Feedback loops are ubiquitous in nature. They generate rich and complex behaviours that dominate many natural processes.
But while this superabundance has fascinated engineers and scientists, the sheer complexity of it has frightened them too. Until recently, they’ve largely avoided like the plague all but the most predictable of feedback mechanisms.
That’s beginning to change and one area that’s set to benefit more than most is computing.
Today, Yvan Paquot at the Universite Libre de Bruxelles in Belgium and a few pals reveal the first incarnation of an exotic new form of computing that exploits this feedback mechanism to perform impressively fast, analogue calculations.
The insight that has driven their work is that a nonlinear feedback mechanism is essentially an information processor. It takes a certain input and processes it to generate an output.
There’s an important difference between this and other types of computation. The feedback loop is a kind of memory that stores information about the system’s recent history. So this kind processing acts on small segment of the recent past.
That’s hugely significant. Many grand challenges in computing, such as speech recognition, depend on processing information from a small window in the recent past.
The type of computing that Paquot and friends have been working on is called reservoir computing. Here the reservoir consists of a reasonably large number of nodes that are connected together at random. Each node is some kind of non-linear feedback loop. The input, or inputs, are fed into random nodes in the reservoir and the output, or outputs, taken from other randomly chosen nodes.
The system is then trained to produce the desired computation by weighting the outputs in a certain way. For example, the input might consist of a waves of certain shapes and the output would be an indication that specific shapes had been recognised.
That sounds much like a neural network and in many ways it is. However, the important difference is that the reservoir workes essentially like a black box. Only the output signals are weighted during training, making this process much simpler than with a neural network, which are notoriously difficult to fine tune.
Paquot and co are not the first to play with reservoir computers. Others have built simple examples using everything from buckets of water to programmable chips to make the feedback nodes. Earlier this year, one group even used their design to carry out a basic form of speech recognition in the form of a standard benchmark test to recognise a few spoken words.
The Belgian group’s advance is to make their feedback nodes out of optoelectronic components that work more or less at the speed of light. Each node is an optoelectronic device in which a voltage modulates the intensity of a beam of light. The output is fed into fibre optic loop that channels the light back to the photodiode producing the voltage that modulates the beam. The result is a classic nonlinear feedback loop in which the ‘memory’ is determined by the length of the fibre optic loop.
Their reservoir consists of 50 nodes, connected together at random, with the readout taken from one node.They trained this device to perform a number of tasks such as distinguishing between sine waves and square waves and even simple word recognition.
This task involves five female voices speaking the digits 1-9, which the computer has to recognise. Paquot and co’s device did this with an error rate of just 0.4 per cent. In other words, 2 errors in 500 recognised words.
That’s not bad but what’s really impressive is the speed at which it does this task. “Our experiment is the ﬁrst implementation of reservoir computing fast enough for real time information processing.” They say it is almost 6 orders of magnitude faster than the earlier attempt and that a further speed increase of 2-3 orders of magnitude should be possible using various new, off-the-shelf optoelectronic components.
One important question is whether this approach will ever match purely electronic computing techniques. Paquot and co say that for now, this is an open question. But with more of this kind of work in the pipeline, that’s something we’re likely to find out much more about in the near future.
Ref: arxiv.org/abs/1111.7219: Optoelectronic Reservoir Computing
This new data poisoning tool lets artists fight back against generative AI
The tool, called Nightshade, messes up training data in ways that could cause serious damage to image-generating AI models.
Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist
An exclusive conversation with Ilya Sutskever on his fears for the future of AI and why they’ve made him change the focus of his life’s work.
The Biggest Questions: What is death?
New neuroscience is challenging our understanding of the dying process—bringing opportunities for the living.
Data analytics reveal real business value
Sophisticated analytics tools mine insights from data, optimizing operational processes across the enterprise.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.