The research comes from the University of Michigan. Here computer scientists have shown that a different form of neural network, composed of memristors, can improve the efficiency of teaching machines to think in a way that is closer to humans. Central to this is the network, a type of reservoir computing system. This network can predict words before they are said during conversation and also help to predict future outcomes based on the present situation.
Reservoir computing is a framework for computation whereby an input signal is fed into a fixed (random) dynamical system called a reservoir. The dynamics of the reservoir map the input to a higher dimension. Following this, a readout mechanism can be trained to read the state of the reservoir and map it to the desired output. Liquid-state machines and echo state networks are both types of reservoir computing, and are classed as third generation of neural network models.
Memristors are a types of resistive devices that are capable of performing logic and storing data. This is very different to conventional computer systems, which have processors that perform logic separate from memory modules. For the research, the computer scientists used a special memristor that can memorize events only in the near history.
The advantage is that reservoir computing systems, built with memristors, can skip most of the standard machine training process and provide the network the capability to remember at a more efficient rate.
The trials to date have been with tests of handwriting recognition. For these tests. numerals were broken up into rows of pixels, and fed into the computer with voltages like Morse code, with zero volts for a dark pixel and a little over one volt for a white pixel. The initial experiments have shown that by only 88 memristors as nodes to identify handwritten versions of numerals (a conventional network that would require thousands of nodes), the reservoir achieved 91 percent accuracy. This is a good foundation for further research.
The research has been published in the journal Nature Communications, with the peer-reviewed study titled “Reservoir computing using dynamic memristors for temporal information processing.”
