Search Papers | Poster Sessions | All Posters
Poster B94 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
The Effect of Attention on Contrast Response Functions in Convolutional Neural Networks
Sudhanshu Srivastava1 (), Miguel P Eckstein1; 1UC Santa Barbara
Covert attention results in contrast-dependent influences on neuronal activity with three distinct signatures (Carrasco, 2011): response gain, contrast gain, and baseline shift. Convolutional Neural Networks (CNNs) have recently been used to model covert attention tasks, capturing some known results from psychology and neurophysiology (Srivastava et al., 2024b). This work uses CNNs to understand emergent attention-related neuronal Contrast Response Functions (CRFs). We trained 10 CNNs on the Posner cueing task to optimize target detection with a central cue pointing to the likely target location, with varying contrasts of the display elements. With no explicit attention mechanisms built in, the networks show a behavioral cueing effect and all three gain types emerged in the deeper layer neurons of the networks: Response Gain (17.4%), Contrast Gain (2.3%), and Baseline Shift (22.8%). Using ROC analysis for each neuron, we assessed whether different gain types are associated with different target and cue sensitivities. Response gain neurons had the highest target sensitivity and the lowest cue sensitivity. Baseline shift neurons had the highest cue sensitivity and the lowest target sensitivity. Contrast gain neurons had target/cue intermediate range sensitivities. Together, we show that the diversity of neuronal gain types reported in the literature might arise as an emergent property of task optimization and neurons with different CRFs neurons might be associated with representations of the predictive cue and target.
Keywords: Convolutional Neural Network Covert Attention Contrast Response Functions