Evolving TinyML—The First Neuromorphic Analog Signal Processor Beside a Sensor
Israeli Polyn’s neuromorphic IC has been successfully packaged and evaluated, taking TinyML, or tiny machine learning, technology another step to becoming more mainstream.
TinyML, or the optimization of Machine Learning (ML) models to run on resource-constrained devices, is one of the fastest emerging subfields of ML. To achieve this ultra-low-power, high-performance computing needed by TinyML (or sometimes called TinyAI), engineers have explored many new and exciting technologies.
Overview of TinyML's place in edge computing. Image used courtesy of Signoretti et al
Capitalizing on this trend, this week, Israeli company Polyn announced that its newest TinyML/TinyAI processor, a neuromorphic analog signal processor, has successfully been packaged and evaluated.
In this article, we’ll take a look at the technology offered by Polyn to see the impact it may have on TinyML as a whole.
Neuromorphic Computing for AI
In the pursuit of lower power, higher performance AI computing hardware, one of the exciting emerging technologies is neuromorphic computing.
The concept of neuromorphic computing is that the human brain is the most power-efficient computing device known to man. When trying to run AI applications, it would be advantageous to create computing hardware that mimics the biological processes in the brain as closely as possible. Though it sounds like a daunting task, engineers can attempt this recreation through a combination of hardware and software.
Implementation of a neuromorphic solution. Image used courtesy of Balaji et al
From a hardware perspective, a neuromorphic chip seeks to imitate the brain through circuit elements that act as neurons, axons, and the weighted connections between them.
To further mimic the brain, this hardware is often implemented via analog circuitry, which also helps improve performance and power efficiency. Neuromorphic computing then relies on specialized neural networks such as spiking neural networks and electric signal modulation to mimic the brain signal variation.
With this basic understanding in mind, let's look at Polyn's new technology.
Polyn’s NeuroSense and NASP Technology
This week, Polyn made headlines when it announced that its proprietary neuromorphic computing chip called NeuroSense had been packaged and evaluated for the first time. Called its neuromorphic analog signal processor (NASP) technology, the NASP technology is designed to be a real-time edge sensor signal processor.
The NASP demo chip. Image used courtesy of Polyn
According to Polyn, their technology leverages a unique platform that takes a trained neural network as an input and uses mathematical modeling to synthesize the network into a true neuromorphic chip. Its NASP chip utilizes analog circuitry, where neurons are implemented using operational amplifiers, while axons are implemented by thin-film resistors.
They claim that the synthesized chip produced by its platform comes fully laid out and ready for fabrication.
NASP design process. Image used courtesy of Polyn
This newly packaged and evaluated NeurorSense chip was implemented in a 55 nm CMOS technology. In addition, it is said to behave as an edge signal sensor capable of processing raw sensor data using neuromorphic computing without any digitalization of the analog signals.
For this reason, the company is calling it the first neuromorphic analog TinyML chip that can be used directly next to the sensor without the need for an analog-to-digital converter (ADC).
While much of the technical specs are unknown, it is stated that, for always-on applications, Polyn's NASP offers a power consumption of 100 uW with "double the accuracy" of conventional algorithms.
Taking TinyML Chips to the Future
For now, Polyn is encouraged by its developments, stating that the successful packaging and evaluation of its chip validates its technology and the NASP system as a whole. In the future, Polyn says that it hopes to provide the chip to customers in the first quarter of 2023 as a wearable device coming integrated with photoplethysmography (PPG) and inertial measurement unit (IMU) sensors.