TransparNet

Projektseite auf Deutsch

Project description

The TransparNet project aims to provide understanding, predictability, and thus transparency in the field of reservoir computing, a cutting-edge information processing methodology in bio-inspired machine learning conceived in the early 2000s. In general, in artificial neural networks, information processing and propagation are still insufficiently understood. A consequence is the difficult and often very limited predictability or controllability of learning outcomes and thus a high degree of intransparency.

The central feature of reservoir computing is that learning and adaptations occur exclusively in a readout layer. The readout layer is a downstream entity from the reservoir. The reservoir is uninvolved in the learning process and therefore remains structurally constant. This allows more direct access to the collective dynamics that occur between the input signals and the output of the result than is the case in other artificial neural networks.

The proposed project will apply methods from nonlinear dynamics of complex systems, network dynamics, and statistical physics to the reservoir. In particular, we plan to combine insights into information routing in and linear response theory in networks to elucidate the emergence of learned weights of the readout layer as a function of the input signals, the circuit structure of the network, the information processing task, and the nonlinearity in the readout layer.

A successful project would make information processing and forwarding in reservoir computing systems more understandable, predictable, and thus more transparent. These insights also potentially help to conceive and develop other novel methods in the future to make other artificial neural or otherwise bio-inspired, network-based, or general artificial information processing systems more transparent.

Project data