Dutch neuromorphic computing startup Innatera Nanosystems has completed a seed funding round, raising €5 million (around $6 million).
Innatera, a spin-out from the Delft University of Technology, is developing an analog chip designed to run spiking neural networks, a type of neural network often used in neuromorphic computing that is inspired by the way the brain works. Like other neuromorphic computing approaches, the benefits are dramatic improvements in power consumption and latency – Innatera claims its chip will allow sensor data to be processed 100x faster and with 500x less energy than using conventional digital processing.
Innatera CEO Sumeet Kumar told EE Times that the company is targeting the sensor-edge, that is, applications inside or very close to the sensor, where processing is always-on and power budgets are tight.
“A number of [neuromorphic] companies target cameras and vision applications today, however, neuromorphic compute has a far wider application scope across sensing: microphones, radars, lidars, ultrasonic,” Kumar said. “There is vast potential for value addition in sensing in general, and we’re working in many of these areas with solutions that outperform conventional implementations.”
Applications that Innatera intends to target include intelligent speech processing in human-machine interfaces, vital signs monitoring in wearable devices, target recognition in Radar and Lidar, and fault detection in industrial and automotive equipment.
Innatera’s technology spans both neuromorphic algorithms and hardware. The company’s analog chip will process spatial and temporal patterns in sensor data.
“Our hardware is built to run neuromorphic spiking neural networks with a high degree of temporal fidelity,” said Kumar. “Innatera’s chip is a programmable array of analog-mixed signal spiking neurons and synapses. The architecture allows the fine-grained temporal processing capabilities of these components to be leveraged in a highly flexible manner. The architecture itself is inherently sparse, event-driven, and massively parallel.”
Innatera’s analog hardware is the key to low power dissipation, but it requires the development of dedicated spiking neural networks. These networks cannot be derived from mainstream neural network algorithms, but they are typically many times smaller than their conventional counterparts. For example, Kumar said that in a recent development carried out with an unnamed customer, Innatera’s spiking neural network outperformed a conventional neural network implemented on a “state-of-the-art analog accelerator” by a factor of 40 in terms of latency, and a factor of 49 in terms of energy per inference.
There are a handful of companies (Prophesee and others) working on neuromorphic hardware today, though most are developing event-based camera systems. Kumar said that Innatera’s approach will differentiate it from its competitors in this space.
“Our silicon architecture is built for performance scalability, robustness and flexibility, not traditionally something that competing neuromorphic approaches can deliver within the power envelope of the sensor-edge,” he said. “Perhaps most important of all, we’ve been working with customers since our inception, and our architecture was built with tight feedback from them.”
Existing customers had made Innatera a revenue-funded operation until now; the injection of funds from the seed round will be used primarily for R&D.
“We’re scaling up our team with more analog and digital designers to accelerate the development of our product chips, as well as to extend the capabilities of our SDK,” said Kumar. “Concomitant to this, we’re working with a number of customers on applying our spiking neural network approach to some very cool use-cases.” The company, however, declined to explain what such use-cases are.
Samples of Innatera’s neuromorphic chip will be available to early access customers in the second half of 2021, Kumar said.
- EE Times Special Project: On a Path to Artificial Brains
- Neuromorphic Promises Better AI
- Sony, Prophesee Open ‘Pandora’s Box’ in AI Sensing
- Will Neuromorphic Computing Compete with Today’s AI Accelerators?
- Prophesee Touts Toolkit for Event-based Vision
The post Neuromorphic Startup Targets Applications Beyond Cameras appeared first on EETimes.