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

## TL;DR

A neural network model successfully decodes and reconstructs visual stimuli from brain activity in monkeys, aligning with known visual functions without supervision.

## Contribution

The study introduces a novel data-driven ANN framework for unsupervised functional alignment of visual processing in neural systems.

## Key findings

- The ANN reconstructs complex visual scenes and aligns with canonical cortical functions for shape, color, and motion.
- The model achieves reliable decoding despite low train-test dataset correlation at the recording-site level.
- Inverting the network architecture reveals a reciprocal relationship between encoding and decoding processes.

## Abstract

Artificial neural networks (ANNs) offer a data-driven approach to reveal brain regional functions without explicit supervision. Here, we demonstrate that an ANN trained to decode visual stimuli from multi-unit spiking activity in monkeys, can not only reconstruct complex and dynamic scenes, but also spontaneously align with canonical cortical visual functions. Without any region-specific functional priors, the model identifies key brain areas associated with shape, color, and motion processing. We provide strong evidence that, despite low train-test dataset correlation at the recording-site level, the ANN-based model is able to learn task-relevant representations embedded at a high-dimensional population level and achieve reliable decoding performance. Furthermore, by inverting the architecture and retraining, we show that the same network can predict region-specific spiking patterns from visual input, indicating a reciprocal relationship between encoding and decoding. These findings shed light on ANN-based visual decoding as a powerful framework for unsupervised functional alignment in neural systems.

A data-driven artificial neural network model decodes dynamic visual stimuli from monkey multi-unit spiking activity, reconstructing shapes, colors, and motion with high fidelity. It spontaneously aligns with canonical cortical visual functions without region priors, and its inverse architecture predicts neural spiking from visuals, revealing reciprocal encoding-decoding in neural systems.