A New Spiking Neural Network-Based Chip Shows Promise for Preventing Drone Collisions
Described as the first of its kind, Imec's new SNN-based chip is designed for anti-collision radar systems for drones.
Imec has announced a new chip that uses spiking neural network technology, which the company claims mimics the way “groups of biological neurons operate to recognize temporal patterns.” Purposed to process radar signals—specifically for drone anti-collision—the Imec chip is said to use one one-hundredth the power and one-tenth the latency of other chips.
Imec’s spiking neural network-based chip. Image (modified) used courtesy of Imec
Usually, artificial neural networks (ANNs) are used in radar-based automotive anti-collision systems. But ANNs are power-hungry devices that don't ameliorate latency issues. This latency is caused by data signals traveling to the seat of the AI inference algorithm for analysis and decision making.
What are Spiking Neural Networks?
As described by Imec's Ilja Ocket, "SNNs operate very similarly to biological neural networks, in which neurons fire electrical pulses sparsely over time, and only when the sensory input changes. As such, energy consumption can significantly be reduced.”
Thus, current "spikes" only when activity spikes—namely, when something needs to be registered and analyzed. Otherwise, only minimal power is required.
The system functionality of SNNs. Image used courtesy of Laxmi R. Iyer and Arindam Basu
As for latency, a lot of the “thinking” takes place right on the spot. As described by AAC contributor Chantelle Duboise, spiking neural network (SNN) sensor elements communicate directly with other sensor elements. These, in turn, react partially on the basis of their own levels of excitation. So a lot of the processing takes place locally, inside the sensor. Time-consuming trips to the AI engine are minimized, slashing latency.
Further, Imec claims that what is described as the chip’s “spiking neurons” can interact in a manner that can digest information, learn temporal patterns, and remember them. Thus, Imec’s design channels the chip’s behavior into reactions that have been precisely predicted by simulation.
Drones often operate independently, or at least semi-independently, of human control. In order to avoid collisions, they must make decisions about speed and direction in a matter of milliseconds, so there is no room for latency.
Drones often use on-board radar systems to garner the raw information necessary for the aircraft to chart a safe flying course. The ability of the SNN chip to do much of the information processing close to physical radar makes for far faster response time and better decisions.
Example of a drone vision reference design. TI emphasizes "minimal power requirements" as a key feature. Image used courtesy of Texas Instruments
Additionally, tiny drones need the same “brainpower” as larger devices, but the onboard power available to “feed” that brain is greatly constrained. These factors make drones an apt use case for the new SNN chip.
According to Ocket, “One scenario we are currently exploring features autonomous drones that depend on their on-board camera and radar sensor systems for in-warehouse navigation, keeping a safe distance from walls and shelves while performing complex tasks.”
Other Use Cases
Before the company zeroed in on radar applications, the initial targets for this device was speech processing in power-constrained applications and electrocardiograms (ECG). The chip’s generic architecture allows it to be reconfigured for radar, sonar, and LiDAR.
Imec’s new SNN chip is part of a trend in artificial intelligence to “front load” the system. Just as in business organizations, these networks arrange for decisions to be made closer to the seat of the action.
In the case of a drone, if the decision to turn or change speed is made physically closer to the radar or LiDAR device, the device can cut latency and save time. As such, Imec anticipates the new device as a useful addition to autonomous factor vehicles and cobots, where fast, infallible decisions are mandatory for safety.