Big Data and Algorithms

Do you know how old algo­rithms real­ly are? Or where they came from? What is an algo­rithm? Are there dif­fer­ent types of algo­rithms…? Muham­mad al-Khwariz­i­mi was a 9^th cen­tu­ry Per­sian astronomer and math­e­mati­cian. His text­book On Indi­an Numer­als was trans­lat­ed into Latin four cen­turies lat­er. It is con­sid­ered the most impor­tant source for the Indo-Ara­bic numer­al sys­tem and writ­ten arith­metic. His sur­name even sounds pho­net­i­cal­ly like an algo­rithm “Algo­ris­mi”. Ada Lovelace was the first per­son to write down an algo­rithm intend­ed for a com­put­er in 1843. She is con­sid­ered to be the first female programmer.

In the lat­est of her Duet inter­views, Dr Cal­daro­la, author of Big Data and Law, and IT-Expert Prof. Dr Katha­ri­na Zweig talk about algo­rithms and heuris­tics and their dai­ly use.

Let’s start our duet by explain­ing what an algo­rithm is.

Prof. Dr Katha­ri­na Zweig: We have heard a lot about algo­rithms in the last decades. How­ev­er, it is a spe­cif­ic term, with a slight­ly dif­fer­ent mean­ing from how it is often used: an algo­rithm is a sequence of instruc­tions that can lead to a proven solu­tion of a math­e­mat­i­cal prob­lem. For exam­ple, there are mul­ti­ple ideas about how to sort things on a com­put­er: a bank might want to sort all of their bor­row­ers by the lev­el of their cred­it. Giv­en any sequence of instruc­tions to sup­pos­ed­ly solve this prob­lem, this needs to be proven math­e­mat­i­cal­ly to be called an algo­rithm.  In the media, how­ev­er, any type of soft­ware is nowa­days called an algo­rithm. This is a bit unfor­tu­nate, as AI does in almost all cas­es not use real algo­rithms, but only so-called heuris­tics. A heuris­tic is a sequence of instruc­tions that is like­ly to yield a solu­tion, maybe even a good one. But in con­trast to an algo­rithm, we do not have a math­e­mat­i­cal guar­an­tee that the solu­tion will be opti­mal for a giv­en problem.

Is an algo­rithm more objec­tive than the human brain and, if so, why?

Again, we need to dif­fer­en­ti­ate between real algo­rithms and those that are actu­al­ly heuris­tics. Using the def­i­n­i­tion above, an algo­rithm is objec­tive: it is proven to find the best solu­tion to a giv­en prob­lem. For exam­ple, when you use a nav­i­ga­tion sys­tem, the algo­rithm behind it will find the absolute­ly short­est path between your loca­tion and some oth­er loca­tion. How­ev­er, if you want to find the fastest solu­tion giv­en real traf­fic, no algo­rithm can guar­an­tee the fastest route. Instead, the sys­tem will use AI to pre­dict which street will be con­gest­ed at which time point. There are heuris­tics for these pre­dic­tions and we can mea­sure in hind­sight, how well they per­formed. But there is no objec­tive guar­an­tee that the heuris­tic will find the fastest way. Thus, algo­rithms are always to be pre­ferred, but in most cas­es, no algo­rithm exists and we need to resort to heuris­tics which are not objective.

How much data do you need to devel­op an algo­rithm? Does the qual­i­ty of the algo­rithm depend on the quan­ti­ty and qual­i­ty of the data used? What quan­ti­ty and qual­i­ty of data is required?

The inter­est­ing thing is that the devel­op­ment of an algo­rithm in the strict def­i­n­i­tion men­tioned above does not need any data. It needs math­e­mat­i­cal cre­ativ­i­ty, pens, a black board, time, frus­tra­tion tol­er­ance, and a bit of luck. Find­ing the best sequence of instruc­tions to solve a giv­en prob­lem is an abstract, the­o­ret­i­cal process. But now comes the twist: many inter­est­ing prob­lems can­not be solved effi­cient­ly by an algo­rithm. There are two rea­sons:  either because we haven’t found one yet, or what we have found is too slow.  Some­times we have an algo­rithm, but it would lit­er­al­ly take longer than 1000 years to wait for the result.

