Photonic communication channels---which code information on waves of light---compose both the vast networks that underlie the internet and the fiber optic links that connect data centers together. Electronics, in contrast, has dominated computing, driving the landscape for nearly sixty years with an exponential progression towards better processors in a phenomenon known as Moore’s law. Electronic computing, however, is running up against fundamental limits that are increasingly harder to circumvent. Today, artificial intelligence is bottlenecked by electronic processing power, and its needs outstrip the available demand.
In contrast, optical channels are well known to be energy efficient over long distances, and have an incredible capacity for information---an optical wire less than a millionth of a meter in width can carry more than a terabit of information per second. Light also has one more property that makes it especially useful for artificial intelligence: the ability to simulate neural networks efficiently. Our brains are composed of billions of individual cells called neurons, which communicate along millions of billions of channels called synapses. Modern artificial intelligence relies on models that are inspired by the structure of our brains. Because light can be routed in complex configurations efficiently, we can use optics to simulate large networks using little energy.
My thesis explores how photonics can be used to represent neurons that behave much like the cells in our brains. The fundamental model I explore is an on-chip laser, which underlies modern photonic communication. I investigate how its dynamics can lead to interesting operations both in theory and in experiment, culminating in a device that encodes information using timed pulses---or spikes---in very much the same way that the brain encodes information, but more than ten million times faster. I also look at the implications of other photonic neuron models, and conclude with a general analysis of how optical systems may fare against electrical ones in the artificial intelligence systems of the future.