In fairy tales and mythology, heroes would seek out mystics and fortune tellers to ask their most pressing questions. These fortune tellers would generally dim the lights, speak in a husky voice, and pull out a crystal ball to repeat the questions into. We know that a universe in which you could pull out a pocketable crystal ball which could answer any questions you might have seems far fetched but humour us.
Search engines are kind of the real-world equivalent to crystal balls – you ask a question and a magical object relates that question to an answer or about 1 100 000 000 (0.6 seconds). Only, there’s no magic. ML is ideally suited to these types of problems because every query might not have a directly determinable answer.
Applied machine learning is the realisation of a model capable of mapping an input-output space without any knowledge of what the map looks like. The outcome of these types of models is a map or rulebook relating inputs to outputs – enabling the prediction of a system output given a certain input with a high accuracy. Because applied ML can be used to create these maps without any knowledge of what the actual map looks like, things can get pretty weird and wonderful. Theoretically, an algorithm could (and would) determine a relationship between the price of oranges in Johannesburg today and how far Perseverance moved four sols ago. Now, is something like this worth building? No. Should you build it anyway? Absolutely!
Most fairytale princesses (and princes) have ugly step-siblings and applied ML is no different. Theoretical ML is the study of new algorithms and the understanding of the underlying math, statistics and probability. Step-siblings, while generally horrendous, are wonderful in their own right. Without theoretical ML, we would never have scratched the surface of deep learning. Our modeling capability would have a very hard limit in terms of accuracy and new, weird and wonderful ML algorithms would never be investigated to the point at which they become implementable.
Why focus on applied ML? This ties back to bringing direct and measurable impact to clients, people and the world. Sure, theoretical leaps and bounds need to be made. But, value is extracted by interfacing the algorithm and the data. These data are then turned into actionable insights based on clear, repeatable maps of the underlying dynamics. It is these insights that bring the magic of ML into the real-world.
Follow your curiosity,
The Afrobots @ Afrobotix.