Artificial Intelligence (AI) is omnipresent in the cloud, but soon, we will hear and see more about leveraging AI sensors and devices at the network edge. Before this becomes a pervasive reality, however, significant barriers must be removed. These include:
- Expensive and large data sets for training do not account for the fact that the physical world has widely varying conditions, which would trigger expensive upgrade cycles.
- Transferring data to the cloud requires significant power and expense, even with low power WAN protocols, which again forces expensive data transfers or charging batteries.
- For many applications, network latency and availability limits real time response for robotics, drones, and surveillance.
Another concern is the ability to snoop, and hack networks in cars, homes, and medical de-vices, causing physical damage. One solution is to move more of the AI processing to the edge devices, but this will require a step function reduction in power consumption of the plat-form deploying the AI. In fact, the AI algorithms running on the devices will need to be re-thought to achieve this level of improvement in efficiency.
One solution is using event driven processor technology running spiking neural networks (SNN). These platforms consume energy only when switching, and automatically sleep when no data is present. These SNNs are generalizable and can operate on multiple modalities thus enabling multiple customer specific applications including sensor fusion. Another important feature is that these models can exploit parallel computation but operate individually at low speed which allows our processor to run at the lowest possible voltage. This has a huge impact on reducing the power required for a deploying an AI solution at the edge. Our current suite of voice processing solutions demonstrates this always-on logic technology for sensors and devices at the edge.