Artificial intelligence — science fiction or already reality?

Artificial intelligence — science fiction or already reality?

Artificial intelligence — science fiction or already reality?

One of the most excit­ing top­ics with­in the diverse aspects of digi­ti­sa­tion is AI — Arti­fi­cial Intel­li­gence. Many peo­ple still place this in the realm of sci­ence fic­tion, but it is already part of our real­i­ty. Numer­ous researchers world­wide are work­ing on the devel­op­ment of AI sys­tems for almost all areas of our every­day and work­ing life. But exact­ly how far has the devel­op­ment of such arti­fi­cial intel­li­gence actu­al­ly pro­gressed? In order to find answers to these and oth­er ques­tions, we have inter­viewed an expert in this field, Prof. Dr. Tonio Ball from the Med­ical Fac­ul­ty of the Uni­ver­si­ty of Freiburg, who is inves­ti­gat­ing AI sys­tems and their appli­ca­tion in med­i­cine.

Prof. Dr. Ball, could you please start by briefly explain­ing in which area exact­ly you are con­duct­ing research at AI?

Prof. Dr. Ball: The main top­ic of our research group at the Uni­ver­si­ty Hos­pi­tal of Freiburg is the use of AI algo­rithms in the med­ical field. This main­ly involves diagnostics, i.e. the analy­sis of data sets, a task that cur­rent­ly still has to be car­ried out by doc­tors and which is not only time-con­sum­ing but also prone to errors. A pro­gram able to detect devi­a­tions on its own would be a solu­tion that would not only relieve the doc­tors’ work­load, but also reduce the error rate sig­nif­i­cant­ly. A good exam­ple of this are already rel­a­tive­ly high­ly devel­oped algo­rithms that are to be used in radi­ol­o­gy and serve to ana­lyze images after brain scans and deter­mine whether the patient is healthy or sick. But we want to take the whole thing even fur­ther and use AI sys­tems not only for clas­si­fi­ca­tions, but also devel­op algo­rithms that can be used to con­trol robots, e.g. as assis­tance sys­tems for par­a­lyzed patients, or to opti­mize hos­pi­tal orga­ni­za­tion­al process­es. For exam­ple, a pro­gram that mon­i­tors the num­ber of drugs in stock in a hos­pi­tal could, over time, inde­pen­dent­ly take care of reorder­ing them in time by rec­og­niz­ing pat­terns, i.e. how quick­ly which drugs are used up. In this way, there is a very wide range of pos­si­ble appli­ca­tions of such meth­ods in hos­pi­tals, which ulti­mate­ly aim to sim­pli­fy process­es and thus help the staff and free up time for more impor­tant activ­i­ties.

When do we actu­al­ly start talk­ing about “arti­fi­cial intel­li­gence”, i.e. what is the exact dif­fer­ence between a con­ven­tion­al algo­rithm and an algo­rithm that would be clas­si­fied as arti­fi­cial­ly intel­li­gent?

Prof. Dr. Ball: This is a good ques­tion to which, unfor­tu­nate­ly, there is no real­ly sat­is­fy­ing answer. This is sim­ply because there is no gen­er­al def­i­n­i­tion of what exact­ly intel­li­gence is. For exam­ple, where does intel­li­gence begin in the ani­mal king­dom. Is a mouse intel­li­gent only because it can make inde­pen­dent deci­sions? Is there no intel­li­gence in a sin­gle work­er bee because it only per­forms its defined func­tion in the swarm? Can some­thing that is not alive even pos­sess intel­li­gence? These are all dif­fi­cult ques­tions to which there are no gen­er­al­ly accept­ed answers out­side the tech­ni­cal field.
When we talk about arti­fi­cial intel­li­gence in our research area, we gen­er­al­ly refer to pro­grams that are capa­ble of learn­ing. If you now have an algo­rithm that can only han­dle one task, no mat­ter how com­plex it is, once it has been pro­grammed, it is done. That would be a reg­u­lar pro­gram, where every deci­sion that can be made was pre­de­ter­mined by the pro­gram­mer. Most of the algo­rithms that would now be described as arti­fi­cial­ly intel­li­gent are relat­ed to so-called “machine learn­ing”. Anoth­er fea­ture that, togeth­er with machine learn­ing, often leads to algo­rithms being placed in the cat­e­go­ry of arti­fi­cial intel­li­gence is the abil­i­ty to plan abstract­ly, or to rep­re­sent com­plex rela­tion­ships.

