In a new Nature Machine Intelligence study, HBP partner Dutch National Research Institute for Mathematics and Computer Science (CWI) researchers Bojian Yin and Sander Bohté demonstrate a significant step toward artificial intelligence that can be used in local devices like smartphones and VR-like applications while protecting privacy.

They demonstrate how brain-like neurons and innovative learning algorithms may train large-scale, energy-efficient spiking neural networks. Wearable AI, speech recognition, and AR are possibilities.

Modern artificial neural networks, which power the AI revolution, are only loosely inspired by biological neural networks like the human brain. However, the brain is bigger, more energy-efficient, and can react quickly to external stimuli. Spiking neural networks behave more like biological neurons, which exchange electrical pulses sparsely.

Learning brain simulation

These algorithms struggle to learn like our brains. New experiences can instantly change or create new brain connections. The brain uses less energy and requires fewer instances to learn. “We wanted to develop something closer to the way our brain learns,” explains Bojian Yin.

Yin explains: if you make a mistake during a driving instruction, you learn from it quickly. You change immediately. “You learn by absorbing fresh knowledge. We gave each neural network neuron updated information to imitate that.

The network learns how information changes and doesn’t have to remember everything. Current networks must adapt to earlier changes. Current learning needs massive computational power, memory, and energy.”

Six-million neurons

The new online learning method lets data-driven spiking neural networks grow. Bohté and Yin showed this in an item recognition and location system alongside TU Eindhoven and Holst Centre researchers. Yin displays a video of a busy Amsterdam street where SPYv4, a spiking neural network, can detect bikers, pedestrians, and autos and indicate their whereabouts.

“Previously, we could train neural networks with up to 10,000 neurons; now, we can do the same quite easily for networks with more than six million neurons,” explains Bohté. This lets us train powerful spiking neural networks like SPYv4.

Future

Then what? With such strong AI solutions based on spiking neural networks, processors are being designed that can run these AI programs at extremely little power and will be used in many smart products like hearing aids and AR/VR glasses.