Main Menu (Mobile)- Block

Main Menu - Block

New computational model could help shed light on how we see

custom | custom
custom | custom

07/02/25 | New computational model could help shed light on how we see

node:body | entity_field

 

 

To build a simplified model of the primary visual cortex, Janelia researchers first recorded neural activity in more than 29,000 neurons in the primary visual cortex of a mouse as it viewed up to 65,000 images of natural textures, like leaves and rocks.This video shows the presented natural images at the top and the raw calcium imaging data recorded across four imaging depths at a rate of 30 Hz at the bottom. The video is shown at 0.5x speed. Credit: Du et al.

 

Neuroscientists want to understand how individual neurons encode information that allows us to distinguish objects, like telling a leaf apart from a rock. But they have struggled to build computational models that are simple enough to allow them to understand what individual neurons are doing.

To address this challenge, researchers in the Stringer and Pachitariu labs at Janelia set out to create a simpler model to explain what’s going on in the primary visual cortex – the first stop in the brain for visual data.

“We are trying to build a model that can predict the visual responses of each individual neuron,” says Fengtong Du, a graduate student in the Stringer Lab who led the new research.

Determining what happens in individual neurons in the visual cortex is an important first step in understanding visual processing and could help researchers figure out how other parts of the brain are carrying out more complicated computations.

“If you think about your visual system, we’re processing all this information all the time, and there’s all these really complicated visual computations going on all the time, and it all has to be built from this core set of neurons from the primary visual cortex,” says Janelia Group Leader Carsen Stringer. “It’s a ton of neurons, it’s this really big set of features, that then all these other brain areas could use for computation. So if we have a better handle on that, we can understand how all these complicated visual computations are implemented.”

Building a simplified model

simplified model minimodel _fig3 (450 x 563 px).pngTo build their simplified model, the team first recorded neural activity in more than 29,000 neurons in the primary visual cortex of a mouse as it viewed up to 65,000 images of natural textures, like leaves and rocks. Then, they tested different models to find the simplest one that could reproduce the visual information.

They homed in on one model that could reproduce 75 percent of the visual information – a big improvement over previous models that reproduced about 50 percent of the information. The new model also achieved this high level of performance with fewer convolutional layers.

These layers act like filters, allowing the model to extract features that it puts together to detect an image. As the number of layers increases, the features become more abstract. Additional layers make the model better at extracting information and parsing out visual features, but that also makes it more difficult to know what the model is doing and what features it is using.

By making their layers wider and increasing the receptive size of each artificial neuron, the team found they could achieve the same high performance as larger four-layer models with only two layers: a small first layer and a second larger layer. They discovered that all the neurons in the network could share features from the smaller first layer, then individual neurons could combine those features with additional features gleaned from the larger second layer.

This allowed the team to create “minimodels” for each individual neuron, where the weights or combination of features are specific for each individual neuron. Overall, they found that these single neuron “minimodels” are just as powerful as large models, giving researchers an accurate and interpretable way to study visual computation.

“We found the simplest model that can achieve state-of-the-art performance, and we can use a ‘minimodel’ that’s trained on individual neurons to explain the visual feature selectivity in single neurons,” Du says.

###

Citation:

Fengtong Du, Miguel Angel Núñez-Ochoa, Marius Pachitariu & Carsen Stringer. “A simplified minimodel of visual cortical neurons.” PMID: 40593666