Machine Learning: can we correct biases?

What are the major chal­lenges cur­rently facing arti­fi­cial intelligence?

In my area of expert­ise, which is machine learn­ing (ML), the three top­ics that I am cur­rently pas­sion­ate about, and that could poten­tially be con­sidered as the great chal­lenges in this field, are bias and fair­ness, weak sig­nals, and learn­ing on net­works. But this is only a par­tial view of the chal­lenges in AI, which is a very broad and mostly inter­dis­cip­lin­ary field. AI is a set of tools, meth­ods and tech­no­lo­gies that enable a sys­tem to per­form tasks in a quasi-autonom­ous way, and there are dif­fer­ent ways of achiev­ing this.

ML is about the machine learn­ing from examples, train­ing itself to per­form tasks effi­ciently that it will later under­take. The great suc­cesses in this area are com­puter vis­ion and auto­mat­ic listen­ing, used for applic­a­tions in bio­met­rics for example, and nat­ur­al lan­guage pro­cessing. One of the ques­tions that cur­rently arises is how much con­fid­ence can be placed in ML tools, as deep learn­ing requires very large volumes of data, which very often come from the web.

Unlike data­sets pre­vi­ously col­lec­ted by research­ers, web data is not acquired in a “con­trolled” way. The vast quant­ity of this data can some­times lead to the meth­od­o­lo­gic­al ques­tions that should be asked to exploit the inform­a­tion it con­tains being ignored. For example, train­ing a facial recog­ni­tion mod­el dir­ectly from web data can lead to bias, in the sense that the mod­el would not recog­nise all types of faces with the same effi­ciency. In this case, the bias may stem from a lack of rep­res­ent­at­ive­ness in the faces used.

If, for example, the data cor­res­ponds mainly to Caucasi­an faces, the sys­tem developed may recog­nise Caucasi­an faces more eas­ily than oth­er types of faces. How­ever, the dis­par­it­ies in per­form­ance may also be due to the intrins­ic dif­fi­culty of the pre­dic­tion prob­lem and/or the lim­it­a­tions of cur­rent ML tech­niques: it is well known, for example, that deep learn­ing does per­form as will in the recog­ni­tion of the faces of new-borns as it does for adult faces. How­ever, there is cur­rently no clear the­or­et­ic­al insight into the link between the struc­ture of the deep neur­al net­work used and the per­form­ance of the mod­el for a giv­en task.

You say “cur­rently”. Does that mean that these biases could one day be removed, or their effect could diminish?

There are dif­fer­ent types of bias. They can be rel­at­ive to the data, there are the so-called “selec­tion” biases, linked to the lack of rep­res­ent­at­ive­ness, “omis­sion” biases, due to errors through endo­gen­eity, etc. Biases are also inher­ent in the choice of the neur­al net­work mod­el, the ML meth­od, a choice that is inev­it­ably restric­ted to the state of the art and lim­ited by cur­rent tech­no­logy. In the future, we may use oth­er, more effi­cient, less com­pu­ta­tion­ally intens­ive rep­res­ent­a­tions of inform­a­tion, which can be more eas­ily deployed, and which may reduce or elim­in­ate these biases, but for the moment they exist!

What role does the qual­ity of the data­sets used for train­ing play in these biases?

It is very import­ant. As I said, giv­en the volume of data required, it is often sourced from the web and there­fore not acquired in a suf­fi­ciently con­trolled way to ensure rep­res­ent­at­ive­ness. But there is also the fact that this data can be ‘con­tam­in­ated’, in a mali­cious way. This is cur­rently an issue for the com­puter vis­ion solu­tions that will be used in autonom­ous vehicles. The vehicle can be deceived by manip­u­lat­ing the input inform­a­tion. It is pos­sible to modi­fy the pixel image of, say, a traffic sign so that the human eye sees no dif­fer­ence, but the neur­al net­work ‘sees’ some­thing oth­er than the traffic sign.

ML is based on a fre­quent­ist prin­ciple and the ques­tion of the rep­res­ent­at­ive­ness of the data in the learn­ing phase is a major issue. Using autonom­ous driv­ing as an example, we now see many vehicles on the Saclay plat­eau, equipped with sensors to store as much exper­i­ence as pos­sible. That being said, it is dif­fi­cult to say how long it will be before we have seen enough situ­ations to be able to deploy a suf­fi­ciently intel­li­gent and reli­able sys­tem in this field, enabling us to deal with all future situations.

There are cer­tainly applic­a­tions for which the data avail­able today allows ML to be imple­men­ted in a sat­is­fact­ory man­ner. This is the case, for example, for hand­writ­ing recog­ni­tion, for which neur­al net­works are per­fectly developed. For oth­er prob­lems, in addi­tion to exper­i­ment­al data, gen­er­at­ive mod­els will also be used, pro­du­cing arti­fi­cial data that account for adverse situ­ations but without claim­ing to be exhaust­ive. This is the case for ML applic­a­tions in cyber­se­cur­ity, in an attempt to auto­mat­ic­ally detect mali­cious intru­sions into a net­work for example.

Gen­er­ally speak­ing, there are many prob­lems for which the data avail­able is too sparse to imple­ment ML in a simple way. This is often the case in anom­aly detec­tion, par­tic­u­larly for pre­dict­ive main­ten­ance of com­plex sys­tems. In some cases, the hybrid­isa­tion of ML and sym­bol­ic tech­niques in AI could provide solu­tions. These aven­ues are being explored in the civil and mil­it­ary avi­ation sec­tors, as well as in med­ic­al ima­ging. In addi­tion to their effect­ive­ness, such approaches may also enable machines to make decisions that are easi­er to explain and interpret.

What is driv­ing evol­u­tion in AI today?

The field of math­em­at­ics con­trib­utes a lot, espe­cially in terms of effi­cient inform­a­tion rep­res­ent­a­tion and algorithms. But it is also tech­no­lo­gic­al pro­gress that is driv­ing AI for­ward. The math­em­at­ic­al concept of neur­al net­works has been around for many dec­ades. Recent tech­nic­al advances, par­tic­u­larly in the field of memory, have made it pos­sible to suc­cess­fully imple­ment deep neur­al net­work mod­els. Sim­il­arly, dis­trib­uted com­put­ing archi­tec­tures and ded­ic­ated pro­gram­ming frame­works have made it pos­sible to scale up learn­ing on large volumes of data. What remains to be done is to design more frugal approaches, so as to reduce the car­bon foot­print of com­pu­ta­tions, which is a very top­ic­al issue!