Emergent human-like covert attention in feedforward convolutional neural networks

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Covert attention enables visual selection without eye movements, with benefits often explained by limited resources.
• Existing theories of attention are challenging to reconcile with neuronal population activity.
• Understanding the neural basis of attention is crucial for explaining perception and cognition.

Purpose of the Study:

• To investigate whether human-like covert attention behaviors can emerge in a feedforward convolutional neural network (CNN) without explicit attention mechanisms.
• To determine if optimizing for target detection accuracy alone can replicate attentional cueing and contextual effects.
• To explore the implications of these findings for theories of attention and its presence in simpler organisms.

Main Methods:

• A feedforward CNN was trained on image datasets to optimize target detection accuracy.
• The CNN was evaluated on its ability to utilize predictive cues and contexts in standard covert attention tasks (Posner cueing, set size, contextual cueing).
• CNN performance was compared to human accuracy data and a Bayesian ideal observer model, with analyses across different training schemes and cue types.

Main Results:

• The CNN successfully learned to utilize cues and contexts, mirroring human accuracy patterns in covert attention tasks.
• These cueing and contextual effects were robust, generalizing across training variations, cue types, and task measures.
• The CNN's effects persisted even when network resources were reduced and generalized to novel cue instances.

Conclusions:

• Human-like behavioral signatures of covert attention may be an emergent property of optimizing task accuracy in neuronal populations.
• Limited attentional resources may not be necessary to explain these attention phenomena.
• The findings provide a potential explanation for attention-like effects observed in organisms lacking a neocortex.