When you speak now of ” adap­tive algo­rithms”, does that mean that these algo­rithms are sim­ply fed with infor­ma­tion and thus grad­u­al­ly become more and more “intel­li­gent”, or how do you have to imag­ine that?

Prof. Dr. Ball: Whether more data auto­mat­i­cal­ly leads to a bet­ter func­tion­ing of the whole thing is anoth­er ques­tion, but often it is. The prin­ci­ple behind this “machine learn­ing” is often arti­fi­cial neur­al net­works con­sist­ing of inter­con­nect­ed arti­fi­cial neu­rons, i.e. nerve cells. The con­nec­tions between the sin­gle neu­rons have a cer­tain weight, they tell us how much influ­ence one neu­ron has on the oth­ers. Depend­ing on how these weights are now set, the abil­i­ties and func­tion of the entire net­work also change. So when we speak of a “learn­ing process”, this is ulti­mate­ly noth­ing more than the sys­tem­at­ic opti­mi­sa­tion of these weights. A typ­i­cal exam­ple would be the clas­si­fi­ca­tion of images into the cat­e­gories “dog”, “cat”, “mouse”, “flower”. The net­work starts to clas­si­fy images that it receives into one of these cat­e­gories, while spe­cial­ly devel­oped algo­rithms, depend­ing on the result, opti­mize the weights of the arti­fi­cial neu­rons and thus slow­ly ensure that the net­work deliv­ers increas­ing­ly bet­ter results over time, i.e. “learns”. How­ev­er, this approach does not have a 100% suc­cess rate. It is not unusu­al for such a learn­ing process to fail and for the net­work not to learn how to mas­ter the giv­en task, since this area of research can give rise to many prob­lems that can­not be iden­ti­fied or solved at the moment.

Do you think that arti­fi­cial intel­li­gence could even­tu­al­ly be devel­oped far enough that it could com­pete with the intel­lect and exper­tise of a human being, or are there lim­i­ta­tions to such learn­ing meth­ods?

Prof. Dr. Ball: In some areas such arti­fi­cial net­works are already far ahead of humans. A well-known exam­ple is the vic­to­ry of an AI against the world cham­pi­on of the Asian board game “Go”. While the bat­tle between man and machine in chess had been lost since the end of the 1990s, since even the chess pro­grams of that time were able to cal­cu­late all pos­si­ble moves very far and make the move with the high­est cal­cu­lat­ed chance of vic­to­ry, Go was for a long time con­sid­ered a bas­tion of human supe­ri­or­i­ty. There are so many dif­fer­ent moves and posi­tions in Go that not even the most pow­er­ful com­put­er in the world could cal­cu­late all the pos­si­bil­i­ties fast enough . In chess it is also much eas­i­er to esti­mate for which play­er a cer­tain board posi­tion or move is favourable than in the case of much deep­er Go. Since con­ven­tion­al­ly pro­grammed com­put­ers are over­whelmed when it comes to strate­gic, intel­li­gent think­ing, it was long assumed that no pro­gram would be able to beat a Go pro­fes­sion­al in the fore­see­able future. That’s why it was a big sur­prise when in 2016 an AI pro­gram that was also trained with the “deep learn­ing” method, a form of machine learn­ing, defeat­ed the world’s best play­er . And now, only three years lat­er, there are already fur­ther devel­oped and even faster learn­ing AI sys­tems that can eas­i­ly beat the pro­gram of that time.
How­ev­er, such devel­op­ments, as with the Go pro­gram, are very dif­fi­cult to pre­dict. Only a few years before 2016, hard­ly any­one would have thought it pos­si­ble that Go could be defeat­ed by a com­put­er in such a short time. But even if future pre­dic­tions in this area are very dif­fi­cult, I do not see any lim­i­ta­tions in gen­er­al for the devel­op­ment of AI sys­tems that are supe­ri­or to humans in many areas. Sim­i­lar to the Go exam­ple, it is quite plau­si­ble that with the help of e.g. deep learn­ing, the pro­grams will grad­u­al­ly become bet­ter and will be able to per­form more and more tasks and activ­i­ties bet­ter and more effi­cient­ly than humans. But whether the AI sys­tems will be lim­it­ed in some areas is an excit­ing ques­tion that can­not be answered at the moment. How­ev­er, I believe that devel­op­ment in this area is pro­gress­ing so quick­ly that we may not have to wait as long for the answer as many peo­ple still think.

