In self-supervised representation learning, a common idea behind most of the state-of-the-art approaches is to enforce the robustness of the representations to predefined augmentations. A potential issue of this idea is the existence of completely collapsed solutions (i.e., constant features), which are typically avoided implicitly by carefully chosen implementation details. In this work, we study a relatively concise framework containing the most common compo- nents from recent approaches. We verify the existence of complete collapse and discover another reachable collapse pattern that is usually overlooked, namely dimensional collapse. We connect dimensional collapse with strong correlation between axes and consider such connection as a strong motivation for feature decorrelation (i.e., standard- izing the covariance matrix). The capability of correlation as an unsupervised metric and the gains from feature decor- relation are verified empirically to highlight the importance and the potential of this insight.
“The ability to ask beautiful questions, often in very unbeautiful moments, is one of the great disciplines of a human life. And a beautiful question starts to shape your identity as much by asking it as it does by having it answered. And you don’t have to do anything about it. You just have to keep asking. And before you know it, you will find yourself actually shaping a different life, meeting different people, finding conversations that are leading you in those directions that you wouldn’t even have seen before.”
— David Whyte, On Being interview “The Conversational Nature of Reality”