Two-shot learning of continuous interpolation using a conceptor-aided recurrent autoencoder

Guillaume Pourcel, Mirko Goldmann, Steven Abreu, Miguel C. Soriano

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Conceptors, Few Shot Learning, Recurrent Neural Networks, BPTT, Motion Modelling, Low Dimensional Dynamics

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a conceptor based regularization of Backpropagation through time that facilitates few shot learning of continuous interpolation between distinct temporal patterns.

Abstract: Generalizing from only two time series towards unseen intermediate patterns poses a significant challenge in representation learning. In this paper, we introduce a novel representation learning algorithm, "Conceptor-Aided Recurrent Autoencoder" (CARAE), which leverages a conceptor-based regularization to learn to generate a continuous spectrum of intermediate temporal patterns while just being trained on two distinct examples. Here, conceptors, a linear subspace characterization of neuron activations, are employed to impose a low-dimensional geometrical bottleneck on the neural dynamics. During training, CARAE assembles a continuous and stable manifold between the two trained temporal patterns. Exploiting this manifold in the inference, CARAE facilitates continuous and phase-aligned interpolation between temporal patterns that are not linked within the training data. We demonstrate the effectiveness of the CARAE framework through comprehensive experiments on temporal pattern generation tasks and the generation of novel complex motion patterns based on the MoCap data set.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9329