So the pro­gramme that beat the Go world cham­pi­on in 2016 did not cal­cu­late its moves in advance, but react­ed to new sit­u­a­tions and inde­pen­dent­ly searched for an answer or solu­tion?

Prof. Dr. Ball: Exact­ly, this was not, as in some chess pro­grams, sim­ply “Which move is the best”, but the sys­tem learned based on a lot of test games, which are strate­gi­cal­ly good moves that make sense in the long-term course of the game and then decid­ed over and over again for the move it thought was best.

Of course, you are right in say­ing that it is dif­fi­cult to pre­dict future direc­tion of such a com­plex issue: How long do you think it will be before AI pro­grams have an impact on the work­place and estab­lish their pres­ence on the mar­ket?

Prof. Dr. Ball: Well, part­ly that has already begun. Pro­grams already exist for cus­tomer ser­vice cen­ters that can inde­pen­dent­ly trans­fer a caller to the right pro­cess­ing point. Inten­sive work is also being done in the legal field on auto­mat­ed legal advice sys­tems that could soon take over part of the work of para­le­gals. The devel­op­ment of all these pro­grams is advanc­ing very quick­ly and the first of them are already on the mar­ket or will soon be. Even though we are not yet feel­ing the changes so strong­ly at the moment, it is con­sid­ered quite real­is­tic that the increas­ing use of AI in the work­place will in the long term have con­se­quences sim­i­lar to those of the first wave of indus­tri­al­i­sa­tion at the end of the 19th cen­tu­ry. How­ev­er, this should not nec­es­sar­i­ly be seen as some­thing neg­a­tive; espe­cial­ly in our field of exper­tise, the inter­ac­tion of AI sys­tems and employ­ees will ulti­mate­ly ben­e­fit many patients. Today, when a doc­tor is talk­ing to a patient, it is often the case that the doc­tor is already sit­ting at the com­put­er and typ­ing in symp­toms and oth­er infor­ma­tion dur­ing the con­ver­sa­tion. Espe­cial­ly in sit­u­a­tions like these, an intel­li­gent pro­gram that lis­tens to the con­ver­sa­tion and is able to auto­mat­i­cal­ly fil­ter and inde­pen­dent­ly trans­fer rel­e­vant infor­ma­tion would be extreme­ly prac­ti­cal. By sup­port­ing doc­tors in such a sit­u­a­tion by AIs, they would be able to con­cen­trate bet­ter on their patients and, above all, save a lot of time, which is par­tic­u­lar­ly impor­tant in hos­pi­tals. What I mean to say is that the changes that arti­fi­cial intel­li­gence brings with it can be pos­i­tive, it depends to a great extent on what we make of it and how we shape the changes that are com­ing.

So do you think that arti­fi­cial intel­li­gences, as in your exam­ple with the doc­tor, will act more as a kind of sup­port­ing help sys­tem for human employ­ees, or do you also see the risk that many jobs may be lost due to the increased use of AI sys­tems?

Prof. Dr. Ball: I think that both of these cas­es will occur. The poten­tial appli­ca­tions of AI sys­tems are so broad and so effi­cient in their imple­men­ta­tion that some, if not many, jobs will def­i­nite­ly be threat­ened by AI. On the oth­er hand, in many areas, a coop­er­a­tion sce­nario will emerge in the first place. It remains to be seen to what extent the divi­sion of labour will shift to the AIs in the future. Of course, human resources are indis­pens­able in areas such as med­i­cine. In oth­er sec­tors, how­ev­er, it could also hap­pen very quick­ly. For exam­ple, as soon as autonomous dri­ving sys­tems can dri­ve safe­ly enough and are finan­cial­ly viable, we would sud­den­ly need far few­er or no more taxi and bus dri­vers.

My spe­cial thanks go to Prof. Dr. Ball for his will­ing­ness to be inter­viewed and for his high­ly inter­est­ing remarks.

Author: Linus Plesni­la, Glasford Inter­na­tion­al Deutsch­land Research & Ana­lyt­ics