This is Science: What's inside a neuromorph chip?

After the recent announcement of the neuromorphic chip from IBM on Habré, it is time to get acquainted with how the work of real neurons is transferred to the iron of neuromorphic chips. And help us in this article, published in ACSNano, about a three-dimensional electronic synapse.

Let's start with a little background. Once we have learned where scientists derive information about the device of the brain, then there was a big post with answers to readers' questions about the project The Human Brain, but recently an announcement of a neuromorphic chip from IBM appeared on Habré. In a series of these publications, I would also like to devote time to how neuromorphic chips are arranged at the basic level.

One of the main components of neuromorphic computing, in general, and neuromorphic chips, in particular, is the synapse or excitation transfer system from the neuron to the neuron, because the neuron itself is often just a metal band, a conductor. A synapse in the nervous tissue is the point of contact between two neurons or between a neuron and a signaling effector cell that serves to transmit a nerve impulse.
')

The main elements of the nerve cell

By the mechanism of transmission of such a pulse, synapses can be divided into chemical - that is, using neurotransmitter molecules - and electrical - that is, due to the "breakdown" of the intercellular fluid by an electric pulse. Electric synapses are a pair of cell membranes that are at a very close distance (just a few nanometers) from each other due to special proteins that perform the functions of excitation transfer.

Typically, data on a particular device or model of a neuromorphic chip used and the type of neuron connections in it to the network is the know-how and subject of the NDA (not-disclosure agreement), but in purely scientific, non-industry publications one can learn a lot about their device.

So, a group of Chinese and American scientists from Peking University, Stanford and the University of Arizona published an article in the ACSNano journal devoted to the development of a new three-dimensional electronic synapse that consumes ultra-small amounts of electricity, which is shown in the figure:


(a) A conventional 2D array for an electrical neural network, where each synapse is located at the intersection of the conducting lines of the pre-neuron and the post-neuron. (b) A circuit diagram of a concept with synapses based on a switching resistance device (resistive switching device). (c) The most compact synaptic layout (high-density application). (d) Synaptic layout for high-accuracy computation. (e) TEM image of the cross section of the resulting electrical synapse.

The basis for the development of steel materials with switchable resistance, for example, HfO _x or HfO _x / AlO _y , which, depending on the duration and amplitude of the applied voltage can change their resistive properties over a wide range - potentially more than 3 orders of magnitude from 10 ³ to 10 ⁶ ohms. Such switching occurs due to the migration and restructuring of oxygen vacancies inside the oxides.

And in order to check the electronic synapses in action, a two-layer neuromorphic chip was created, the first layer of which consisted of 32 x 32 neurons sensitive to the brightness of pixels, and the second one of 16 cortex neurons connected by ordinary or three-dimensional electrical synapses. The results on the face: a three-dimensional electronic synapse gives a better recognition compared with the usual due to the smaller deviation of the resistance. At the same time, the training takes place at a reduced power consumption by pulses of 50 ns, at voltages below 2.5V and currents below 0.3 μA


(a) The abundance of oxygen vacancies leads to low resistance and, conversely, (b) their lack means high resistance and low current. (c) Deviation of resistance depending on energy expended. (d) Pattern used to train the system. (ef) Patterns obtained using a normal neural network and created on the basis of three-dimensional synapses, respectively. (g) Recognition accuracy.

Compared with the primitive two-dimensional neural network of synapses, the developed device looks incredibly difficult, but the materials used are relatively cheap and ubiquitous in the electronics industry, which, according to the authors of the paper, makes it possible to manufacture such neuromorphic chips with extremely low cost.

Original ACSNano article (DOI: 10.1021 / nn501824r)

---

Full list of published articles This is Science on GeekTimes:
This is Science: Simple and Cheap Solar Energy
This is Science: Is graphene life or death?
This is Science: Blow and Get Electricity
This is Science: Silicon Electronics: Bend Me Completely!
This is Science: Resilient quantum dot display
This is Science: Putting triboelectricity at the service of humanity
This is Science: 3D optical printing moves to the micro level
This is Science: What's inside a neuromorph chip?
This is Science: News from graphene fields
This is Science: 3D lithography to masses
This is Science: The discharge of alkaline batteries or why the battery jumps
This is Science: microcannons and nano-core
This is Science: wearable electronics and triboelectricity. Part 1
This is Science: wearable electronics and triboelectricity. Part 2

---

Sometimes briefly, and sometimes not so much about the news of science and technology, you can read on my Telegram channel - welcome;)