Data-driven ANN-based visual decoding enables unsupervised functional alignment.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning in Neuroscience

Background:

• Artificial neural networks (ANNs) provide a data-driven method for unsupervised discovery of brain functions.
• Understanding visual processing in the brain remains a key challenge in neuroscience.

Purpose of the Study:

• To demonstrate that an ANN trained for visual decoding can spontaneously align with canonical cortical visual functions.
• To investigate the ability of ANNs to learn high-dimensional neural representations for reliable decoding.
• To explore the reciprocal relationship between neural encoding and decoding using ANNs.

Main Methods:

• Training an ANN to decode visual stimuli from multi-unit spiking activity in monkeys.
• Analyzing the ANN's identified brain regions for functional alignment with known visual processing areas (shape, color, motion).
• Inverting the ANN architecture to predict neural activity from visual input.

Main Results:

• The ANN successfully reconstructed complex visual scenes and identified brain regions involved in shape, color, and motion processing without explicit functional priors.
• Despite low train-test correlation at the recording-site level, the ANN learned task-relevant representations at a high-dimensional population level, achieving reliable decoding.
• Inverting the network demonstrated a reciprocal relationship between encoding and decoding of neural activity.

Conclusions:

• ANN-based visual decoding is a powerful framework for unsupervised functional alignment in neural systems.
• This approach can reveal underlying neural representations and functional organization without pre-defined region-specific information.
• The findings highlight the potential of ANNs to bridge the gap between neural activity and cognitive function.