If your nervous system learned the way most artificial networks do, you would need a nuclear power plant wired to your head. Max Talanov says so, and then jokes about why you don't have one: "Mom forgot?" No, he answers himself, mom was fine. You just don't need it.
The standard training method, backpropagation, rewrites every weight in the network on every pass. A real nervous system is thriftier. It changes synapses slowly, and only the ones that have to change. The rest it leaves alone.
Talanov is a researcher at the Neuralesi Clinical Research Center in Messina and a senior researcher at an AI institute in Novi Sad. He builds spiking networks, which fire like neurons instead of crunching numbers. His bet is that this thrift can be copied in hardware, and the payoff he points to is a body: a chip implanted in a paralysed person that helps with neurorehabilitation or returns lost functions.
The part that makes this plausible is the memristor, a resistor with a memory. Its resistance depends on the current that has already flowed through it, so it behaves like a synapse that remembers its past. It spends energy mostly at the moment it changes state.
Talanov first read about memristors in 2012, in a popular science magazine, and was struck by the reader comments asking how anyone would program such processors. Why program them at all, he thought. They work like a brain, so they can learn alongside you. When he later met the memristor researcher Viktor Erokhin, his first proposal was to build a mouse brain. Erokhin was horrified, Talanov says, and agreed anyway.
The obstacle was scale. Memristors are erratic devices. The high-resistance state comes out different in every experiment, which makes them poor flash drives. For years nobody could make chips with more than a handful. Then came chips with half a million, and a million.
Here Talanov gives his numbers. Prototypes have shown that a million memristors run on microwatts. "Microwatt," he stresses. "Not megawatt." That is several orders of magnitude less than conventional hardware, and it moves toward the roughly 20 watts a human brain uses.
A million synapses, by his count, is about the synapse volume of an entire biological spinal cord. Hence the line he wants people to remember:
We can potentially build the whole spinal cord on a chip that draws microwatts. Already now. Мы потенциально можем реализовать весь спинной мозг в чипе, потребляющем микроватой энергии. Уже сейчас.
— Max Talanov, Think Like a Scientist, 107:30
He insists it is not science fiction. It can be done now. He is careful about the next step, though: the implant for a paralysed person is what he wants "in the ideal case," not something that exists. Neurons need building too, since they also burn energy, and he says there are memristor-based ways to make them.
His own group has taken a small step in that direction. A February 2023 paper in Frontiers in Neuroscience described a circuit of just five neurons, two of them with memristive elements. It produced electrical responses with the timing pattern of spinal responses to epidural stimulation during walking, an early burst around 10 milliseconds and a late one after 30.
Why does the wattage matter so much? Because a chip inside a body has to get rid of its own heat. A November 2024 study in PNAS Nexus put the problem plainly: for implants that process signals in real time, power and heat are the limiting constraints. It found that neuromorphic processing cut the data needed by more than three orders of magnitude.
Talanov's version is simpler. A chip that draws microwatts has nothing to radiate. It can sit at body temperature and keep working, and for a long time.
