The diversity-accuracy duality in ensembles of classifiersd

Abstract

Horizontal scaling of Machine Learning algorithms has the potential to tackle concerns over the scalability and sustainability of Deep Learning methods, viz. their consumption of energy and computational resources, as well their increasing inaccessibility to researchers. One way to enact horizontal scaling is by employing ensemble learning methods, since they enable distribution. There is a consensus on the point that diversity between individual learners leads to better performance, which is why we have focused on it as the criterion for distributing the base models of an ensemble. However, there is no standard agreement on how diversity should be defined and thus how to exploit it to construct a high-performing classifier. Therefore, we have proposed different definitions of diversity and innovative algorithms which promote it in a systematic way.

We have first considered architectural diversity with an algorithm called WILDA: Wide Learning of Diverse Architectures. In a distributed fashion, this algorithm evolves a set of neural networks that are pretrained on the target task and diverse w.r.t. architectural feature descriptors. We have then generalised this notion by defining behavioural diversity on the basis of the divergence between the errors made by different models on a dataset. We have defined several diversity metrics and used them to guide a novelty search algorithm which builds an ensemble of behaviourally diverse classifiers. The algorithm promotes diversity in ensembles by explicitly searching for it, without selecting for accuracy. We have then extended this approach with a surrogate diversity model, which reduces the computational burden of this search by eliminating the need to train each network in the population with stochastic gradient descent at each step. These methods have enabled us to investigate the role that both architectural and behavioural diversity play in contributing to the performance of an ensemble.

In order to study the relationship between diversity and accuracy in classifier ensembles, we have then proposed several methods that extend the novelty search with accuracy objectives. Surprisingly, we have observed that, with the highest-performing diversity metrics, there is an equivalence between searching for diversity objectives and searching for accuracy objectives. This contradicts widespread assumptions that a trade-off must be found by balancing diversity and accuracy objectives. We therefore posit the existence of a diversity-accuracy duality in ensembles of classifiers. An implication of this is the possibility of evolving diverse ensembles without detriment to their accuracy, since it is implicitly ensured.