Machine Learning


  1. How Do Machi­nes Learn?
  2. What Can Machi­ne Learning Be Used For??
  3. Fur­t­her terms around machi­ne learning




How do Machines Learn?


Arthur Samu­el, one of the pioneers of arti­fi­ci­al intel­li­gence defi­ned ‘machi­ne learning’ in 1959 as, ‘[A] field of stu­dy that gives com­pu­ters the abi­li­ty to learn without being expli­cit­ly pro­gram­med.’

The defi­ni­ti­on has not lost any of its vali­di­ty sin­ce then. Howe­ver, the gene­ral con­di­ti­ons, pos­si­bi­li­ties and are­as of app­li­ca­ti­on for machi­ne learning have chan­ged. Today, machi­ne learning sys­tems can pro­cess enor­mous amounts of data, make very pre­cise pre­dic­tions, and be app­lied much more wide­ly.

The abi­li­ty of a machi­ne to reco­gni­ze pat­terns, inter­pret them cor­rect­ly, and respond to them cor­rect­ly does not come pre-instal­led on a com­pu­ter. Ins­te­ad, it is a pro­cess. This can take place in two ways.

In ‘super­vi­sed learning’, the machi­ne also for­mu­la­tes the cor­rect out­put for each input. A good examp­le of super­vi­sed machi­ne learning is the reco­gni­ti­on of hand­writ­ten let­ters. Images of a hand­writ­ten ABC are fed into the machi­ne learning sys­tem. For each of the­se images, the meta­da­ta con­ta­ins the infor­ma­ti­on for­mu­la­ted by a human ‘tea­cher’ as to which let­ters are invol­ved. The sys­tem learns in the trai­ning pha­se the dif­fe­rent forms in which humans wri­te a let­ter. Once the trai­ning is com­ple­te, the sys­tem should be able to abs­tract from what has been lear­ned and clear­ly reco­gni­ze hand­wri­ting.

The­re is also an approach cal­led ‘unat­ten­ded learning’. While super­vi­sed learning dis­co­vers pat­terns whe­re the sys­tem has a set of ‘cor­rect ans­wers’, in ‘unsu­per­vi­sed learning’, machi­ne learning sys­tems find pat­terns whe­re we do not. This may be becau­se the cor­rect ans­wers are not obser­v­a­ble or pos­si­ble, or becau­se the­re is no right ans­wer for a par­ti­cu­lar pro­blem. The algo­rithms are able to reco­gni­ze struc­tures and simi­la­ri­ties in the data and crea­te a learning model. As a result, such sys­tems are often a black box in which even pro­gramm­ers are not sure about how the sys­tem learns exact­ly.

In prac­tice, unsu­per­vi­sed machi­ne learning is used, for examp­le, whe­re pro­vi­ders of mar­ke­ting data scan huge data sets and clas­si­fy users into spe­ci­fic clus­ters.


Applications and Uses of Machine Learning


The foun­da­ti­ons for machi­ne learning were laid down in the midd­le of the last cen­tu­ry. Howe­ver, it is only in recent years that the poten­ti­al of this approach has come to the fore. Two trends have favou­red the rapid fur­t­her deve­lop­ment of the field and the app­li­ca­ti­on pos­si­bi­li­ties of machi­ne learning in recent years. Com­pu­ters are beco­m­ing increa­singly per­for­mant in order to pro­du­ce lar­ge amounts of data. Machi­ne learning has also expe­ri­en­ced a sur­ge sin­ce com­plex gra­phics pro­ces­sors with several cores have allo­wed par­al­lel cal­cu­la­ti­ons.

On the other hand, the lar­ger the data volu­me, the grea­ter the need to inter­pret, orga­ni­se, and pro­cess the data. Sci­en­ti­fic deve­lop­ments quick­ly found their way into com­mer­ci­al app­li­ca­ti­ons.

  • Spam detec­tion in e-mails
  • Image search, e.g. in search engi­nes like Goog­le
  • Text trans­la­ti­on of app­li­ca­ti­ons such as DeepL
  • Text clas­si­fi­ca­ti­on, for examp­le in online pri­ce com­pa­ri­sons or news por­tals
  • Speech reco­gni­ti­on / speech assi­stant as in Siri, Cor­ta­na, Ale­xa etc.
  • Fraud detec­tion, e.g. for payment pro­vi­ders and in eCom­mer­ce


Further terms around machine learning


The more com­pre­hen­si­ve the phe­no­me­na sur­roun­ding arti­fi­ci­al intel­li­gence and machi­ne learning beco­me in people’s ever­y­day lives, the more fre­quent­ly the rela­ted terms appe­ar in the media.

The con­cepts of arti­fi­ci­al intel­li­gence, machi­ne learning and deep learning are often blur­red and con­fu­sed. The con­nec­tion is basi­cal­ly simp­le, with the con­cepts it’s like a set of Rus­si­an dolls. The lar­gest, outer doll is AI. Machi­ne learning is a branch of AI, while deep learning is con­s­i­de­red a sub­di­sci­pli­ne of machi­ne learning.

The ‘Deep’ in Deep Learning refers to the num­ber of lay­ers through which the data is trans­for­med. Deep learning uses hier­archi­cal lay­ers or a hier­ar­chy of con­cepts in the pro­cess of machi­ne learning. The arti­fi­ci­al neural net­works used are struc­tu­red like the human brain, with the nodes con­nec­ted via a net­work.

Auto­ma­tic machi­ne learning sim­pli­fies the set­up of a sys­tem through exten­si­ve auto­ma­ti­on. Eit­her the ent­i­re pro­cess or selec­ted steps are auto­ma­ted so that no expert is requi­red for each indi­vi­du­al modu­le. For examp­le, data pre­pa­ra­ti­on, fea­ture selec­tion, or model selec­tion can be auto­ma­ted. The Goog­le Cloud AutoML ser­vice, for examp­le, pro­mi­ses that the solu­ti­on will enab­le even users with litt­le pro­gramming know­ledge to crea­te machi­ne learning sys­tems such as trans­la­ti­on models or models for natu­ral lan­guage.

Ano­t­her term often used in con­nec­tion with machi­ne learning is ‘data mining’. Machi­ne learning and data mining have many simi­la­ri­ties, but also signi­fi­cant dif­fe­ren­ces. Machi­ne learning is based on known cha­rac­te­ris­tics and focu­sed on pre­dic­tions. In con­trast, data mining focu­ses on dis­co­vering unknown pro­per­ties in the data.