New PNAS paper using AI to identify new neuron types and further the understanding of covert attention

Covert attention refers to the selection of part of the visual scene without moving our eyes.  Imagine you’re driving on a busy highway. While your eyes are fixed on the car ahead, you’re simultaneously tracking the movement of other vehicles in your peripheral vision. Or consider a social gathering where you discreetly pay attention to someone’s conversation without making direct eye contact.   When attention is directed to a location, people's performance in detecting an object is more accurate and faster.  Traditionally, scientists have studied how 

A recent study from our lab turns this long-held assumption on its head. A simple artificial intelligence (a Convolutional Neural Network, CNN), given a basic task and no instructions about attention, spontaneously developed the same behavioral signatures. The findings suggest that this remarkable cognitive ability may not be so special after all.

1. Attention Isn't as "Advanced" as We Though. It's an Emergent Property

The most counter-intuitive finding is that covert attention appears to be an "emergent form of intelligence." Rather than being a pre-programmed module unique to complex brains, it arises naturally when a system is simply optimized to perform a task more efficiently.

This new perspective is supported by growing evidence from the animal kingdom. The behavioral signs of covert attention have been found in a wide range of animals, including crows, rodents, archer fish, and even bees—creatures that lack the mammalian brain structures, specifically, the parietal lobes in the neocortex, previously thought to be required for this function. This widespread presence hinted that a simpler, more fundamental principle was at play.

2. An AI Taught Itself to "Pay Attention" Without Any Instructions

To test their hypothesis, we trained a standard convolutional neural network (CNN)—a type of AI mode on a straightforward task: detect a faint, tilted line hidden within a noisy image. Most of the time, a visual cue (a small box) appeared in the image, indicating the likely location of the target line.

Crucially, the CNN was never explicitly told what "attention" was, what the cue meant, or that it should prioritize information from the cued location. The model was only trained to get better at one thing: correctly detecting the target.

In the process of optimizing its performance, the network spontaneously developed human-like attentional behaviors. It became significantly more accurate at finding the target when it appeared at the cued location, effectively demonstrating covert attention without ever being programmed for it. Its performance was so effective that it was comparable to a Bayesian ideal observer, a theoretical model that represents the highest possible accuracy for the task, despite the AI having none of the pre-programmed statistical knowledge the ideal observer is given.

3. The AI Predicted New Types of Brain Cells That Actually Exist

Using neuroscience-style analytical tools, we peered inside the 180,000 "neurons" of ten trained CNN (1.8 million units across all networks) to understand the internal mechanisms that emerged in the CNNs.

The CNN predicted the existence of specific neuron types that neurophysiologists had not previously highlighted in attention research. These included cue-inhibitory cells (which diminish their response when a cue is present), location-opponent cells (which are excited by a stimulus in one location but inhibited by one in another), and neurons that pool information across space (summation neurons)

We then went back and re-analyzed previous brain recordings from the superior colliculus in mice (Wang, Herman, and Krauzlis, 2022), performing a similar attention task. We found evidence of these exact neuron types that the CNN had predicted. The model's internal workings provided a guide that revealed previously overlooked mechanisms in a real, biological brain.

4. Why we think this is important

This research pioneers a powerful methodological shift. Instead of treating AI models as inscrutable black-box function approximators, scientists can now use them as analyzable model systems.

The benefit of this approach is immense. A relatively simple, fully observable AI can generate concrete, testable hypotheses about how vastly more complex biological neural circuits might function. The CNN acts as a guide, suggesting what neurophysiologists should look for in the billions of neurons that make up a living brain, accelerating the pace of discovery.

Conclusion: Rethinking Covert Attention

This study reshapes our understanding of a fundamental cognitive function. It suggests that complex abilities like covert attention may not be intricately designed, pre-programmed modules. Instead, they can be the natural, emergent result of a system simply optimizing its ability to process information and perform a task well.  There is much to assess whether this simple form of emergent attention can account for all forms of behavioral attention in humans, but the work pushes the field to rethink what might be the added benefits of a specialized attention network and feedback present in the human brain.