Thus, nowa­days, com­put­er sci­ence answers some ques­tions by design­ing heuris­tics that try to find sta­tis­ti­cal­ly robust pat­terns in large amounts of data – inde­pen­dent of the con­crete prob­lem to solve. This approach is called machine learn­ing. Machine learn­ing con­sists of many dif­fer­ent heuris­tics, based on some human intu­ition of how learn­ing by see­ing mul­ti­ple exam­ples works. For exam­ple, the machine could be shown infor­ma­tion on a bank’s bor­row­ers and on whether they paid  back their loan or not. The machine could then be tasked to find the best pat­terns to iden­ti­fy those that are not cred­it­wor­thy. These ideas about learn­ing can be sim­ple math­e­mat­i­cal equa­tions like a line through some data, or the famous neur­al net­works. Neur­al net­works are just huge sets of math­e­mat­i­cal equa­tions with an even larg­er set of para­me­ters or weights, that mod­el which of the pre­sent­ed infor­ma­tion is most impor­tant. All these ideas of learn­ing have in com­mon that they read data from the past to find pat­terns to make deci­sions for the future. But they iden­ti­fy and store these pat­terns in dif­fer­ent struc­tures. In the above exam­ples infor­ma­tion is stored in the weights in the respec­tive math­e­mat­i­cal equations.

The prob­lem to solve is then to find the best pat­terns in the data to make future deci­sions. For exam­ple, find the best line through the data or the best weights in the neur­al net­work such that they can iden­ti­fy cred­it­wor­thy future appli­cants. How­ev­er, only the sim­plest of these prob­lems can be solved by an algo­rithm, name­ly find­ing the best pos­si­ble line. How­ev­er, the line is very often a very poor rep­re­sen­ta­tion of the data. Thus, it can­not be used in real­i­ty. All oth­er ideas about learn­ing cre­ate math­e­mat­i­cal prob­lems for which we do not have any algo­rithms, only heuris­tics. In short, we only have algo­rithms with proven opti­mal solu­tions for the sim­plest learn­ing tasks. For more com­plex learn­ing tasks, we only have heuris­tics that might not result in the best mod­el to rep­re­sent the data.

It is believed that enough data will make it very like­ly that the heuris­tics can come close to a very good solu­tion. How­ev­er, in most cas­es we do not know exact­ly how much data we need.

Are AI sys­tems devel­oped with raw data or do they require processed data? If the lat­ter is the case, what ‘mold’ is required and how is it ‘made’?

My favourite quote by Geof­frey Bowk­er answers that question:

My favourite quote:

“Raw data is both an oxy­moron and a bad idea”.

Geof­frey Bowker

The moment some­one decides to store some infor­ma­tion is the moment the data is “cooked” accord­ing to some flavour of choice. What we select and what remains undoc­u­ment­ed is a choice. There is no such thing as raw data, all data is cooked.

AI sys­tems work with prob­a­bil­i­ties and cor­re­la­tions, while our life, sci­ence and law fol­low causal­i­ty. How is this compatible?

In my view, machine learn­ing is inter­est­ing as a tool to find hypothe­ses about how the world works. How­ev­er, the pat­terns found may be inci­den­tal. Hun­dreds of years of epis­te­mol­o­gy, the sci­ence of sci­ence or how we can deduce facts from obser­va­tions, have shown us that iden­ti­fy­ing cor­re­la­tions is not enough. Thus, if an AI sys­tem actu­al­ly does show supe­ri­or deci­sion-mak­ing it is nec­es­sary to iden­ti­fy why that is the case. For this, we have meth­ods from the nat­ur­al sci­ences: We need to iso­late pos­si­ble causal fac­tors and find out whether they are indeed caus­es or not. I see no way to short­cut this process. The ques­tion of what is a fact can­not be left to the meth­ods of machine learn­ing alone.

AI sys­tems are tout­ed for their effi­cien­cy, mea­sur­a­bil­i­ty, stan­dards, automa­tism and much more. Is that at the expense of human­i­ty because, for instance, peo­ple are seen as “poor” deci­sion-mak­ers or because peo­ple are pre­dictable based on past behaviour?…

