Retrospective for the Dynamic Sensorium Competition for predicting large-scale mouse primary visual cortex activity...

Area of Science:

• Computational Neuroscience
• Machine Vision
• Artificial Neural Networks

Background:

• Biological visual system processing is complex due to nonlinear input-response relationships.
• Artificial neural networks offer tools for modeling biological and machine vision.
• Lack of standardized benchmarks hinders progress in dynamic models of the mouse visual system.

Purpose of the Study:

• To establish a standardized benchmark for evaluating dynamic models of the mouse visual system.
• To create a large-scale dataset of neuronal responses and behavior from the mouse primary visual cortex.
• To foster advancements in predictive modeling for neuroscience.

Main Methods:

• Established the SENSORIUM 2023 Benchmark Competition with dynamic input.
• Collected a large dataset from 78,853 mouse neurons over 2 hours, including behavioral data.
• Ranked over 160 model submissions from 22 teams on in-domain and out-of-distribution prediction tasks.

Main Results:

• The competition successfully ranked dynamic models based on predictive performance.
• New neural network architectures were proposed, improving state-of-the-art by 50%.
• Models were evaluated on their ability to generalize to novel stimuli.

Conclusions:

• The SENSORIUM benchmark provides a vital resource for advancing computational models of the visual system.
• The competition spurred innovation in neural network architectures for neuroscience.
• The dataset and infrastructure are available for continued research.