Interpreting Video Representations with Spatio-Temporal Sparse Autoencoders

Video models have the architecture to support encoding temporal aspects of the input, i.e. concepts that require more than one frame to define, like object movement. However, because motion and appearance change are entangled, a feature that appears to encode temporal concepts like object movement may instead encode non-temporal concepts like spatial presence at different positions whose activations merely correlate with motion. To resolve this, we introduce the No-False-Positives (NFP) test, a statistical test with a formal guarantee of statistical specificity for identifying features encoding temporal concepts. Using specially designed stimuli, we apply this test to representations produced by VideoMAE, V-JEPA2, and TimeSformer, and observe that temporal features exist in each model. We validate the test first by showing it yields no temporal features for the DINOv2 image-encoder. Then, using a synthetic representation space, we show that detection rate increases with a direction's projection onto the temporal subspace, and that Sparse Autoencoder features test-flagged as temporal project much more of their norm onto the temporal subspace than unflagged features. Interventions on temporal SAE features yield mixed results across models, suggesting that intervening on temporal features may not outperform a non-temporal feature baseline.

Video representation learning has seen tremendous progress in recent years. This has been driven by many factors, including the scale of training and the success of self-supervised models trained on next-frame prediction. While these factors have pushed the boundaries of what video models can do, they also introduce their own set of limitations. First, scaling video models can reach prohibitive costs, with recent models needing hundreds of years of video to be trained. Second, learning to predict the next frame (or embedding of the frame) naturally focuses on spatial information and as a result, video models still struggle with temporal understanding. In this paper we propose a novel approach that uses motion as a modality to alleviate both of these core issues. Given the motion in a video in the form of point tracks, we mask some of the tracks and use a masked autoencoder to reconstruct the missing tracks. This allows us to learn a representation in a self-supervised manner which we call TIME (Temporally-Informed Motion Embedding), and is focused on capturing temporal information. Also, as motion is inherently appearance-invariant, TIME needs far fewer examples to generalize well. As a result, without bells and whistles, on temporal tasks TIME performs on par with state-of-the-art models, using up to 4 orders of magnitude less training data. For general tasks, TIME can be used in combination with existing representations, and we observe that it leads to a significant improvement for V-JEPA 2, RVM and VideoMAE on standard benchmarks such as SSV2, EgoExo4D and Diving48. These results point to a new promising video paradigm for both more temporally-aware as well as more scalable models.

Sparse Autoencoders (SAEs) began as an unsupervised tool for decomposing neural representations into sparse, interpretable features, and are increasingly used not only for passive analysis but also for active interventions such as unlearning, bias mitigation, and concept editing. A central challenge for these editing and steering methods is reliably matching features to target concepts; most current approaches address this by computing post-hoc scores over an already-trained, frozen dictionary. We instead introduce ClasSAE, a novel method that both automatically assigns classes to features and guides the encoder toward class-separable representations during training. Specifically, we apply a differentiable top-$k$ operator to a trainable feature--class affinity matrix with per-feature budgets, coupling the features selected for each sample to the classes they are trained to represent. Because gradients flow through the selection of active features rather than only through their magnitudes, the encoder and the affinity matrix co-adapt rather than being fit in separate stages. The result is a dictionary that is both class-separable and class-annotated, with no need for post-hoc probing. We propose three variants for enforcing sparsity within this framework, which achieve comparable overall performance with slightly different trade-offs. Using CLIP ViT-L/14 embeddings on ImageNet, we show that the learned affinity matrix agrees closely with an independently estimated post-hoc feature-class matrix computed on held-out data. The model also supports direct class prediction from the encoder and affinity matrix alone, without a separately fitted classifier, and its more class-aligned encoder yields improved separation in Targeted Probe Perturbation evaluations. https://github.com/St0pien/ClasSAE.