Abstract
We propose the factorized action variational autoencoder (FAVAE), a state-of-the-art generative model for learning disentangled and interpretable representations from sequential data via the information bottleneck without supervision. The purpose of disentangled representation learning is to obtain interpretable and transferable representations from data. We focused on the disentangled representation of sequential data since there is a wide range of potential applications if disentanglement representation is extended to sequential data such as video, speech, and stock market. Sequential data are characterized by dynamic and static factors: dynamic factors are time dependent, and static factors are independent of time. Previous models disentangle static and dynamic factors by explicitly modeling the priors of latent variables to distinguish between these factors. However, these models cannot disentangle representations between dynamic factors, such as disentangling "picking up" and "throwing" in robotic tasks. FAVAE can disentangle multiple dynamic factors. Since it does not require modeling priors, it can disentangle "between" dynamic factors. We conducted experiments to show that FAVAE can extract disentangled dynamic factors.
Introduction
In machine learning, "disentangled representation" means that a single latent variable in a model controls a single, interpretable factor of the data (e.g., one variable controls a face's smile, another controls hair color). While models like $\beta$-VAE perform well for static images, sequential data such as video or time-series is significantly harder to disentangle. This difficulty arises because sequences contain both static factors (time-independent, like a character's shirt color) and dynamic factors (time-dependent, like movement).
Recent models attempt to separate these factors by explicitly hard-coding different time dependencies into the model's graphical priors. However, this approach fails when a sequence contains multiple dynamic factors occurring simultaneously (e.g., a robotic arm's trajectory shape vs. its speed). Because these dynamic factors share the same time dependency, existing models often lump them together. FAVAE addresses this gap by disentangling multiple dynamic factors without relying on explicit prior modeling or supervised labels.
Methodology
FAVAE extends the standard Variational Autoencoder (VAE) framework through two primary innovations:
- Information Bottleneck (IB) with Capacity Constraint: Instead of manually engineering graphical priors, FAVAE utilizes the Information Bottleneck principle applied to sequences ($x_{1:T}$). It employs a loss function similar to $\beta$-VAE to compress the latent space but introduces an "information capacity" parameter ($C$). This parameter linearly increases during training, preventing a common failure mode where high regularization forces the model to ignore latent variables, which typically degrades reconstruction quality.
- Ladder Network with Time Convolutions: To process temporal data, FAVAE incorporates 1D time convolutions (similar to Quasi-Recurrent Neural Networks) within a hierarchical "Ladder" architecture. As data passes through successive layers, it becomes increasingly abstracted, allowing the model to disentangle representations at different levels of temporal scope.
Experiments and Results
The authors evaluated FAVAE on several sequential datasets, including 2D Reaching (point movement to a goal), 2D Wavy Reaching (complex trajectories with local dynamic factors), and the Sprites Dataset (video containing various static and dynamic factors).
Key Insights
- No Need for Prior Engineering: By relying on the Information Bottleneck, FAVAE separates dynamic factors that share the exact same time dependency—a significant advancement over previous sequential VAEs.
- Hierarchical Sorting: The ladder network naturally sorts factors by their temporal abstraction. Long-term, global factors (like a final goal position) are captured by deeper layers, while short-term, local factors (like specific curve segments) are captured by earlier layers.
- Quantitative Superiority: On the complex 2D Wavy Reaching task, FAVAE achieved a significantly higher Mutual Information Gap (MIG) score (0.67) compared to the FHVAE baseline (0.22), while maintaining much lower reconstruction loss.
Conclusion
FAVAE demonstrates a robust ability to extract disentangled dynamic factors from sequential data without supervision. While the current 1D time convolution approach may need further scaling for real-world complexity, the model's success in separating overlapping dynamic factors is a major step forward. Future applications include using FAVAE in Reinforcement Learning to extract "macro-actions," potentially reducing action spaces and accelerating agent training.
Summary updated