From reconstructions of neuronal circuits to anatomically realistic artificial neural networks

Introduction

Artificial neural networks (ANNs) have found applications in a wide variety of real-world problems. Despite this tremendous success, artificial intelligence systems still face major challenges due to their reliance on extensive training and large datasets. Recent reports indicate that the architecture of ANNs could be a prime target for reducing their training and data requirements.

Hypothesis

We hypothesize that such architectural features can be identified from neuronal networks in the brain, which have evolved to efficiently perform highly specialized functions. Recent advances in electron microscopy will soon provide detailed reconstructions of large-scale neuronal networks from different brain areas, species, developmental stages, and/or pathological conditions.

Challenges

However, even if such data become available, directly transforming neuronal network reconstructions into ANNs will raise problems of:

1. Interpretability, due to their enormous complexity.
2. Generalizability, due to high inter-individual variability.

Methodology

Here, we will resolve these challenges by implementing a set of computational approaches that allow:

• The extraction of rules that explain the wiring properties underlying dense connectomics data.
• The transfer of these anatomical principles into the design of ANN architectures.
• The evaluation of how these principles impact performance on a battery of deep learning tasks.

This unique methodology will lay the foundation for groundbreaking insights into how different network architectures facilitate specific brain functions, and also how the underlying anatomical principles can inform the development of more effective and efficient artificial intelligence systems.

Accessibility

Our methodology will be publicly accessible online to scientists, but also to companies and non-profit organizations that seek to improve the performance or reduce training data requirements for applications of deep learning.