Neural Networks Uncover Previously Unreported Neuron Types Related to Emergent Covert Attention

What is it about?

Covert attention is the ability to select part of the visual world without looking directly at it, for example, monitoring cars in a neighboring lane while driving or attending to someone in a social setting without shifting your gaze. Attention improves detection and identification at the attended location. Traditionally, this process has been viewed as an internal brain mechanism (“attention”) that enhances visual processing at the attended location by increasing neural activity. However, we show that a Convolutional Neural Network (CNN) with no in-built attention mechanism, trained to optimize detection of targets, not only exhibits emergent human-like attentional behaviors, but its inner components show an emergent hierarchical organization of neuronal-like properties paralleling findings in neuroscience. We show how the CNN’s emergent attention behaviors arise from a plurality of neuronal unit mechanisms jointly tuned to targets and the environmental cues predictive of the target. Critically, the CNN shows unit types unreported or underemphasized by neurophysiologists, including location-opponent units that are excited by one location and inhibited by another, and units showing attentional inhibition. Re-analysis of mouse neural data performing an attention task reveals these same neuron types, suggesting they may be part of a plurality of neuron types mediating covert attention.

Featured Image

Photo by Sumaid pal Singh Bakshi on Unsplash

Photo by Sumaid pal Singh Bakshi on Unsplash

Why is it important?

The work is timely and unique in three key ways. First, it shows that not only can attention-like behaviors emerge in CNNs despite the absence of any built-in attention mechanism, but many neuronal properties seen in animal brains also emerge as a consequence of task optimization. These findings may help explain why attention-like behaviors, once thought to be primarily human or primate abilities, have been observed across a wide range of species, including crows, mice, zebrafish, bees, and fruit flies in recent decades. Second, the work illustrates how analyzing the internal workings of an AI system can reveal neuron properties that neuroscientists may never have hypothesized or explicitly searched for because CNNs make it possible to analyze every single neuron, a scale hard to achieve in the brain. It provides an example of how AI can guide neuroscience by suggesting new neuron types to look for and by complementing hypotheses generated from traditional psychological and neurobiological theory. Third, this paper brings together human behavioral data, mice neuronal data, Bayesian models, and neural networks, a rare combination of techniques in one paper.

Perspectives