Neuromorphic Chips Increasingly Mimic Brain Plasticity
How close are we to reverse-engineering the human brain's plasticity? New research in computational neuroscience and nanotechnologies indicates that neuromorphic chips can read synaptic connectivity with greater sensitivity.
Replicating the human “connectome,” or the physical wiring of the nervous system, remains one of the main challenges of neuromorphic computing. Despite scientists' increased understanding of how the brain's wiring executes higher-level functions, reverse-engineering the brain on solid-state devices (SSD) is still out of reach.
Key elements of a neural network. Image used courtesy of Neuromorphic Computing and Engineering
Neuromorphic computing attempts to appropriate the brain’s biological connectome—specifically, by digitally copying biological neuronal network models onto electronic devices. Neuro-electronic interfaces are brain-computer interfaces that transfer information between the brain and external devices. Currently, no adequate neuro-electronic interfaces exist, making such digital brain mimicry hard.
However, replicating the human neuronal network is a more viable reality with recent development in ferroelectric devices, nanowire networks, organic materials, and new memory hardware.
Artificial Neural Networks and Natural Neuronal Networks
So far, the approach to neuromorphic engineering and computing has bifurcated into two main directions: artificial neural networks (ANNs) and natural neuronal networks (NNNs).
ANNs are well-known as the basis of machine learning and AI applications, whose function is best realized with CPUs, GPUs, and deep-learning processors — NPUs (neural processing units) and TPUs (tensor processing units).
Analog-aided ANNs may not be “naturally intelligent” but are considered neuromorphic because their design is inspired by the brain’s in-memory computing.
An example of artificial neural networks is Intel chips Loihi 1 and Loihi 2. Loihi 1 and 2 and the associated LAVA API have provided a new mechanism for information coding and processing based on spiking neural networks (SNN) algorithms that work in an event-based asynchronous fashion using parallel sparse computing.
Loihi 2 chip. Image used courtesy of Intel
Natural neuronal networks are powered by electrochemical reactions, similar to the brain synapses. They show superior deductive capabilities, learning autonomy, adaptation, and cognition. The manner in which NNNs organize information is quite peculiar and disparate from ANNs. For example, early models of NNNs were capable of imitating early visual processing with a silicone retina model based on brain sensory peripherals.
The Human Brain Project Tackles Large-scale Neuromorphic Computing
Large-scale neuromorphic computing platforms that are a part of the Human Brain Project (HBP) are still in the testing phase—although certain chips have been available to early adopters for testing since 2018.
The architecture of the HBP is built upon two principles: first, a many-core “SpiNNaker machine” based in Manchester, UK. This machine includes one million Arm processors organized in a spiking neural network. Second, a physical model machine located in Heidelberg with analog electronic models, which consist of four million neurons and one billion synapses on 20 silicon wafers.
As a short-term benefit, the HBP expects its neuromorphic platform to improve speech and recognition capabilities in smartphones. In the long term, it can be used in various applications such as autonomous vehicles and home robots.
Samsung Develops “Copy-paste” Neuromorphic Modeling
The recent Samsung advancement goes a step further, bridging the gap between brain-inspired and brain-mimicking chips. The “copy-paste” neuromorphic chip by Samsung aims to replicate the functional synaptic connectivity map of a mammal's neuronal network onto a solid-state 3D memory.
A research team comprised of engineers from Samsung Advanced Institute of Technology (SAIT) and Harvard University developed a CMOS nanoelectrode array (CNEA) that could replicate intracellular parallel recording and capture synaptic signals with high sensitivity.
Neuromorphic scaling using 3D integration and packaging technology. Image used courtesy of Nature Electronics
This latest version of the Samsung CNEA chip that can map functional synaptic connectivity integrates 4,096 electronic channels with 4,096 vertical nanoelectrodes. The highly-sensitive nanoelectrode array can measure active potential (AP) and post-synaptic potential (PSP), find synapses, and measure their connection strength. The nanoelectrode array is the “copying” part of the model.
The determined synaptic connectivity map produced from in-vitro cultured neurons is then “pasted” onto a memory network that consists of conductive memories, where each memory stores a conductance that replicates the biological synaptic connection.
According to the Samsung research team, the best contenders to store reverse-engineered neuronal networks are four types of memories, both volatile and non-volatile. Non-volatile memories include:
- Resistive random access memories (RRAM)
- Phase-change random access memory (PRAM)
- Spin-transfer torque MRAM (magnetic random access memory)
- Dynamic random access memory (DRAM)
One unique aspect of Samsung's neuromorphic chips is neuromorphic scaling. The three-dimensional memory packaging mimics the 100 billion neurons of the human brain. For example, a 128-layer 3D integration would reduce the computer-extracted connectivity map area from 30 × 30 cm2 to an area of 26 × 26 mm2.
Image of rat neurons on CNEA (CMOS nanoelectrode array). Image used courtesy of Samsung News
This idea is not without its challenges; although the NNN represents a snapshot of the functional connectivity at the moment of copying, it still keeps a stable structure for well-defined tasks over generations.
But a question remains whether the complex synaptic activity can be represented by memory conductance values. Additionally, neuron protein membranes are poles apart from silicon circuits.
The Future of Neuromorphic Engineering and Computing
Despite the growing optimism, neuromorphic computing raises questions of practicality, as well as ethical, social, and legal issues.
The Neuromorphic Computing and Engineering journal recently published a 2022 roadmap for the future of this field. In it, researchers note that a major driver for neuromorphic computing will be developing new calculations for complex data sets at low power.
Research in neuromorphic hardware, algorithms, and materials may open computing opportunities for a span of applications: self-driving cars, robotics, biohybrid systems, and regenerative medicine to name a few.
While non-volatile memories are at a solid maturity level, the same does not ring true for nanomaterials. In addition, neuromorphic circuits—including photonic and quantum computing—are still hotbeds for new research.