The first ques­tion is: Under which cir­cum­stances are humans poor deci­sion mak­ers? Of course, psy­chol­o­gists and behav­iour­al econ­o­mists have shown ample exam­ples in recent years where humans have made biased deci­sions. Some­times these can also be called irra­tional deci­sions. The book Noise by Kah­ne­mann, Sibony and Sun­stein sum­maris­es many of these stud­ies in which experts are asked to come up with deci­sions, but it turns out that they do not agree in their find­ings very well. This can be cost­ly, e.g., if the risk of some insur­ance case is wrong­ly cal­cu­lat­ed. The authors also pro­vide good rules to reduce the noise. So, can AI sys­tems help to make bet­ter deci­sions than humans? I do not know many stud­ies in which this has actu­al­ly been proven. I am open to the idea, but I would like to see stud­ies. Fur­ther­more, the ques­tion is what would we con­clude from such a find­ing? The thing is that machine learn­ing is based on iden­ti­fy­ing sta­tis­ti­cal cor­re­la­tions, not nec­es­sar­i­ly causal rela­tions. Thus, we do not know whether the find­ing can be gen­er­alised and applied to future data or not. Can a machine replace a human deci­sion under this con­di­tion? The answer to this ques­tion is what I con­sid­er in my new book. My gen­er­al answer is a machine can only safe­ly replace humans in deci­sion mak­ing if the machine comes to an answer that can be ver­i­fied inde­pen­dent­ly. For all kinds of judg­ments, like the suit­abil­i­ty of a job can­di­date, it cannot.

We talk about errors in an AI sys­tem as a mat­ter of course. But when is the result of such a sys­tem actu­al­ly “good”? The word “good” cer­tain­ly dif­fers accord­ing to var­i­ous cul­tures, ethics, morals, etc. How does that go hand in hand with the world­wide use of an algorithm?

A deci­sion is good if it is cor­rect. If we can­not define what a cor­rect deci­sion is then we should not use AI sys­tems. The deci­sion of a machine is cor­rect if an inde­pen­dent ver­i­fi­ca­tion process comes to the same con­clu­sion. If there is no such ver­i­fi­ca­tion process, we should not use AI sys­tems which some­body just claims that they are good to go. Unless in sit­u­a­tions that are not very harm­ful, such as prod­uct rec­om­men­da­tion systems.

How does a con­sumer of a prod­uct or ser­vice that con­tains an AI sys­tem get the nec­es­sary evi­dence or the train­ing data behind it to be able to show that the algo­rithm con­tains an error? Does the aver­age con­sumer have the nec­es­sary com­pe­tence for this and isn’t the AI ​​sys­tem rather “opaque”?

If there is a ver­i­fi­ca­tion process, even if it is expen­sive to use it, the per­son can at least iden­ti­fy incor­rect deci­sions. For exam­ple, if a machine pre­dict­ed a Coro­na infec­tion based on the cough­ing sound of a patient, a PCR could be used to ver­i­fy that result. For cred­it­wor­thi­ness, it is already much more com­pli­cat­ed: Of course, it can be checked in 10 years whether per­sons deemed cred­it­wor­thy by the sys­tem, actu­al­ly did pay their loans back or not. But we can­not know whether the per­sons deemed unwor­thy of cred­it would have paid back their loans or not. Thus, for risk pre­dic­tions, we already have a prob­lem of know­ing whether the sys­tem is accu­rate or not. 

I assume from your answer that it takes a lot of knowl­edge and expe­ri­ence to detect an error in an AI sys­tem. And I also assume that it takes even more exper­tise to find the cause for it after the fact. If an employ­ee in a com­pa­ny has detect­ed an error in an AI sys­tem, it will cer­tain­ly be dif­fi­cult to find a col­league who is com­pe­tent in terms of IT (for the AI ​​sys­tem) and the spe­cial­ist area (e.g., the area of ​​appli­ca­tion tax­es) who can ver­i­fy the opin­ion. It is even more dif­fi­cult to per­suade col­leagues to stop using the sys­tem or even “repair” it, because for the com­pa­ny this means effort, time, loss of sales, pos­si­bly loss of rep­u­ta­tion and so on. Is that an unsolv­able dilemma?

An AI sys­tem is a process used by some­one. If that process is faulty and the con­se­quences of this faulty process are high, soci­ety needs to find ways to stop them. I do not think that there is some­thing inher­ent­ly dif­fer­ent with respect to whether AI is used or not. How­ev­er, it is dif­fi­cult to find out whether they are faulty or not, as already dis­cussed. Thus, those sys­tems for which we have a hard time prov­ing their fault­i­ness, should not be used for deci­sions with a high social impact.

Prof. Dr Zweig, thank you for shar­ing your insights on algo­rithms, heuris­tics and their impact on dai­ly life.

Thank you for hav­ing me.