
IMEC’s spiking neural network aids drone navigation
Spiking neural networks (SNNs) are closer analogues of the biological neural networks than common weighted-synapse neural networks. In SNNs data is moved as a series of electrical pulses over time but only when the sensory input changes. IMEC’s chip consumes 1 percent of the power of traditional artificial NNs while providing a tenfold reduction in latency.
Radar for military and avionic applications has also had a traditional requirement for high-speed and high-performance digital signal processing and is increasingly being considered for road traffic and autonomous driving.
IMEC’s chip has 336 neurons. Part of this neuron fabric is fully feed-forwaed connected, part has reconfigurable feed-back and feed-forward recurrent connections. It operates at a nominal 1.1V and was implemented in 40nm CMOS by TSMC. It can classify micro-Doppler radar signatures using only 30-microwatts of power although the architecture and algorithms can be applied to variety of one- and two-dimensional sensor data, including electrocardiogram, speech, sonar, radar and lidar streams. The first use case for IMEC’s spiking neural network is an anti-collision radar system for drones.
Although artificial neural networks are used in automotive industry, IMEC points out that in power-constrained environments such as battery-powered drones the power consumption is too burdensome. In addition the time taken to move data from sensor and through the AI inference algorithm is too long.
Next: World’s first
“Today, we present the world’s first chip that processes radar signals using a recurrent spiking neural network,” said Ilja Ocket, program manager of neuromorphic sensing at IMEC, in a statement. Ocket made the point that SNNs can be connected recurrently – turning the SNN into a dynamic system that learns and remembers temporal patterns.
That avoids keeping lengthy, data intensive training separate from inference. “The technology we are introducing today is a major leap forward in the development of truly self-learning systems,” said Ocket.
She said that by doing processing much closer to the sensor the radar should be able to distinguish more quickly and more accurately between approaching objects.
Kathleen Philips, program director of IoT cognitive sensing at IMEC, added: “For [the chip’s] creation, we rallied experts from various disciplines within IMEC – from the development of training algorithms and spiking neural network architectures that take neuroscience as a basis, to biomedical and radar signal processing and ultra-low power digital chip design.”
The implication is that IMEC has used near-threshold voltage circuit design in relatively-aggressive CMOS manufacturing process.
Related links and articles:
News articles:
Here comes a 77GHz CMOS radar transceiver from China
Hold your breath! IMEC 8 GHz radar will detect your heartbeat
Analog Devices’ lidar is flying under the radar
