Image recognition: the promise of a new approach to energy-efficient AI

To com­mu­nic­ate with oth­er cells, such as those in muscles and vari­ous glands, neur­ons emit elec­tric­al sig­nals. The amp­litude of each of these elec­tric­al sig­nals is approx­im­ately 100 mil­li­volts; a read­ing that remains con­stant regard­less of the neur­on or the inform­a­tion. How­ever, the neur­on will not emit any elec­tric­al sig­nal if the stim­u­lus it is exposed to is too weak. In oth­er words, it is the pres­ence or absence of an elec­tric­al sig­nal that con­sti­tutes inform­a­tion in itself, rather than the vari­ation in its amplitude.

“But elec­tric­al sig­nals are extremely rare,” says Simon Thorpe. “For me, what mat­ters in trans­mit­ting an image and its com­pos­i­tion from the ret­ina to the brain is, above all, the order in which the elec­tric­al sig­nals travel from the 100 mil­lion neur­ons in each ret­ina to brain cell.” The first elec­tric­al sig­nals will determ­ine the bright­est or highest-con­trast points, then the image will gradu­ally take shape, mov­ing from the bright­est to the dim­mest, until the stim­uli are too weak to trig­ger an elec­tric­al sig­nal from the neurons.

It was on this the­or­et­ic­al found­a­tion, developed as early as the 1980s by Simon Thorpe, that he built the basis for his own arti­fi­cial intel­li­gence mod­el for recog­nising objects with­in land­scapes. Indeed, the edges of an object with­in an image can be com­pared to abrupt changes in bright­ness. This makes it eas­ily iden­ti­fi­able by his mod­el. “And, if there is a face in this image, we can detect it, for example, by identi­fy­ing an upper edge – the top of the skull – a lower edge – the chin – and a middle edge – the mouth,” adds Simon Thorpe.

In 1999, Simon Thorpe co-foun­ded Spiken­et Tech­no­logy, an image-pro­cessing com­pany based on this tech­no­logy. One of its main aims was to identi­fy logos along the route of sport­ing events such as cyc­ling or For­mula 1, to help mar­ket­ing depart­ments optim­ise their ad place­ments. “It was a mar­ket that was grow­ing rap­idly at the time, both at phys­ic­al events and in digit­al spaces. Pro­fes­sion­als wanted to use this type of tech­no­logy to optim­ise the place­ment of their adverts, so that they would reach the audi­ence with s them feel­ing over­whelmed,” explains Jean-Gab­ri­el Ganas­cia, a pro­fess­or at the Sor­bonne and an expert on these issues.

From 2010 to 2017, a com­pet­i­tion well-known with­in the field, known as the ImageN­et Large Scale Visu­al Recog­ni­tion Chal­lenge (ILSVRC), brought togeth­er teams of research­ers from around the world, who sub­mit­ted their mod­els to clas­si­fy image data­bases. In 2012, a mod­el based on stat­ist­ic­al sci­ence won this com­pet­i­tion. This event helped to estab­lish this type of mod­el – which forms the basis of LLMs (Large Lan­guage Mod­els) – as the dom­in­ant approach on the glob­al stage. “Indeed, mod­els based on stat­ist­ic­al sci­ence were then con­sidered to have a lower mar­gin of error than those based on neur­os­cience,” explains Jean-Gab­ri­el Ganascia.

“For mod­els like mine, the momentum had passed,” says Simon Thorpe. “But – as recent events clearly show us – we are real­ising that these mod­els are very energy-intens­ive.” This is mainly due to the astro­nom­ic­al quant­it­ies of images they must pro­cess dur­ing the train­ing phase to learn to dis­tin­guish a sauce­pan from a drum. “But that’s not how human beings work. Oth­er­wise, 10 years of edu­ca­tion wouldn’t be enough to pass your A‑levels!” exclaims Simon Thorpe.

Simon Thorpe’s mod­el oper­ates in two suc­cess­ive phases: first, the iden­ti­fic­a­tion of an image based on the order of the dis­charges rather than their amp­litude, in order to con­struct the mean­ing of an image. Next, this image is stored in a sys­tem known as ‘silent neur­ons’. These are the digit­al equi­val­ents of inact­ive neur­al con­nec­tions in the brain that learn and then go into standby mode – to pre­vent the brain from becom­ing over­loaded. They can then be react­iv­ated when the image reappears before the eyes.

In fact, it fol­lows in the tra­di­tion of so-called ‘frugal’ AI mod­els, which seek to achieve their object­ives with the min­im­um of resources, rather than with max­im­um pro­cessing power. For example, the mul­tina­tion­al Microsoft and the French com­pany Mis­tral AI have also released their own frugal AI mod­els: Phi 3 and Mis­tral 7B. These mod­els train on lim­ited data­sets and are based on stat­ist­ic­al science.

Guénolé Boillot