Multi-directional continuous association with input-driven neural dynamics

We present an input-driven dynamical system approach to continuous association. Previous formulations of associative reservoir computing networks and associative extreme learning machines are unified and generalized to multiple modalities. Association in these networks proceeds by externally driving parts of the network. Through continuous variation of driving inputs, a continuous association of output patterns is achieved. Robust association in this scheme requires to cope with potential error amplification of feedback dynamics and to handle differently sized input and output modalities such that the outcome of association is controlled by the driving inputs. We propose a dendritic neuron model in combination with a regularization technique to address both issues. The presented method allows for tuning contributions from each modality to the hidden representation by prescribed factors while the regularization of network weights mitigates the problem of error amplification. The scalability of the approach to high-dimensional applications is demonstrated in image and audio processing scenarios.