The Dynamic Sensorium competition for predicting large-scale mouse visual cortex activity from videos.

Area of Science:

• Computational neuroscience
• Systems neuroscience
• Artificial intelligence

Background:

• Modeling biological visual systems is complex due to nonlinearities between neuronal responses and high-dimensional input.
• Artificial neural networks aid understanding by bridging biological and machine vision.
• Existing benchmarks focus on static images, but animals process dynamic environments.

Purpose of the Study:

• Establish a standardized benchmark for dynamic neural models of the mouse visual system.
• Introduce the Sensorium 2023 Benchmark Competition with dynamic input.
• Facilitate research into brain function under dynamic conditions.

Main Methods:

• Collected a large-scale dataset of neuronal responses from mouse primary visual cortex to dynamic stimuli.
• Organized a competition with main and bonus tracks for predictive modeling of neuronal responses.
• Provided video stimuli, behavioral data, code, tutorials, and baseline models.

Main Results:

• The competition dataset includes responses from over 78,000 neurons to over 2 hours of dynamic stimuli per neuron.
• Participants will develop models for dynamic input (video) and evaluate performance on out-of-domain data.
• The benchmark aims to measure progress in large-scale neural system identification.

Conclusions:

• The Sensorium 2023 Benchmark Competition addresses the need for dynamic input models.
• This initiative will advance the understanding of neural processing in dynamic environments.
• It aims to establish a standard tool for measuring progress in visual system modeling.