An Artificial Neural Network Model of Autism: Understanding the Causes

Overview

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that impairs the ability to communicate and interact. Artificial neural networks (ANNs) can serve as simple models of human brain, allowing us to understand how alterations in brain structures might result in the emergence of certain autistic behavioral traits. Reduced generalization has been observed in individuals with ASD for decades. Recent neuropathological studies suggest that affected individuals have too many brain neurons and synapses. With this knowledge, I hypothesized that an ANN with a high number of neurons simulates reduced generalization seen in people with autism. I also hypothesized that pruning the network would improve generalization. To test my hypothesis, I built an under-complete stacked autoencoder that learnt important features of input images it has been trained on using compression and reconstructed lossy images as the output. Training data consisted of five unfragmented pictures from the Fragmented Picture-Completion task used by Happe et al. (2016) for testing global processing deficits in individuals with ASD. Fragmented pictures were used for testing the model. As the number of neurons in the network increased, image reconstruction loss increased, suggesting that high neuron density accounts for certain autistic traits such as reduced generalization. Pruning the network through regularization induced sparsity, and reduced image reconstruction loss, proving my hypothesis. The results of my experiment show that pruning extra brain synapses could potentially reverse symptoms of autism. Overall, a computational perspective of ASD can help us understand the causes and, thereby, create possible courses of treatment for the disease.

Competition history

• AJAS 2020 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• AJAS None projects
• Every AJAS 2020 project