Polyn Looks to Speed ML Adoption at the Edge

Article By : Saumitra Jagdale

Mimicking brain's behavior makes neuromorphic chips the best candidates for achieving efficient but powerful computing at the edge.

Fabless semiconductor company Polyn has announced the availability of neuromorphic analog signal processing (NASP) models for Edge Impulse, a development platform for machine learning on edge devices. Edge impulse provides a way for developers to compare models and their performance, and Polyn is making its models available on the platform to enable such evaluations, CEO and founder Aleksandr Timofeev said in an interview with EE Times Europe.

“Polyn is comfortable with this comparison, as it is confident in its promise of offering chips that consume 100 microwatts of power, and no other competitor offers the same,” said Timofeev, adding that the company pays a licensing fee to make models available on Edge Impulse.

Current ML implementation methods rely on digitizing the generated data and then running them through digital ML frameworks, a process that involves considerable computational power. Processing raw sensor data in analog form can lead to decreased power consumption and increased accuracy for all applications compared with traditional, digital algorithm-based computing, Timofeev said. This is achieved by designing analog circuitry that mimics neurons in the brain, with the connections between those neurons based on the appropriately trained ML models.

“The brain is an ultra-low–power analog computer,” Timofeev noted, and mimicking its behavior makes neuromorphic chips the best candidates for achieving efficient but powerful computing at the edge.

Timofeev explained the technology behind NeuroVoice, the Polyn model that will soon be available for evaluation on Edge Impulse. The current noise-canceling products on the market are based on DSP algorithms that focus on noise reduction, “whereas Polyn’s chips are based on ML models that are trained to detect the human voice and extract it from the audio input,” making them more accurate than DSP-based solutions, he said.

NASP voice extraction technology

Neural nets to neuromorphic chips

Polyn’s platform converts trained neural networks into analog neuromorphic chips. In essence, models are fed to the platform, which then generates the files necessary to build the neuromorphic chips. This speeds chip development by eliminating the need to design chips from scratch, Timofeev said. The platform “is not currently open source or licensed to third parties,” he said; rather, “it exists as a proprietary technology that the company utilizes to design [its] chips.”

The products designed via NASP platform will “fall into three major categories,” Timofeev said: “one for voice extraction, one for vibration signal processing that would help in predictive maintenance, and a verbal sensor for deployment in smart watches and health-monitoring devices.”

Polyn Looks to Speed ML Adoption at the Edge
NASP chip (Image source: Polyn)


The target applications for the NASP chips have a common theme: They all require always-on data processing, which warrants low power consumption. By design, the Polyn chips run on ultra-low power, on the order of 100 microwatts. These applications also require real-time data processing, which is enabled by the chips’ low latency, on the order of a few microseconds. Timofeev said Polyn’s neuromorphic chips are “built with the application in mind and hence always have adequate processing power as compared to fixed designs of digital frameworks,” which may have unused capacity in some cases and inadequate processing power in others.

Voice extraction

One of the main applications of the NASP technology is voice extraction. In the case of noise-canceling headphones, the NeuroVoice technology’s ability to isolate the human voice allows the user to listen to conversations without background interference. The voice-extraction feature also enhances the experience of people using hearing aids. The chips can just as efficiently extract music from the environment, allowing users to focus on whatever they choose.

Predictive maintenance

Industry 4.0 has deepened and widened the scope of AI and ML applications, such as the power-hungry field of predictive maintenance. In this application, sensors collect massive amounts of data on machine parameters and run them through ML models to predict the life of a machine and help schedule maintenance. Neuromorphic chips can be loaded with ML models that process this generated data without sending them to data centers, minimizing costs, Timofeev said.


Smart wearables require always-on data processing; indeed, under some circumstances, failure of the always-on capability could prove dangerous to the wearer. Hence, these devices need reliable performance and extended battery life. Because of their ultra-low power consumption, and the way they process data, the Polyn devices last longer and are more accurate than their digital counterparts, according to Timofeev.

Polyn Looks to Speed ML Adoption at the Edge
(Image source: Polyn)

The future

Timofeev said the company intends to license the Polyn platform to third-party companies in the next three years as the technology matures. It also plans to develop a hardware family based on its NASP technology and is looking for development partnerships to achieve that goal. The company describes its biggest challenge as one of perception as it works to generate “awareness about technology that can process voice while consuming close to 100 watts of power.”


This article was originally published on EE Times Europe.


Leave a comment