Fibre optic computation based on the human brain
Dendrites are branched structures at the end of nerve cells which receive and process incoming electrochemical signals. Their structure closely resembles that of a tree: many branches where many signals arrive and are locally preprocessed at these dendritic branches before the resulting output from these computations is sent to the cell body (soma) of the neuron for final processing. “These dendrites can be understood as consisting of many ‘local networks’ that take away a lot of the computational load from the ‘central processor’ – the cell body,” explains Florentin Wörgötter, professor of Physics at the University of Göttingen. Modern computer systems partly do the same when using multiple ‘slave processors’ for preprocessing, relieving the master CPU, says Wörgötter. Researchers in the EU-funded ADOPD project sought to harness these ideas to create ultrafast fibre-optical computational units for neuromorphic computing (based on the human brain), which uses significantly less energy than conventional computer systems and could meet the exponentially rising demand for processing speeds. “This technology is particularly interesting due to its low energy uptake and superfast speed,” adds Wörgötter, ADOPD project coordinator.
A fruitful interplay between theoretical modelling and hardware development
The ADOPD project combined theoretical modelling with actual hardware implementation. The researchers started with a known computational principle – a rule for changing the connection strengths within a dendritic network (a so-called synaptic plasticity rule) – and developed the corresponding electro-optical hardware, which allows for fast and efficient signal processing. “Due to the implemented plasticity rule, this system can adapt to changes in the input to some degree,” notes Wörgötter. In parallel, the team continued to develop theoretical models of dendritic signal processing in biophysically realistic neuronal models. “This interplay was fruitful, as at the end of the project the original plasticity rule could be augmented by new computational principles, discovered by our theoretical modelling efforts,” says Wörgötter.
The first fibre-optic synaptic plasticity
The ADOPD project successfully implemented synaptic plasticity using fibre optics, referring to a computational principle similar to that seen in neurons in the human brain. “In the course of doing this, we had also discovered and patented a novel method that allows for very fast synchronisation of different signals,” notes Wörgötter. This method rests on the same plasticity rule and has the potential to compete with conventional hardware solutions. The team is currently using this method to develop a microchip for applications ranging from telecoms to autonomous vehicles.
Creating the basis for next-generation neuromorphic computing
The ADOPD results are feeding into the development of neuromorphic computing, and some of the project partners are actively pursuing this. The remaining difficulties lie in scaling the existing system up to multiple dendrites and larger networks, though there is potential for the ADOPD system to reach commercial maturity within the next 10 years. Another line of research continuation concerns ‘holographic’ methods for fibre-optic-based calculations. One result of ADOPD has been that it is possible to use thicker fibres and inject a signal stream. This then leads to a sequence of complex patterns, similar to how movie frames follow each other in a film. “One can use these patterns for calculations, too, and this offers the advantage of a highly compressed information stream,” Wörgötter explains. “This is a clear extension from the existing results from ADOPD – but it is difficult.”
Keywords
ADOPD, neuromorphic computing, theoretical modelling, hardware, neurons, brain, computing