Mind of an AI: from curiosity to autonomy

How does one meas­ure the intel­li­gence of an arti­fi­cial intelligence? 

It isn’t easy because the term “arti­fi­cial intel­li­gence” is used by the gen­er­al pub­lic to refer to objects developed in this research field, such as soft­ware equipped with a learn­ing sys­tem. In fact, arti­fi­cial intel­li­gence is not a thing. Rather it is a field of study which tries to mod­el func­tions of the human mind, like memory, reas­on­ing, learn­ing or lan­guage. We there­fore can­not meas­ure intelligence.

Moreover, the notion of ‘intel­li­gence’ makes no sense gen­er­ally speak­ing. For example, it’s impossible to say that an earth­worm is more stu­pid than a human. Instead, each liv­ing being has beha­vi­or­al and mor­pho­lo­gic­al char­ac­ter­ist­ics res­ult­ing from an evol­u­tion­ary pro­cess related to its par­tic­u­lar envir­on­ment. Earth­worms can find food in the ground. And, in their own eco­sys­tem, human beings have social inter­ac­tions – lin­guist­ic or cul­tur­al – with oth­ers. An earth­worm wouldn’t know what to do in our eco­sys­tem and a human wouldn’t do any bet­ter in soil.

Tech­no­lo­gies, too, are developed in very spe­cif­ic con­texts. It can­not be said that smart­phone voice recog­ni­tion sys­tems are stu­pid because they don’t under­stand the mean­ing of the sen­tences that they tran­scribe. They were not trained to that end, it’s not part of their ‘eco­sys­tem’. 

The soft­ware you men­tion tran­scribes and learns, could it also understand?

Fun­da­ment­ally, the mean­ing we asso­ci­ate with a sen­tence is per­son­i­fied, it is inter­preted based on the sens­ory and motor exper­i­ences of our body in its envir­on­ment. If a machine does not have access to a body to phys­ic­ally inter­act with our world, it stands no chance of inter­pret­ing sen­tences like we do.

How­ever, we can train lan­guage mod­els with large text data­bases. Machines can then detect stat­ist­ic­al pat­terns and do aston­ish­ing tasks, such as answer­ing a simple ques­tion, by pre­dict­ing the struc­tures of sen­tences accord­ing to a giv­en con­text. These tools are very use­ful in the world of industry, for human-machine inter­face, when machines must inter­pret dir­ect­ives based on the con­text. In order to do this, unlike humans, they don’t neces­sar­ily need to under­stand sentences.

In your research you state that, in humans, part of learn­ing is driv­en by curi­os­ity. Can it be applied to software? 

This issue is at the heart of my team’s research. We study the mech­an­isms of curi­os­ity, or what psy­cho­lo­gists call “intrins­ic motiv­a­tion”. It allows liv­ing beings to under­take inde­pend­ent learn­ing. We devel­op algorithmic mod­els of curi­os­ity to high­light the mech­an­isms involved, such as spon­tan­eous explor­a­tion. This plays a fun­da­ment­al role in the sens­ory, cog­nit­ive and motor devel­op­ment of humans.

We then test our the­or­ies with volun­teers or machines. In doing so, we dis­covered that to effi­ciently explore its envir­on­ment, a robot must pick the areas where it makes the most pro­gress, mean­ing those in which the gap between pre­dic­tion and real­ity tends to decrease. For example, it is in its interest to play with an object with which he tends to make pro­gress rather than anoth­er that he imme­di­ately mas­ters or, on the con­trary, that he can­not use at all. We showed that, in the­ory, this strategy is effi­cient for robots. The ques­tion remains open as to wheth­er humans use this meas­ure to make pro­gress and to guide the explor­a­tion of their surroundings.

But could this meas­ure of pro­gress explain why humans tend to prefer activ­it­ies that they are able to learn easily?

Yes. The mech­an­ism of explor­a­tion accord­ing to pro­gress leads to a “snow­ball” effect: when explor­ing an activ­ity, ini­ti­ated ran­domly or oth­er con­tin­gent factors, we devel­op know­ledge or skills which will make sim­il­ar types of activ­it­ies easi­er to learn. This encour­ages the indi­vidu­al to pur­sue that course of action; also asso­ci­ated with the pleas­ure response in the brain upon explor­ing new activities.

This fun­da­ment­al hypo­thes­is explains the diversity of learn­ing tra­ject­or­ies found in dif­fer­ent people. To con­firm that, we com­pared the beha­vi­or of adult volun­teers to that pre­dicted by our com­pu­ta­tion­al mod­el. These ana­lyses showed that learn­ing pro­gress and per­form­ance for each task are meas­ures used by humans to guide their explor­a­tion. Both work in dif­fer­ent ways: the com­bin­a­tion of these dif­fer­ences and the snow­ball effect men­tioned above sup­ports the idea that there is a diversity in learn­ing path­ways, which explains dif­fer­ences between individuals.

Does this mod­el improve machines?

Our the­or­ies can some­times be imple­men­ted in machines to make them more adapt­able. But the explor­at­ory beha­vi­or of humans is not neces­sar­ily the most optim­al choice. For example, oth­er curi­os­ity mech­an­isms are bet­ter suited for robots destined to autonom­ously explore the ocean floor or the sur­face of Mars – if only to pre­vent these machines from mak­ing dan­ger­ous choices as much as possible.

Can these tools also help humans to learn better?

Yes, there are indeed applic­a­tions in the field of edu­ca­tion. We designed soft­ware that cre­ated cus­tom exer­cise sequences for stu­dents in math­em­at­ics. The object­ive is to adapt a series to each child that will optim­ise both his/her learn­ing and motiv­a­tion. We know that the lat­ter is an import­ant factor in aca­dem­ic fail­ure. Motiv­a­tion­al aspects prompt stu­dents to per­severe and try harder. Using curi­os­ity mod­els, we developed algorithms that inter­act with each child indi­vidu­ally to offer a motiv­a­tion­al series of exer­cises accord­ing to the child’s profile.

In a pre­vi­ous pro­ject named Kidlearn we showed that, on aver­age, a great­er diversity of stu­dents made more pro­gress thanks to our software’s pro­pos­als than with a teach­ing expert. A fig­ure that includes chil­dren with a range of dif­fi­culties or abil­it­ies, too. This bene­fit was asso­ci­ated with a high­er degree of intrins­ic motiv­a­tion. We are now work­ing with a con­sor­ti­um of indus­tri­als in the field of edu­ca­tion tech­no­logy (edTech) in order to trans­fer this approach into a digit­al edu­ca­tion­al sys­tem inten­ded to be used on a large scale in primary schools in France (Adaptiv’Maths pro­ject). My col­league Hélène Sauzéon even showed that this sys­tem can facil­it­ate learn­ing for chil­dren suf­fer­ing from devel­op­ment­al dis­orders such as autism.