Transparent Machine Learning

Machine learning is the process of creating algorithms that infer patterns from data, where some algorithms are commonly viewed as more interpretable than others. The need for transparency has surged to importance in recent years. Healthcare professionals increasingly rely on the aid of image analysis when diagnosing problems, decision support systems often fuel commercial decisions, and as of last year, all businesses must legally comply with the GDPR and the freedom to information. Interpretability encompasses the terms explanation, how an algorithm responds when exposed to a specific instance, and interpretation, how the structure, weights and parameters of a model can highlight important features of the data. This research will look to extensively evaluate and develop both aspects.

A learning algorithm may produce a correct result, but without insight as to why, it is very difficult to assert that the process is correct and logical, or that a solution is always applicable. Model output is increasingly accepted under the assumption that designers and professionals understand the algorithms which they create, but this is seldom the case, as faith is often a substitute for understanding. The amount of internal and external model parameters continues to increase exponentially as the industry advances, and minor setting changes produce into incredibly non-linear outputs. A lack of interpretability usually results in implementing less precise algorithms with poorer accuracy out of necessity; but this compromise conceals relationships, importance and redundancy amongst features.

The key objectives of this research are:
1) To improve the accuracy of interpretable algorithms,
2) To improve the interpretability of complex algorithms,
3) To change how algorithms are intrinsically viewed as being either interpretable or not,
4) To apply these new techniques to a commercial environment,

The research will be novel through a combination of newly created techniques, and techniques that have not been applied across domains before. This will begin with research that bridges subtle gaps, and later expand to largely new fields.