<ul><li>WARP (Weight-space Adaptive Recurrent Prediction) framework unifies weight-space learning with linear recurrence for sequence modeling.</li><li>Unlike conventional RNNs, WARP parametrizes the hidden state as the weights of a separate neural network, allowing higher-resolution memory and gradient-free adaptation at test-time.</li><li>Empirical validation shows that WARP outperforms state-of-the-art baselines on various classification tasks and offers valuable insights into model's inner workings through weight trajectories.</li><li>WARP's efficacy is demonstrated across tasks like sequential image completion, dynamical system reconstruction, and multivariate time series forecasting, showcasing its expressiveness and generalization capabilities.</li></ul>

Weight-Space Linear Recurrent Neural Networks

Discover more