Part 3 The neural network paradigm

 

The past 10 years in machine-learning research have been deeply influenced by smart people turning their attention to the brain and the way it works. With capacious terrestrial computing and graphical processing units (GPUs) that optimize machine-learning code speed and deployment by orders of magnitude, both older models and newer approaches that are computationally difficult to empirically test are widely available, democratized by the cloud, not hidden and usable only by big web companies.

As it turns out, modeling how to think, hear, see, and speak based on the brain, and taking those models and easily deploying them, sharing them, retraining and adapting them, and using them have yielded many advances, such as the intelligent digital assistant in your smartphone or a home assistant device that can order food or turn the channel to your favorite program based on your talking to it and its talking to you.

These machine-learning models are called neural networks. Neural networks are modeled based on your brain, which contains networks of connected neurons. The recent release of models like GPT-3 that automatically generate believable news articles, plays, tweets, and you name it feels like the beginning of the (r)evolution.