← Back to Research

Continuous Manifold Alignment via Dual VAEs

A Zero-Forgetting Framework for Continuous Geometric Regression

The Core Concept

Traditional gesture recognition systems treat input as a rigid, discrete classification task. This project fundamentally reformulates continuous user interaction as an online geometric regression problem between two high-dimensional latent manifolds. By abandoning discrete spatial mappings, the system translates unsegmented, free-form spatiotemporal trajectories directly into a continuous linguistic space.

Architectural Innovations

Designed to run deterministically on edge computing devices, this framework introduces three core machine learning contributions:

  • Frozen Dual-VAE Pipeline: Representation learning is completely decoupled from user adaptation. A Spatio-Geometric CNN-VAE encodes position-invariant polar touch frames, while a Recurrent Linguistic VAE maps sub-word n-grams. Both spaces are frozen to establish geometrically stable topologies.
  • Linear-Complexity Temporal Path Integration: Standard temporal architectures (RNNs, deep signatures) introduce latency or scale exponentially. This architecture utilizes a custom, continuous path-integral transformation modulated by a static harmonic basis. It summarizes variable-length trajectory matrices into a fixed-dimensional vector with a strict linear complexity of O(N · D).
  • Zero-Forgetting Non-Parametric Adaptation: To map inputs across the two latent spaces without triggering catastrophic forgetting during fine-tuning, the framework uses a dynamic k-NN Ball Tree database. It employs an offset-stabilized Inverse Distance Weighting (IDW) interpolation, creating a self-correcting, "magnetic" mapping that allows for instant, one-shot personalization.

Read the Paper

Explore the mathematical formulation, signal preprocessing techniques, and component-level validation in the embedded paper below.