enhanced top-1 test accuracy on unknown samples by 33.18%.

Abstract

The human ability to recognize when an object belongs or does not belong to a particular vision task outperforms all open set recognition algorithms. Human perception as measured by the methods and procedures of visual psychophysics from psychology provides an additional data stream for algorithms that need to manage novelty. For instance, measured reaction time from human subjects can offer insight as to whether a class sample is prone to be confused with a different class -- known or novel. In this work, we designed and performed a large-scale behavioral experiment that collected over 200,000 human reaction time measurements associated with object recognition. The data collected indicated reaction time varies meaningfully across objects at the sample-level. We therefore designed a new psychophysical loss function that enforces consistency with human behavior in deep networks which exhibit variable reaction time for different images. As in biological vision, this approach allows us to achieve good open set recognition performance in regimes with limited labeled training data. Through experiments using data from ImageNet, significant improvement is observed when training Multi-Scale DenseNets with this new formulation: it significantly improved top-1 validation accuracy by 6.02%, top-1 test accuracy on known samples by 9.81%, and top-1 test accuracy on unknown samples by 33.18%. We compared our method to 10 open set recognition methods from the literature, which were all outperformed on multiple metrics.