Preliminary ChipAI sample
There is currently a seemingly endless amount of hype surrounding the AI or Machine Learning fields. While there have undoubtedly been several recent developments that may warrant the hype, they largely require expensive computational and electrical processes to achieve their results. According to a recent paper by researchers at the University of Massachusetts, training high quality AI models can have a surprisingly heavy environmental impact.
The research group with whom I’ve had the opportunity to work and learn from these past months has been developing a possible answer to some of these concerns with their ChipAI project. In a brief and simplistic summary, one of ChipAI's goals is to leverage photonics at a nanoscale in order to overcome some of the costly power requirements of these systems. What does that mean for the future? One scenario might be that rather than your phone reaching out to the cloud to query neural networks on a server somewhere, these operations might conceivably be able to happen locally, meaning faster results, lower consumption and less network traffic.
At a very basic level, the 'way' this will happen is with LEDs similar to those with which we are familiar - but miniscule - arranged in a grid and only emitting a few photons each. These LEDs can communicate with one another as a spiking neural network (SNN). These networks are closest to emulating what actually happens in a brain, and it is what most excites me about working within this paradigm.
I have spent my time here trying to further my understanding about the idea of neuromorphic computing, and what steps are being taken to emulate these processes on a tiny chip. My initial research into the history of AI introduced me to Frank Rosenblatt, who built an early system at Cornell University, called the Mark 1 Perceptron. Rosenblatt, however, did not set out to initially to build an artificial intelligence computer, he was simply interested in the functioning of the eyes of the common housefly and he set out to build an electronic version of those mechanics. This interfacing of the natural world and technological world is where I find the inspiration for my work; nature is, in fact, the highest technology, not silicon transistors.
Thus far, I have spent time with the lead investigator at the Iberian Nanosystems Laboratory (INL) to develop a better sense of how the project is evolving and to understand the scope of some of the anticipated outcomes. I have also had the opportunity to witness certain manufacturing and testing processes in order to become more familiar with the concepts that underlie the work. My plan is to create a set of modular tools that emulate an SNN to then create an environment that reflects upon these connections- one in which 'neurons' will communicate with one another and respond to the particular environment in which they are situated.
- Matthew Biederman