Olfaction-Inspired Machine Learning

A New Inspiration for Artificial Intelligence Most artificial neural networks process information in ways based loosely on the visual system. Now some scientists are hoping to address shortcomings in AI by turning to another neurological source for insight: olfaction. Vision-Inspired Machine Learning This deep network, which consists of many layers of neurons, involves a hierarchical extraction of features. Neurons in lower levels detect specific details and send that information to the next level. With each layer, the representation of inputs gets increasingly abstract. The AI system “learns” by adjusting the connections between its neurons to improve the accuracy of its final output. INPUT: Portrait is broken into pixels LAYER 1 NEURONS IDENTIFY Pixel values L2 Edges L3 Edge combinations L4 Features L5 Feature combinations OUTPUT: “Jane” Olfaction-Inspired Machine Learning In this “shallow” three-layer network, a relatively small number of neurons initially respond to the molecules in an odor. That information gets projected to a much larger number of cells, each randomly connected to several neurons in the previous layer. All but the most active 5 percent of the neurons then get silenced, leaving the brain with a “tag” to identify the odor. INPUT: Odor consisting of many different molecules LAYER 1 NEURONS Detect molecules L2 Receive signals and pass them on L3 Respond to and encode odor OUTPUT: “Skunk” tag Most active (top 5%)

An educational infographic compares vision-inspired machine learning with olfaction-inspired machine learning using a portrait and a skunk odor. The joke comes from the absurd contrast between the complex neural-network explanation and the blunt final output identifying the smell as “Skunk,” making it useful as a reaction to overcomplicated explanations with an obvious conclusion.

• @profunctor_io / 6407 2021-03-07