Core Concepts
Explicitly modeling generality into the objective of self-supervised learning to improve the model's ability to learn general representations and achieve superior performance on various unseen tasks and domains.
Abstract
The paper aims to address the limitations of existing self-supervised learning (SSL) methods in achieving true generality. It first provides a theoretical definition of generality in SSL, which involves discriminability, transferability, and generalization. Based on this, the paper proposes a σ-measurement to quantify the generality of SSL models.
To explicitly model generality into SSL, the authors propose a novel SSL framework called GeSSL. GeSSL learns general representations through a bi-level optimization process:
The first level optimizes the SSL model to quickly adapt to each training task, corresponding to the learning generality.
The second level further refines the learned representations to capture general knowledge across various tasks, corresponding to the evaluation generality.
The second-level optimization is guided by a self-motivated target based on the proposed σ-measurement, which encourages the model to update towards the optimal direction for generality.
The paper provides theoretical analysis on the rationality of the task construction and the performance guarantee of GeSSL. Extensive experiments on various benchmarks, including unsupervised learning, semi-supervised learning, transfer learning, and few-shot learning, demonstrate the superior robustness and generalization of GeSSL compared to state-of-the-art SSL methods.
Stats
"The model fθ trained on task Ttr
i can achieve competitive performance quickly on task Ttr
j through few samples Dtr
i."
"The trained model f*θ can achieve comparable performance with all the optimal task-specific models on all the target tasks Tte through minimal additional data Dte
min."
Quotes
"The generality of SSL can be reflected in two aspects: learning generality and evaluation generality."
"We explicitly model generality into self-supervised learning and propose a novel SSL framework, called GeSSL."
"GeSSL introduces a self-motivated target based on σ-measurement, which enables the model to find the optimal update direction towards generality."