Application-agnostic Machine Learning
In this thust I particularly focus on understanding the behavior of ML model
The current powerhorse of ML is backpropagation discovered by Paul Werbos in 1981 and then reinvented for ML by Geoffery Hinton. The method is still the main tool to train ML models from DQN to chatGPT. Moreover, the method was built in the context of pure mathematics and addresses the challenge of credit assignment backward in time. The key insight is that, methods that are designed to address specific challenges fundamentally and methodically can stand the test of time and solve key problems. The goal of this vertical to understand, develop tools that are agnostic to applications and provide mathematically grounded solutions to several key challenges.