Self-Supervised Relational Reasoning for Representation Learning
This work addresses the challenge of costly annotation in machine learning by enhancing self-supervised representation learning, though it appears incremental as it builds on existing relational reasoning concepts.
The paper tackles the problem of building useful representations in self-supervised learning without manual annotation by proposing a novel relational reasoning method that trains a relation head to discriminate intra- and inter-reasoning, resulting in an average 14% accuracy improvement over the best competitor and 3% over the most recent state-of-the-art model.
In self-supervised learning, a system is tasked with achieving a surrogate objective by defining alternative targets on a set of unlabeled data. The aim is to build useful representations that can be used in downstream tasks, without costly manual annotation. In this work, we propose a novel self-supervised formulation of relational reasoning that allows a learner to bootstrap a signal from information implicit in unlabeled data. Training a relation head to discriminate how entities relate to themselves (intra-reasoning) and other entities (inter-reasoning), results in rich and descriptive representations in the underlying neural network backbone, which can be used in downstream tasks such as classification and image retrieval. We evaluate the proposed method following a rigorous experimental procedure, using standard datasets, protocols, and backbones. Self-supervised relational reasoning outperforms the best competitor in all conditions by an average 14% in accuracy, and the most recent state-of-the-art model by 3%. We link the effectiveness of the method to the maximization of a Bernoulli log-likelihood, which can be considered as a proxy for maximizing the mutual information, resulting in a more efficient objective with respect to the commonly used contrastive losses.