FYI.

This story is over 5 years old.

Tech

What We Can Learn from Programming a Computer to Be Autistic

A new neuro-simulation framework offers insights into the physiological roots of autism and maybe even schizophrenia and aging.
Image: Wiki

Neuroscientists from Baylor University have successfully programmed an artificial neural network to exhibit autistic behavior. In doing so, they've shown how a theorized general root of autism—related to a concept known as divisive normalization—in the brain leads to the disorder's specific symptomatic manifestations.

One of the main hypotheses of autism is that the disorder has to do with a neurophysiological excitation-to-inhibition (E/I) imbalance. An autistic brain is overexcited, generally.

Advertisement

The whole E/I system works like this. Neurons in the brain act as gates, constantly receiving excitatory signals from other neurons while only passing them on to the next cell in line if more excitatory signals are reached than inhibitory signals, which are also being transmitted to the cell/gate in question. Inhibitory signals act to raise and lower this gate, making it easier or harder to pass along an excitatory signal. Below a certain threshold, the excitatory signals are withheld. The whole process is pretty neat.

The E/I balancing act also acts as a filter of sorts. As in any electrical or electro-chemical system there is background noise, which, for the brain to properly function, needs to be filtered out of its circuitry. With an inhibitory gate in place, a neuron will only send a message of sufficient strength, which is usually much greater than any background noise that might be present. So, the E/I system ensures that only meaningful signals are transmitted from neuron to neuron. In those suffering from autism, the hypothesis goes, this balance is off.

The result is that the autistic brain is in a state of constant excitation. There is some physical evidence for this in the increased prevalence of seizures, which are strongly linked to over-excitation in the brain, in autistic patients, yet it remains a difficult hypothesis to prove. This difficulty has to do with the lack of clear connections between I/E imbalances and the actual lived symptoms of autism. And this is where artificial neural networks come in as a way of simulating that imbalance and following it to its likely symptomatic expressions.

Advertisement

"Here, we propose that autism symptomatology reflects alterations in neural computation," the Baylor neuroscientists report in the Proceedings of the National Academy of Sciences. "Using neural network simulations, we show that a reduction in the amount of inhibition occurring through a computation called divisive normalization can account for perceptual consequences reported in autism, as well as proposed changes in the extent to which past experience influences the interpretation of current sensory information in individuals with the disorder."

"The simulations further establish a link between divisive normalization and high-level theories about how autism may alter the influence of past experience on the interpretation of current sensory information," they continue.

It's worth briefly expanding on the concept of divisive normalization. It's a computation that takes the activity of an individual neuron and divides it by the average activity of the neurons in its general cluster. Normalization is still a sort of new idea, but is thought to be implicated in signalling activity all over the brain, particularly when it comes to sensory inputs. Basically, it's a simple way of taking a big dump of data and filtering out meaningful or "ideal" messages. If we can determine the average background noise of a given channel, then we can more easily determine which signals are meaningful. This is basic information theory.

"Specifically, autism-like symptomatology arises in the model when the influence of the population on the activity of individual neurons is reduced, in essence making the neurons less 'social,'" the neuroscientists write. Moreover, divisive normalization may be able to account for the overall diversity of autism phenotypes, or how the disorder is actually expressed.

While the simulations conducted by the Baylor group offer some powerful evidence for one (of several) autism hypothesis, the researchers note that it may be the general experimental framework developed here that winds up being the most important result just given the historic difficulty in coming up with falsifiable conclusions related to autism and other disorders.

"The computational framework described in this study provides a formalism for investigating how alterations in neural computation may give rise to autism symptomatology," the Baylor group concludes. "Adaptations of this approach may provide insights into other mental health disorders such as schizophrenia, and perhaps some aspects of aging. The results of our simulations further suggest that behavioral assays combined with computational modeling may be useful in identifying altered physiological pathways in individuals, and thus in facilitating the development of individualized treatment plans."