Guillaume Pourcel (RUG)

During its PhD, ESR2 focussed on novel algorithms to leverage recurrent neural networks to do novel type computation. The starting point was “conceptor” a general framework to add another level of control on top of RNN. A conceptor embodies a “higher” cognitive control module that regulates the RNN. Inspired by biological brains, it allows attentional focusing, context setting, or generating predictive expectations.

During the POST-DIGITAL project, it was demonstrated that conceptors could enhance the stability of RNNs in various scenarios. Firstly, they effectively stabilize RNN dynamics against potential input pattern alterations. Secondly, they address partial failures within the network. By integrating a conceptor-derived control loop, as depicted in Figure 1, the system’s inherent resistance to perturbations is significantly bolstered. This methodology could potentially be employed in the future to counteract hardware perturbations.

 

Figure: A simple recurrent neural network is augmented with a conceptor control loop. The RNN is in a closed loop mode, i.e., driven by its own inputs. b) Geometrical interpretation of the conceptor control loop. The geometry of the dynamic is estimated online (grey ellipsoid), compared to a desired reference (red ellipsoid), and pushed toward it.

Another advancement with Conceptor that was developed is its application to enhance the machine learning capabilities of RNNs, specifically for interpolating between time series from a similar category. For instance, this could involve data from two sine/Fourier series of varying frequencies or motion capture data capturing both walking and running movements of a human. While numerous strategies using reservoir computing exist for such interpolations, they typically need a multitude of distinct time series. However, with the proposed method, the conceptor framework enables interpolation between just two unique datasets, eliminating the need for multiple samples.

Building on the successes with reservoir systems, the conceptor control loop was integrated with the established backpropagation algorithm to train the entire network’s weights. The goal was to enable precise control and interpolation using RNNs, even with complex datasets like motion modeling (MoCap) and motor control (MoCapAct).