Machine Learning More Like a Brain: No Central or Digital Processing

What is it about?

Most computers use a central controller (processor) to perform digital calculations (0's and 1's), and store them in memory. This top-down system makes machine learning inefficient, and is very different from how brains function. There, every neuron performs its own calculations and stores its own memory; that is, learning is a collective, cooperative phenomenon, rather than a centralized algorithm. We refer to this process as "emergent learning" which we recreate in experiment in a (greatly) simplified system. We construct a new kind of flexible system that learns on its own in a collective manner, similar to the brain, but without directly modeling neurons. Rather, we create an analog electronic network whose elements self-adjust in response to being shown training data. A good analogy is a network of pipes with water (current) flowing through them, and the width of these pipes changes in response to water being pumped through the network. Input data is encoded in the pressure of water being pushed into the network at different points, and outputs are pressures read out at other points in the network. (In this analogy, voltage = pressure, pipe width = electronic conductance.) The result is a system whose elements adjusts their own structure in such a way that the system as a whole learns to perform a chosen task. And it does it all without any outside instruction except being presented with training examples. Because the network is analog and self-adjusting, it is already quite power efficient. Because it is built with standard electronic components, it is scaleable. Our hope is that networks like these will teach us about collective learning phenomenon in brains and provide energy-efficient machine learning at scale.

Featured Image

Photo by JJ Ying on Unsplash

Photo by JJ Ying on Unsplash

Why is it important?

Understanding the brain is one of the main challenges of modern science, and AI energy costs are doubling every 4 months. Our system provides a (much) simpler example of "emergent learning" -- learning without centralized control -- present in biology but not in standard computers, and therefore a new angle on understanding the brain. Our system is also scaleable and energy-efficient, making it a potentially useful platform for sustainable AI.