Analysis of Trustworthiness in Machine Learning and Deep Learning

Abstract

Trustworthy Machine Learning (TML) represents a set of mechanisms and explainable layers, which enrich the learning model in order to be clear, understood, thus trusted by users. A literature review has been conducted in this paper to provide a comprehensive analysis on TML perception. A quantitative study accompanied with qualitative observations have been discussed by categorizing machine learning algorithms and emphasising deep learning ones, the latter models have achieved very high performance as real-world function approximators (e.g., natural language and signal processing, robotics, etc.). However, to be fully adapted by humans, a level of transparency needs to be guaranteed which makes the task harder regarding recent techniques (e.g., fully connected layers in neural net-works, dynamic bias, parallelism, etc.). The paper covered both academics and practitioners works, some promising results have been covered, the goal is a high trade-off transparency/accuracy achievement towards a reliable learning approach.