Technology can help to build small com­put­ers inspired from the brain, and so here is the moti­va­tion for my work. Making our proces­sors small­er and faster has turned them into ener­gy ogres. If we were to increase their speed, they would self-destruct, heat­ing local­ly as much as a nuclear reactor.

If we want to con­tin­ue increas­ing the per­for­mance of our com­put­ers, we need to rethink the way we com­pute. And our brains are won­der­ful proof that impres­sive com­pu­ta­tions can be car­ried out with a very low pow­er bud­get. We are able to rec­og­nize peo­ple we bare­ly know in just a frac­tion of a sec­ond, and we do this much faster than any super­com­put­er while con­sum­ing one thou­sand times less pow­er. Therefore it is very rel­e­vant to take inspi­ra­tion from our brain to build fast and energy-efficient cog­ni­tive com­put­ing hardware. 

But then, what is the dif­fer­ence between our brains and our com­put­ers? Well, every­thing. Their build­ing blocks, their archi­tec­ture, and their prin­ci­ples of com­pu­ta­tion are dif­fer­ent. Our proces­sors are made of bil­lions of tran­sis­tors. So inside, mem­o­ry and pro­cess­ing are spa­tial­ly sep­a­rat­ed. And data is con­tin­u­ous­ly trans­port­ed back and forth, which con­sumes a lot of pow­er. In addi­tion, our com­put­ers are sequen­tial. They do most­ly one thing at a time, slow­ing them down.

In com­par­i­son, in the brain, there are a hun­dred bil­lion neu­rons and one thou­sand times more synaps­es. Memory, pro­vid­ed by the synaps­es, is entan­gled with the pro­cess­ing pro­vid­ed by neu­rons, which all work in par­al­lel. And so this entan­gle­ment and this par­al­lel pro­cess­ing are two rea­sons the brain is low-power and fast.

So, our com­put­ers are excel­lent at pre­cise arith­metic. Our brains excel at cog­ni­tive tasks. So how can we build cog­ni­tive hard­ware that will be com­pat­i­ble with our cur­rent com­put­ers? How can we catch the essence of bio­log­i­cal synaps­es and neu­rons, and map it to hardware? 

For synaps­es, there is some con­sen­sus. We know that it is very impor­tant to repro­duce their plas­tic­i­ty. In oth­er words, their abil­i­ty to recon­fig­ure the strength with which they con­nect to neu­rons. Indeed, we know that it is this plas­tic­i­ty that allows mem­o­ries to be formed and to be stored in the brain. So a chip with plas­tic synaps­es could learn and adapt to the chang­ing world. For exam­ple, the exact same chip could learn to detect dis­eases if pro­vid­ed with med­ical infor­ma­tion, or to detect traces of pol­lu­tants if pro­vid­ed with water com­po­si­tion. This would be very useful.

So, how can we fab­ri­cate elec­tron­ic nano-synapses, and [not] research plas­tic­i­ty? It is very inef­fi­cient to imple­ment these prop­er­ties with tran­sis­tors. There are some sub­stan­tial efforts to devel­op nov­el mate­ri­als and nov­el physics to emu­late arti­fi­cial nano-synapses. The best solu­tion today are nano-resistors called mem­ris­tors.” These nano-fuses can reg­u­late the elec­tri­cal cur­rent flow between between two neu­rons. If they let a lot of cur­rent flow, the synapse is active and a mem­o­ry is imprint­ed. If they let just a trick­le of cur­rent flow, the synapse is inactive.

Researchers know how to fab­ri­cate tens of these mem­ris­tors. But the chal­lenge is to build dense arrays of hun­dreds of mil­lions of these arti­fi­cial nano-synapses, and to con­nect them to elec­tron­ic neu­rons. And the main trend today to fab­ri­cate these elec­tron­ic neu­rons is to assem­ble tens of tran­sis­tors, which takes a lot of space on silicon.

So how can we fab­ri­cate nano-neurons? Well, the approach we are using in my lab­o­ra­to­ry is to devise tiny elec­tron­ic oscil­la­tors. Indeed, the brain has its own rhythms, and we know that for some com­pu­ta­tions, neu­rons behave as small oscil­la­tors that vibrate dif­fer­ent­ly depend­ing on the sig­nals incom­ing from oth­er neu­rons. We want to build chips that can learn and process infor­ma­tion through the com­plex vibra­tions of these elec­tron­ic oscil­la­tors, cou­pled with mem­ris­tor synapses.

Are you famil­iar with djem­bes, the African drums? Well, when the drum­mer strikes the edge or the cen­ter of the drum, the skin vibrates dif­fer­ent­ly, pro­duc­ing a dif­fer­ent sound. So you could say that these drums are able to clas­si­fy. By lis­ten­ing to the sound, which is a result a the com­pu­ta­tion, you can know if the drum­mer has struck the cen­ter or some­where around the edge. So, our chips would work in a sim­i­lar way, but with nan­otech­nol­o­gy and pro­duc­ing much more com­plex and crazy vibra­tions, allow­ing much more com­put­ing power.

So, this is a field of research that is still in its infan­cy. But if we suc­ceed, we could make sense of Big Data with a very low ener­gy cost. We could give robots a small embed­ded brain, to make them more autonomous. And we could build smart pros­the­ses for med­ical appli­ca­tions. And in addi­tion, by try­ing to build sys­tems inspired from the brain, we could learn more about the brain itself.

So the ques­tion I would like to ask you today is, what would you do? How can you use in your every­day life or in your busi­ness, a small chip that can do pow­er­ful pat­tern recog­ni­tion with a very low pow­er budget?

Thank you.

Help Support Open Transcripts

If you found this useful or interesting, please consider supporting the project monthly at Patreon or once via Cash App, or even just sharing the link. Thanks.