POLYN Technology's Eugene Zetserov
Tiny AI is a buzzword for the recent trend of reducing the size of math models and consumed power. Tiny AI is sometimes referred to as Tiny ML, which aims to miniaturize the devices and optimize machine-learning algorithms while maintaining high levels of inference accuracy.
But Tiny AI goes further.
On a practical basis, Tiny AI means an always-on solution that processes various types of raw data on-device, with little power consumption. In the new era of Tiny AI, the Industrial IoT, automotive, manufacturing, smart home, and healthcare will benefit from processing data at the thin edge with low latency and high power efficiency.
Many areas in consumer and industrial markets, especially at the thin edge, could do well with AI in general and the neural network paradigm in particular, but the so-called memory bottleneck and lack of true parallel computation are the most significant challenges neural networks face if performed in a traditional way on standard CPUs or even GPUs.
The solution is a neuromorphic architecture mimicking the human brain and sensory system. For use cases with a continuous sensor signal, special-purpose processors such as Neuromorphic Analog Signal Processing (NASP) are required.
A good example is monitoring the health of machines to predict probable failure of components and systems, a practice that is central to digital transformation and receiving much attention. Condition monitoring and predictive-maintenance solutions naturally rely on massive amounts of data collected from vibration, pressure, flow, temperature, force and torque sensors. The sensor data must be processed and analyzed by machine-learning algorithms. To send all this data to a center for analysis would be more trouble than it is worth.
Even with low-bandwidth communication technologies such as NB-IoT or LoRa, the data transmitted from a sensor, multiplied by the vast number of sensors that are planned to be deployed, would result in a traffic tsunami. This can only be addressed by on-sensor data pre-processing, for which a neuromorphic processor could extract data patterns detected by sensors for optimal machine-health monitoring.
Another power-hungry application is wearables, with heartrate tracing and human-activity recognition, where PPG/IMU sensors constantly generate data and processing consumes significant battery power. A similar example could be voice extraction for hearing aids, two-way radio, and other applications that require voice/noise optimization. And in general, for devices that require truly always-on supervised and unsupervised learning, NASP is an ideal solution with ultra-low 100uW power consumption, very low latency, and better accuracy compared to traditional algorithms.
Tiny AI addresses all these and many other use cases by smart pre-processing of raw data directly on sensor nodes, opening new opportunities for the whole industry.
Eugene Zetserov is vice president of marketing and business development at POLYN Technology
