Reproduce “Unsupervised Feature Learning from Temporal Data” on OpenResearch

About this paper

Ross Goroshin, Joan Bruna, Jonathan Tompson, David Eigen, Yann LeCun

A new unsupervised learning framework extracts robust features from temporal data like video by combining a convolutional pooling autoencoder with slowness and sparsity regularization. This method addresses limitations of purely contrastive approaches by incorporating a reconstruction objective, promoting informative and invariant representations.

Key insights

• Integrating a reconstruction objective into unsupervised feature learning is critical for ensuring features are not only discriminative but also inherently informative about the input data.
• Leveraging temporal coherence through a slowness prior, combined with spatial invariances from convolutional pooling, allows for the extraction of robust and general-purpose features from unlabeled video.
• An L1 sparsity penalty on hidden activations effectively promotes the learning of more distinct, localized, and interpretable filter structures in the decoder dictionaries.

Summary by alphaXiv.