AutoML Solutions for Generative Models

Nowadays artificial intelligence (AI) is one of hottest and the most debated topics in applied science. Being it applied to a plethora of different and possibly delicate use cases, both the technical performances and the understanding, explainability and trustworthiness of such technology are crucial in order for this to fit in human lives and become pervasive in the society. Despite the leap forward in performances occurred during the last few years, most of the machine Learning (ML) and deep learning (DL) algorithms still involve manual dataset-specific fine tuning, making the models' predictions tightly coupled with the input and the entire pipeline heavily problem-specific, hence requiring an high degree of domain expertise usually provided by ML experts. The rise of AutoML is trying to fill this gap by automating several steps in the process with the goal of providing good off-the-shelf optimized models agnostic of the input data and more accessible. This work, enclosed in the broader context of explainable AI, adopts AutoML techniques together with generative models, in particular variational autoencoders, with the goal of building an accurate, low latency and dataset-agnostic optimization framework. Such a generalized and near real-time setting represented also the main limitation of this solution as it induced the trade off between accuracy and time of execution at multiple stages of the thesis. Different optimizations approaches, primarily related to hyperparameter optimization, have been tested spanning from standard techniques to meta-heuristics and evolutionary algorithms. As in the context of this work the neural networks in use were rather shallow due to the latency requirements and the input data peculiarities, a side analysis involving neural architecture search techniques was implemented. Finally, the different methodologies being studied have been tested and evaluated taking into account the following performances: accuracy, latency and explainability. Experimental evaluations such as comparing the different optimization techniques induced insights and considerations on the better performing optimization techniques for a given input and shone a light on which directions the scientific literature could take to further automate ML algorithms and their generative models applications in the context of trustworthy and human-centered AI.