Emergence of Neuromorphic Chips to Mimic Human Brain

Article By : Maurizio Di Paolo Emilio

A neuromorphic chip can mimic the brain to process data effectively, far surpassing existing machines, which struggle to accommodate the demands of big data, AI, and machine learning. Neuromorphic chip processing is also expected to play a crucial role in non-medical areas...

The electrical properties of biological cells have long been studied to understand intracellular dynamics. The difficulty of measuring microscopic parameters that control the dynamics of ionic currents and the nonlinearity of ionic conductance have hindered efforts to construct quantitative computational models. The growing attention paid to implantable bioelectronics for the treatment of chronic diseases is driving technology toward low-power solid-state analog devices that accurately mimic biological circuits.

Image: Ceryx

The human brain processes information and stores it instantly through more than 100 billion neurons. The neurons communicate with each other through more than 100 trillion synapses that are connected in parallel, allowing the network to perform memory, computation, reasoning, and computing simultaneously at low power (about 20 W).

Neurons determine signals as part of networks that produce collective oscillation patterns that are extremely sensitive to the neurons’ properties. One objective of neuromorphic chips is to be able to integrate nonlinear electrical characteristics and offer shallow power with the ability to process a considerable volume of signals in real time.


Recommend

Accelerating Your Machine Learning Compute for the IoT and Embedded Market with Arm


Solid-state neurons implemented through a microelectronic layout respond almost identically to biological neurons under stimulation from a wide range of current injection protocols. The optimization of nonlinear-equation models demonstrates an effective method for programming analog electronic circuits. This approach offers a way to repair diseased biocircuits and emulate their function with biomedical implants that can adapt to biofeedback.

Neuromorphic chips represent a promising technology for implanted brain-machine interfaces, with many research projects now under way. An application example includes solutions to improve visual prosthetic systems or deep brain stimulation settings. “Neuromorphic chips are low-power and compact, and they potentially can adapt, through on-chip online learning circuits, to the changes that the body undergoes with time,” said Giacomo Indiveri, director of the Institute of Neuroinformatics at the University of Zurich and ETH Zurich. “Typically, neuromorphic chips are designed to couple to the neural circuits they are being interfaced to by using the same dynamics and then detecting anomalies in the activity of the neural populations they are talking to” — for example, to detect the onset of a seizure.

Neuromorphic chips work similarly to the human brain, conserving energy and working only when needed. Many researchers and analysts believe that these chips are likely to be the future not only for artificial intelligence but also for developing low-energy cryptographic evaluation systems.

“The chips would meet the needs of [patients with] degenerative neuron disease by substituting diseased biocircuits with synthetic ones,” said Alain Nogaret, a professor of physics at the University of Bath. Nogaret is part of a research team that worked with cardiologists to show that “neuron chips can reverse the effects of heart failure by restoring the function of … respiratory neurons” at the base of the brain.

In a paper1 on their findings, the authors state their reluctance to extrapolate the results to other diseases, “as the only extensive trials we have conducted on animal models of disease so far are animal models of heart failure.” But according to Nogaret, “diseases that come to mind” as candidates for the approach include “Alzheimer’s disease and diseases of ion channels in the neuron membrane (channelopathies). Epilepsy [patients] could also benefit, as some forms of epilepsy reported in the literature are associated with specific ion channels.”

Indiveri said, “Another application area in this domain that is more mature is cochlear implants. The main advantages of neuromorphic chips are their lower power consumption, their compactness, and their potential for ‘speaking the same language’ [as the] spiking neurons they are being interfaced to, i.e., that of action potentials and neural dynamics.”

Neuromorphic chips could also be used to listen to motor neuron activity and to decode the expected muscle-activation pattern (e.g., for control of a prosthetic device).

The chips are analog devices, typical of nonlinear dynamic systems. This means they read raw nerve signals and output neuron oscillations as analog voltages. “No ADC/DAC is needed,” said Nogaret. “Asynchronous chips can thus integrate complex, noisy synaptic inputs in real time. The main difficulty lies in conditioning the chip — through its parameters, gate biases, etc. — to respond identically to a specific type of biological neuron. This is the purpose of the parameter estimation methods that our lab and others are beginning to develop.”

The chip must be minimally invasive in terms of biocompatibility, adapting to signals with virtually zero power consumption by maximizing the use of energy-harvesting sources. The design constraints are the same as those imposed on the electronic circuits and systems currently being used in implants such as cardiac pacemakers. “To a large extent, these conditions are already accessible, thanks to the 60-or-so years’ experience we have in producing VLSI circuits,” said Nogaret. “The challenge of optimizing these chips further to optimize bioimplants will be met through incremental engineering progress and feedback from targeted trials.”

VLSI circuits realized with CMOS technology are a strategic technology for the development of digital systems; continual increases in microelectronics integration have enabled systems of rising complexity.

The development of VLSI systems has resulted in highly specialized technologies. Integration at the packet and chip level is more practical for the implementation of VLSI systems because of their compact size and short signal interconnection. The growing complexity of chips creates a need for improved design methodologies and more powerful CAD environments. Further research and development are required before complete adoption is possible, however.

Figure 1: This sewing-machine–like robot inserts electrodes into the brain (Image: Neuralink)

Among the companies working on application cases, Ceryx Medical is developing biolectronic central pattern generators (CPGs) to imitate the body’s nerve centers. CPGs produce rhythmic outputs in the absence of rhythmic inputs. In medical applications, the devices could help control involuntary and voluntary rhythmic processes such as peristalsis, heart rate, and even gait, restoring proper functioning when natural rhythmic processes have been impaired by disease or injury.

Startups Neuralink and Paradromics are also working to optimize neuromorphic solutions. Neuralink is building an implantable wireless system that has far more electrodes so that it can record signals from more neurons (Figure 1).

Paradromics is bringing to market the first high-data-rate brain computer interface (Figure 2). The implantable system can be used for practical health-care applications by vastly increasing data rate, portability, and durability. The startup is focused on enabling an even higher density of probes over the face of its neural implant by integrating more, smaller electrodes.

Figure 2: The Paradromics system (Image: Paradromics)

Future challenges for neuromorphic devices are to increase the efficiency of the response and the improvement of the model through deep-learning tools, with the aim of transforming the brain into an increasingly digital one. The key application for such solutions is a digital cure for Alzheimer’s disease and other cognitive disorders.

A neuromorphic chip can mimic the brain to process data effectively, far surpassing existing machines, which struggle to accommodate the demands of big data, AI, and machine learning. Neuromorphic chip processing is also expected to play a crucial role in non-medical areas, including voice/face recognition and data mining, learning accurately from evolving data.
This article was first published on EE Times Europe

Leave a comment