Brain chip that 'learns' packs triple punch
A computer chip that can ‘learn’ could improve our understanding of how the brain works, lead to a better human-machine interface and pave the way for artificial intelligence.
The chip has the ability to mimic the brain’s plasticity – the malleable quality that allows neurons to adapt in response to new information – which is necessary for learning.
Using approximately 400 transistors, the researchers built circuits on the silicon chip that mimic sub-cellular mechanisms, in particular the way neurons change their connection strength. ‘We built the chip to have the same current and voltage levels as a real biological neuron,’ says Guy Rachmuth, one of the inventors of the chip whose PhD thesis at MIT in Massachusetts, USA, formed part of the research.
‘We were able to mimic the change in synaptic strength by changing the voltage of the circuit – we slowed it down so it [the chip] operates in real biological time, but you can potentially operate it much faster than that.’
The reason this is desirable in a computer chip is that it leads to unsupervised learning – rather than simply receiving commands from the brain. The chip can communicate back and forth, creating feedback that enables the chip to learn, adapt and improve itself.
A chip that can learn would have drastic implications for a number of fields, but most directly, it could be used to help humans interact better with machines – particularly neural prosthetic devices. ‘Because we are able to scale the current voltages to a biological level, you can imagine circuits replacing pathological tissue, so [the chip] could be used for people with spinal cord injuries for example, people with retinal implants or in the future, for those with Alzheimer’s and Parkinson’s. Where tissue is no longer useful, you could replace it with a chip that communicates with the biological tissue.’
Although there are currently neuroprosthetic products that stimulate the brain using electrical current, ‘there is no feedback between the brain and the device,’ says Rachmuth. ‘Our chip would be like a second generation [device], where you have feedback from the biological circuit back through the electronic circuit.’
The MIT team decided to work in analog when developing the transistors of the chip. They operated in the weak-inversion region, which meant they generated a low current to run low power circuits – just like a brain. ‘Using analog allowed us to avoid large currents and therefore a lot of power. We made a decision to use the analog side of the transistor equation because we wanted to work in the exponential relationship between the voltage and current, which is exactly how the brain’s ionic signals work,’ says Rachmuth.
Citing leading companies that use supercomputers to simulate small sections of the brain, he notes, ‘You can never hope to reach the complexity the brain actually has using digital computing’.
He adds that the benefit of using electronic circuits rather than digital computing is that they lead immediately into implementation, into devices and medical equipment.
In terms of fabrication, Rachmuth says that using standard CMOS manufacturing techniques means the chip would be easy to mass produce, but he concedes that it will be a while before it becomes commercially available.
In the meantime, he hopes the chip’s ability to simulate a large section of the brain will enable better understanding of the organ. ‘By mimicking the subcellular ionic channels that are part of the brain and being able to simulate some of the biochemical reactions that the neurons go through, you actually see emergent behaviour that you wouldn’t necessarily be able to with a much more simplified model of a neuron. This is where our breakthrough really is.’