A Minimalist White‑Box Unsupervised Learning Method Using Sparse Manifold Transform

Prof. Ma Yi and Turing‑Award laureate Yann LeCun jointly presented a paper at ICLR 2023 describing an extremely simple and interpretable unsupervised learning method that requires no data augmentation, hyper‑parameter tuning, or other engineering tricks, yet reaches performance close to that of state‑of‑the‑art self‑supervised learning (SSL) methods.

The approach leverages a sparse manifold transform that integrates sparse coding, manifold learning, and slow feature analysis. Using a single deterministic sparse manifold layer, the method attains 99.3% KNN top‑1 accuracy on MNIST, 81.1% on CIFAR‑10, and 53.2% on CIFAR‑100.

With a straightforward grayscale enhancement, the CIFAR‑10 and CIFAR‑100 accuracies improve to 83.2% and 57%, respectively, substantially narrowing the gap between this white‑box technique and leading SSL methods.

The authors also provide visual explanations of how the unsupervised representation transformation is formed, noting that the method is closely related to latent‑embedding self‑supervised approaches and can be viewed as the simplest variant of VICReg.

First author Yubei Chen is a post‑doctoral researcher at NYU’s Center for Data Science and Meta FAIR, supervised by Yann LeCun; he holds a PhD from UC Berkeley’s Redwood Center for Theoretical Neuroscience and BAIR. Prof. Ma Yi earned dual bachelor’s degrees in Automation and Applied Mathematics from Tsinghua University, a master’s and PhD in EECS from UC Berkeley, and is now a professor in the EECS department at UC Berkeley, as well as an IEEE, ACM, and SIAM Fellow.

The work aims to build the simplest possible “white‑box” unsupervised learning model without deep networks, projection heads, or complex engineering, thereby offering a principled, fully interpretable framework that may also shed light on unsupervised learning mechanisms in the brain.