The role that analog plays in enabling the TinyML technology at the edge to lower power usage and latency may surprise you.
There is a rapidly growing need for power-efficient artificial intelligence (AI) to run on smaller devices at the edge. Power-hungry edge devices send massive amounts of data to and from the cloud. At the same time, however, tomorrow’s AI at the edge must run on much smaller battery-powered microprocessors and microcontrollers used in smartphones and Internet of Things (IoT) devices. Eliminating the back and forth communications with the cloud will also solve the latency issue.
The answer can be found in the use of TinyML technology at the edge. What may surprise you is the role that analog plays here.
Figure 1 The block diagram shows the use of machine learning at the edge and the broad TinyML ecosystem-as-a-service. Source: MDPI
AI encompasses devices and systems where knowledge is derived from data using techniques like machine learning (ML) and deep learning. TinyML involves building complex AI algorithms into hardware located at the device or sensor. Through this integration, it takes less power for data analytics, as interaction with the cloud is removed. Given that there is no data transfer, outside attacks and security breaches are curtailed as well.
With TinyML models, only the TinyML model parameters are sent to the cloud for processing. First-time learning occurs on the server and more specific learning takes place locally.
Tiny machine learning or TinyML includes hardware, algorithms, and software capable of performing on-device sensor data analytics. This takes place at extremely low power—milliwatt range or less—supporting always-on use cases and battery-operated devices.
Now, let’s add neuromorphic analog signal processing or NASP to TinyML technology at the edge.
Neuromorphic analog signal processing
Neuromorphic computing is where system elements mimic the human brain and nervous system in hardware and software, providing faster computation and lower power consumption. Like the brain, it handles many operations simultaneously.
In April 2022, POLYN announced NeuroSense, a proprietary neuromorphic chip. The neuromorphic analog signal processor (NASP) device has been designed as a real-time edge sensor signal processor. POLYN’s NASP chip uses analog circuitry, where neurons are implemented using operational amplifiers and axons implemented by thin-film resistors. POLYN claims it’s the first neuromorphic analog TinyML chip for use directly next to a sensor without the need for an analog-to-digital converter (ADC).
Here, TinyML uses a technique called embeddings that allows AI computations to be performed directly on the chip and not on the cloud or remote server. These embeddings are representations that a trained autoencoder neural network forms in deep and hidden layers. They contain densely-packed information regarding sensory input data and their use for sensor data preprocessing; it reduces noisy raw data flow by 1,000x, making it appropriate for Industrial Internet of Things (IIoT) use cases.
The NASP chips are, therefore, TinyML implementations that solve power consumption and computing latency challenges.
Figure 2 The use of embeddings for sensor data preprocessing reduces noisy raw data flow by 1,000x, making it appropriate for industrial IoT use. Source: POLYN
It’s worth noting that digital implementation in this environment won’t process massively parallel data optimal for neural network computation. A digital neural network is a model with simulated neurons; it uses standard step-by-step consecutive math operations in the digital processor core. However, digital neural networks on traditional processors are inefficient and, although there has been significant progress in digital neural network implementation over the past two decades, energy consumption remains unresolved.
Figure 3 Power consumption is still an issue in digital neural network implementations. Source: POLYN
It’s neuromorphic analog computing that mimics the brain’s function and efficiency by building artificial neural systems, implementing neurons and synapses to transfer electrical signals specifically via analog circuit design. That’s why analog neuromorphic chips are well adapted for neural network operations.
Here, the NASP approach simplifies chip layout. It combines a fixed weights method with a fixed chip structure, akin to a human visual nerve and retina, and a flexible function that’s responsible for further classification of the received embeddings. In a hybrid concept (Figure 4), a combination of fixed neural networks is responsible for pattern detection, combined with a flexible algorithm responsible for pattern interpretation.
Figure 4 With NASPs, most neural network layers responsible for raw data preprocessing remain unchanged after training epochs, and layers are updated while receiving new data and retraining. Source: POLYN
The fundamental parts of the NASP Hybrid Core concept comprise a fixed neuromorphic analog core featuring ultra-low power and low latency and generating embeddings and a fully flexible digital core for final classification. Providing the optimal answer to the power consumption and computing latency challenges, the solution is divided into fixed and flexible parts. The fixed part is analog, and only a relatively small output, based on embeddings, is sent to flexible digital processing for analytics.
Expanding neuromorphic applications
What spurred the writing of this blog was an announcement about the collaboration between POLYN Technology and Edge Impulse to leverage NASPs, enabling design engineers to develop application-specific chips for one-dimensional signal processing at the sensor level. The goal is to accelerate the adoption of ML at the edge.
Figure 5 POLYN’s neuromorphic analog signal processor (NASP) announced 4/22 mimics the way the human brain perceives and learns. Source: POLYN
Edge Impulse brings to the table its machine learning software, enabling smarter edge products and providing robust automation and low-code capabilities. The company claims its solution makes it easier to create valuable datasets and rapidly develop advanced ML algorithms and industry-specific solutions in weeks vs. years.
POLYN and Edge Impulse are focusing initially on hearables, using POLYN’s voice detection and voice extraction from any ambient environment with the up to 150 uW power consumption, enabling designs for hearing aids, hearing assistance products, earbuds, and other miniature devices including wearables.
Next in line are use cases in industrial IoT, robotics, automated video security, voice assistants, autocorrection, image processing in cameras, autonomous driving, connected healthcare, Industry 4.0, and precision farming.
This article was originally published on Planet Analog.
Carolyn Mathas has 15 years of journalism experience, the last seven of which have been spent on the EE Times Group’s DesignLines, including PLDesignLine, Network Systems DesignLine, Mobile Handset DesignLine, and her current sites, Industrial Control DesignLine and CommsDesign. Prior to joining UBM in 2005, she was a senior editor at Lightwave Magazine and a correspondent for CleanRooms Magazine. Mathas has a BS in marketing from University of Phoenix and an MBA from New York Institute of Technology. She lives in the Sierra foothills and claims that the pine forest, snow, mountain air, bears, and power outages balance her deadline-packed high-tech career.