Blog

Artificial intelligence (AI) is really good at making pictures. From fake photos to phony paintings, they can look almost real. But AI is not very good with videos – and that’s because they can’t keep track of time. Engineers from the University of Michigan tackled this problem by imitating the human brain!

Consider one brain cell: this cell, called a neuron, gets signals from other neurons that change its electrical charge. If enough electrical charge builds up, our neuron fires its own signal. But if not, then the built-up charge simply fades away. Our neuron relaxes.

Not every neuron relaxes in the same amount of time. Having a mix of slow-relaxing and fast-relaxing neurons allows our brain to sense time. Now, engineers have created an electronic part that acts like a neuron. It’s called a memristor.

Memristors can change their resistance, which is how easily electricity flows. When a memristor gets an electrical signal, it lowers its resistance so that it’s easier for the next signal to pass through. But – like a neuron – if the signals stop, then the memristor “relaxes” back to its original resistance. The team invented both slow-relaxing and fast-relaxing memristors, to match the neurons in a brain.

Putting memristors inside computer chips would make AI programs a lot better at working with video. And that means we could use smaller, low-power computer chips to do the job. “Our brand-new material system could improve the energy efficiency of AI chips six times over,” says Sieun Chae, an engineer on the team. And with all the ways we’re using AI, these savings will add up!