- Scientists build transistor-free analog switching devices using silicon germanium.
- The small artificial synapses chip can recognize samples of handwriting, with 95% accuracy.
- It aims to eliminate the limited scalability of current neuromorphic chips and perform complex tasks that are only possible with supercomputers.
At present, it is quite impossible to beat human brain when it comes to processing power. However, MIT engineers in the evolving field of neuromorphic computing have tried to develop a chip that works like the human brain.
Unlike today’s digital hardware that works on binary bits, the new brain-chip would work in an analog manner, exchanging weights (gradient of signals), like neurons that get activated in numerous different ways based on quantity and type of ions, flowing across a synapse.
Neuromorphic engineering is not something new: the concept was developed in the late 1980s that describes the usage of VLSI (very large scale integration) systems to mimic neuro-biological architectures in the nervous system. There are already dozens of big companies and universities working in the area.
Recently, MIT engineers revealed transistor-free analog switching devices that may overcome some limitations like bonded scalability and high power consumption.
Inspired By Human Brain
The adult human brain contains 86±8 billion neurons and 85±10 billion of non-neuronal cells. A single neuron is capable of transmitting instructions to thousands of other neurons through synapses.
A synapse in a nervous system is a structure that allows a neuron to pass a chemical or electrical signal to another neuron. There are over 100 trillion synapses that mediate neuron signaling. Some connections are strengthened while some are pruned, enabling the human brain to remember facts, recognize patterns, and perform learning tasks, at fast speeds.
Like human brain, tiny neuromorphic chips could process million of streams in parallel. So far, this has been possible only with supercomputers, but now researchers have come up with an artificial synapse capable of precisely controlling the strength of electric current flowing through it (like flow of ions between neurons).
They used silicon germanium to build a small artificial synapses chip, which can recognize samples of handwriting, with 95% accuracy. It’s a big step towards developing efficient, portable and low power neuromorphic hardware for pattern recognition and other complex learning tasks.
The Problem With Current Systems
Most of the neuromorphic hardware designs try to mimic the synaptic connectivity between neurons through two conductive layers divided by a switching medium, similar to synapse-like space. Just like how a synapse’s ‘weight’ changes, ions are supposed to move in the switching medium to form conductive filaments when the voltage is applied.
However, in current designs, we can’t control the ions flow precisely because of the switching medium. Since most of these switching medium is made of amorphous materials, it provides ions with countless paths to travel through. These paths make it very hard to predict where ions will move, creating unnecessary non-uniformity in the performance of synapse.
A conceptual schematic of the new design during switching
In amorphous metals, when you apply some voltage to represent data, the ions go in different directions. The stream is ever changing and it is difficult to control. However, in the new artificial neuron, you can entirely erase the data and write it again in the exact same manner.
Building Artificial Synapse
The new design uses single-crystalline silicon, in which atoms are aligned in a continuous manner that enables ions to flow predictably. To make them do so, scientists developed a chicken-wire pattern of silicon germanium on top of the wafer of silicon. The lattice of silicon germanium is quite bigger than silicon. These two materials create a tunnel-like dislocation, forming a single path for ions.
The scientists fabricated a silicon germanium chip that has artificial synapses, with each synapse having a length of approximately 25 nanometers. Applying voltage to each individual synapse exhibits less or more the same current (ions flow), with as little as 4% variation between synapses. This performance is more uniform compared to neuromorphic chip made of amorphous metals.
In order to verify the uniformity of device (a crucial factor to demonstrate artificial neural networks), scientists tested a single synapse under the same voltage over 700 cycles. It exhibited the same current with just 1% variation.
A three-layer Multilayer Perception Neural Network
The team has also performed some practical tests like recognizing handwriting samples. The neuromorphic chips would have ‘input/hidden-layer/output neurons’, each attached to other neurons through silicon germanium artificial synapses. These neural networks can learn certain patterns, the same way human brain does.
They ran artificial neural network simulation with 3 neural layers attached through 2 artificial synapses layers. They executed a handwritten dataset containing tens of thousands of samples, and found that the network successfully recognized those samples 95.1% of the time.
The development of artificial synapses opens an avenue to realize fully functional large neural networks beyond the traditional von Neumann computing algorithm. In addition, they meet the properties required for digital non-volatile memory.
Ultimately, researchers are looking to carry out the recognition task in reality, not in simulation, and perform other complex operations that are only possible with massive supercomputer.