IBM created the world's first stochastic phase-transition neurons


    Artistic interpretation of the appearance of a stochastic neuron from IBM

    Developers from IBM created the world's first stochastic neurons with a phase transition, which promises us the creation of a neuromorphic chip that will significantly accelerate the calculation and processing of information. Attempts to create such a technology were reported back in 2012, but then Intel was dealing with this issue. Four years later, developers from IBM were able to achieve results in this area.

    What is the fundamental difference between a chip made of stochastic neurons with a phase transition from classical silicon?

    Phase transitionin thermodynamics, the transition of a substance from one thermodynamic phase to another with changing external conditions. In fact, the creation of a stochastic neuron with a phase transition will allow the creation of an artificial model of such a biological system as the brain.

    A corresponding study was obtained by Nature magazine back in May 2015, and published in April 2016.

    image
    The stochastic phase transition neuron, IBM

    Like its biological counterpart, IBM's artificial neuron has the same structure. It contains dendrites (inputs), a membrane (lipid bilayer), core and axon (output). A distinctive feature of an artificial neuron is its neuronal membrane. In a real neuron, this is a lipid bilayer, which, in fact, works as a resistor and capacitor: a signal is passed only when a sufficient charge accumulates on the dendrite, which leads to a surge in charge generation and the signal passes further to other neurons.

    image
    Dynamics of the behavior of a neuron during a phase transition, IBM

    The phase transition in an artificial neuron is important in that, with its help, scientists and engineers will be able to emulate the cognitive learning process inherent in the real biological brain. In fact, this development can be actively used in existing projects of neural networks for network training and cognitive computing, which will significantly speed up the processing of information, for example, data analysis on the Internet.

    In the IBM neuron, the membrane is replaced by an alloy of germanium, antimony and tellurium (GeSbTe or GST). GST has already been used in the production of rewritable CDs due to the fact that it is subject to phase transition. This means that it can successfully exist in two different states (crystalline and amorphous) and easily switches between them upon receipt of heat, which is indirectly confirmed by rewritable RW discs. In the case of GST, the material has fundamentally different properties depending on its phase. In crystalline it is a conductor, in amorphous it is an insulator.

    image
    A visual demonstration of the difference in the behavior of a neuron in a few seconds, IBM

    Particular attention was also paid to stochasticity in the operation of the chip. Scientists have suggested that the successful long-term work of our brain with incomparably lower voltage of electromagnetic pulses in comparison with artificial neural networks is ensured by the formation of random connections between neurons. It is this “accident” that was implemented by IBM engineers in the new technology due, inter alia, to the use of materials subject to phase transition, namely GST, in the technology.

    At rest, the shell of artificial neurons is in an amorphous state. When a signal (discharge) is applied, it begins to crystallize, which ultimately makes the neuron from the insulator a conductor. After passing the signal, the neuron passes through a “reset”, i.e. its membrane returns to its amorphous state.

    Where is the stochasticity inherent in living organisms due to noise, the environment, and so on? Stochasticity appears at the reverse amorphization stage after crystallization of the shell. For each neuron, the time to return to the initial state and the total amorphous membrane is always different, which leads to the creation of approximately the same degree of “randomness” in their work as that of biological analogues. That is why engineers can not accurately predict which specific neurons will be ready at the right time and involved in the transfer of information.

    image
    IBM Neuromorphic Computer Project

    IBM engineers have already collected 5 cubes 10 by 10 neurons and combined them into a network of 500 pieces. This block showed the same behavior in terms of population coding as biological neurons, and also circumvented the limitations for processing digital signals formulated in the Kotelnikov theorem .

    Also popular now: