Teburin abubuwan ciki na littafin machine learning: abin da yake faɗi

Foundations of Machine Learning (bugu na biyu, MIT Press, 2018) na Mehryar Mohri, Afshin Rostamizadeh da Ameet Talwalkar yana ma'anar koyo a matsayin amfani da gogewa don yin hasashe daidai, kuma yana gina tsayayyiyar manhaja a kan sample complexity, iyakokin Rademacher da tabbacin ka'ida.

Yawancin littattafai game da machine learning suna farawa da alƙawari game da abin da za ka iya ginawa. Wannan yana farawa da ma'ana da jerin abubuwan da ke yanke ko mai koyo zai yi nasara. Wannan tsari yana nuna abin da marubutan ke ganin batun yake, kuma ina ganin shawara ce daidai. Fannin da ke tallata kansa da hasashe yana da bashin bayyana wa masu karatunsa lokacin da za a iya amincewa da hasashe, kuma wannan bayani ne littafin ya sa a gaba ya bayar.

Wannan labarin yana karanta littafin ta abin da shafukan farkonsa, gabatarwarsa (preface) da shafukan buɗewarsa suka faɗa a zahiri. Wannan aiki ne mafi ƙunci fiye da duba kowane babi, kuma yana gaskiya game da iyakokinsa: majiyoyin nan su ne shafin take, teburin abubuwan ciki, gabatarwa da farkon sashen shimfiɗa (introduction). A cikin waɗannan iyakoki akwai hujja bayyananniya da za a bi. Ta shafi menene machine learning, me ya dogara da shi, me marubutan suka zaɓa su rufe da kuma me suka zaɓa su bari.

Littafi mai asali tabbatacce #

Fara da gaskiyar da ke ɗaure abin. Taken littafin shi ne Foundations of Machine Learning, kuma wannan shi ne bugu na biyu. Yana cikin jerin Adaptive Computation and Machine Learning, wanda editansa shi ne Francis Bach. Marubutan su ne Mehryar Mohri, Afshin Rostamizadeh da Ameet Talwalkar. Mawallafin shi ne The MIT Press, Cambridge, Massachusetts, kuma haƙƙin mallaka na Massachusetts Institute of Technology ne, mai kwanan wata 2018. Bugun hardcover yana ɗauke da ISBN 9780262039406.

Bayanan samarwa guda biyu sun cancanci jumla ɗaya kowanne. Marubutan da kansu ne suka shirya littafin a LaTeX, abin da ya dace da rubutu cike da iyakoki (bounds), ma'anoni da hujjoji, kuma an buga shi aka ɗaure shi a Amurka. Babu ɗayan waɗannan da ke canza yadda za ka karanta lissafin. Amma marubutan da suka shirya littafinsu da kansu sun taɓa kowane shafi, kuma a littafi irin wannan hakan yana kwantar da hankali.

Asalin abubuwan ciki ya fi muhimmanci. Yawancinsu sun samo asali daga wani kwas na digiri na biyu, wanda ake kira Foundations of Machine Learning shi ma, wanda marubuci na farko ya koyar a Courant Institute of Mathematical Sciences na New York University a cikin shekaru goma sha huɗu da suka gabata. Wannan asali ne tabbatacce da za a iya dubawa. Littafin karatu da ya girma daga shekaru goma sha huɗu na koyar da kwas ɗaya mai yiwuwa abubuwan da ɗalibai suka ga suna da wuya sun siffanta shi, kuma motsa jiki (exercises) da ƙarin bayanan da aka bayyana a ƙasa suna karantuwa haka. Zan yi hankali kada in tura hasashen fiye da yadda majiyar ta yarda. Gabatarwar ta faɗi inda abubuwan suka fito, amma ba ta faɗi sau nawa aka gyara kwas ɗin ba ko abin da ɗalibai suka sha wahala da shi. Abin da ta faɗa ya isa: wannan kayan karatun kwas ne, ba bincike da aka haɗa daga wallafe-wallafe ba.

Ma'anar da ke ɗaukar nauyi mai yawa #

Duk abin da ke biye ya dogara da jumla ɗaya. Machine learning, a cewar marubutan, za a iya ma'anarsa a faɗaɗe a matsayin hanyoyin lissafin kwamfuta da ke amfani da gogewa don inganta aiki ko yin hasashe daidai.

Kowane suna a wannan jumla yana da aikinsa. “Hanyoyin lissafin kwamfuta” yana sanya batun a cikin kimiyyar kwamfuta. Sannan an ƙayyade “gogewa”: tana nufin bayanan da suka gabata da ke hannun mai koyo. Jumlar ba ta da'awar cewa mai koyo yana fahimtar wani abu, kuma ba ta faɗi komai game da basira ba. Mai koyo yana da bayanan da suka gabata da kuma manufa, wato ingantaccen aiki ko hasashe daidai, kuma dukan fannin yana game da dangantakar waɗannan biyu.

Zan iya cewa wannan ma'ana ta fi amfani ga masu karatu na fasaha da masu kafa kamfanoni fiye da manyan ma'anonin da ke yawo, domin tana nuna kai tsaye ga abubuwan da kake iko da su. Idan koyo aiki ne na bayanan da suka gabata, to tasirinka a kan sakamako yana cikin irin bayanan da kake tattarawa da yawansu. Da'awar da ke biye a rubutun ta faɗi haka: inganci da yawan bayanai suna da matuƙar muhimmanci ga nasarar hasashen da mai koyo ke yi. An nuna wannan magana a rubutun a matsayin tsari mai sauƙi, ba tare da wani sharaɗi ba.

Marubutan sun kuma bayyana dabarun da kansu. Hanyoyi ne da bayanai ke jagoranta (data-driven) waɗanda ke haɗa muhimman ra'ayoyin kimiyyar kwamfuta da ra'ayoyi daga ƙididdiga (statistics), yiwuwa (probability) da optimization. Wannan jumla tana aiki a matsayin jerin abubuwan da ake buƙata kafin karatu da kuma taswirar maƙwabtan fannin. Ƙungiyar da ke son gina tsarin koyo tana amfani da kimiyyar kwamfuta, ƙididdiga, yiwuwa da optimization tare. Ɗaukar ɗayansu a matsayin zaɓi zaɓi ne da littafin bai goyi baya ba.

Misali ɗaya da aka yi aiki da shi: takardu da batutuwa #

Sashen shimfiɗa yana ɗaure ma'anar a cikin wata matsalar koyo ta zahiri. Ɗauki samfuri mai iyaka na takardu da aka zaɓa ba bisa tsari ba (randomly), kowacce da lakabin batu. Aikin shi ne amfani da wannan samfuri don hasashen batun takardun da ba a gani ba.

Wannan tsari yana biyan karatu a hankali. Samfurin yana da iyaka, don haka mai koyo ba ya taɓa ganin komai. An zaɓi takardun ba bisa tsari ba, abin da ke ba samfuri mai iyaka damar faɗin wani abu game da waɗanda ba a gani ba. Kowacce tana da lakabin batu, don haka mai koyo yana da jagora. Kuma manufar ita ce takardun da bai gani ba, wanda shi ne bambanci tsakanin hasashe da haddacewa.

Sannan marubutan suka fito da abubuwa biyu da ke sauƙaƙa ko wahalar da aikin. Na farko shi ne girma: a bayyane yake, girman samfurin, sauƙin aikin. Na biyu shi ne ingancin lakabi. Wahalar aikin ta kuma dogara da ingancin lakabin da aka ba takardun da ke cikin samfurin, tun da ba lallai ne dukan lakabin su zama daidai ba, da kuma yawan batutuwan da ka iya kasancewa.

Wannan jumla tana ɗauke da maɓallai uku daban-daban. Girman samfuri ɗaya ne. Daidaiton lakabi wani ne. Yawan batutuwan da ka iya kasancewa shi ne na uku. Babu ɗayansu da ya shafi algorithm. Dukansu halaye ne na matsalar da bayanan, kuma su ne ke yanke yadda matsalar take da wuya kafin a zaɓi wata hanya. Ga mai kafa kamfani da ke yanke shawarar inda zai kashe kasafin kuɗi, wannan lura ce mai kaifi. Idan za ka iya inganta lakabin ko ƙara girman samfurin, kana aiki a kan abubuwan da marubutan suka ambata a fili a matsayin masu ƙayyade wahala.

Misalin ya kuma ambaci aikinsa. Hasashen lakabin takarda ana kuma kiransa document classification, kuma marubutan sun yi saurin ƙara cewa ba shi kaɗai ne aikin koyo ba. Wannan taka-tsantsan yana da sauƙin wucewa. Misalin takarda abin hawa ne kawai, kuma da'awar da ke ƙarƙashinsa ta gama-gari ce.

Sarƙaƙiya uku, ba biyu ba #

A nan littafin ya yi matakinsa mafi bambanci, kuma yana da sauƙin rasawa domin yana zuwa a matsayin ɗan gajeren tsokaci. Machine learning, a cewar marubutan, ya ƙunshi tsara algorithms na hasashe masu inganci da daidaito. Kamar a sauran fannonin kimiyyar kwamfuta, time da space complexity muhimman ma'aunai ne na ingancin irin waɗannan algorithms. Duk wanda ya karanci algorithms zai gane waɗannan biyu. Tsawon lokaci nawa yake ɗauka, kuma ƙwaƙwalwa nawa yake buƙata?

Sannan ƙarin ya zo. Machine learning yana buƙatar, ƙari, ra'ayin sample complexity don auna girman samfurin da algorithm ke buƙata don koyon wani iyali na ra'ayoyi. Wannan shi ne ma'aunin da ba shi da kwatankwaci a cikin aikin tsarawa (sorting). Ba a buƙatar faɗa wa algorithm na sorting misalai nawa yake buƙata ba. Mai koyo kuwa yana buƙata, domin abin da yake fitarwa hasashe ne game da lamuran da bai gani ba, kuma yawan gogewar da ke bayan wannan hasashe wani ɓangare ne na abin da kake biya.

Zan faɗe shi haka: time da space suna faɗa maka abin da algorithm ke kashewa wajen gudanarwa, kuma sample complexity yana faɗa maka abin da yake kashewa wajen koyarwa. Mai koyo mai sauri amma mai buƙatar samfuri mai girman da ba zai yiwu ba ba shi da inganci ta kowace fuska da ke da muhimmanci ga mutumin da ke tattara bayanan. Sanya ma'auni na uku kusa da sauran biyun hanya ce ta littafin na cewa bayanai albarkatu ne masu farashi, da za a tsara kasafin kuɗinsu kamar compute da ƙwaƙwalwa.

Rubutun ya ci gaba a hanya ɗaya. A gama-gari, tabbacin koyo na ka'ida ga algorithm ya dogara da sarƙaƙiyar concept classes da ake la'akari da su da girman samfurin horo. Wannan taƙaitacciyar magana ce game da abin da ake gina tabbaci da shi. Akwai shigarwa biyu, yadda iyalin ra'ayoyin yake da wadata da yawan bayanan horo da kake da su, kuma tabbaci magana ce da ke danganta su da aiki. Lura da abin da wannan ke nufi game da tsarin littafin. Babobin da ke biye dole ne su samar da hanyoyin auna sarƙaƙiyar concept class, kuma teburin abubuwan ciki ya nuna cewa sun yi haka.

Marubutan sun kuma bayyana dalilin da ya sa fannin yake inda yake. Tun da nasarar algorithm na koyo ta dogara da bayanan da aka yi amfani da su, machine learning yana da alaƙa ta asali da nazarin bayanai da ƙididdiga. An bayyana tsarin a sarari: nasara ta dogara da bayanai, don haka batun yana cuɗanya da fannonin da ke nazarin bayanai. Ina ganin wannan ita ce jumla mafi gaskiya a sashen shimfiɗa. Ba ta gabatar da ƙididdiga a matsayin abokin hamayya ko kakan fannin ba. Tana gabatar da ƙididdiga a matsayin sakamakon abin da koyo ya dogara da shi.

Fanni	Time Complexity	Space Complexity	Sample Complexity
Ma'anar aiki	Lokacin gudanarwa na asymptotic da ake buƙata don fitar da hypothesis $h$	Ƙwaƙwalwar da ake buƙata lokacin horo da kimantawa	Yawan misalan horo $m$ da ake buƙata don koyon iyalin ra'ayoyi
Ka'idar da ke jagoranta	Computational complexity (P vs NP, wahalar horo)	Iyakokin ƙwaƙwalwar tsari (architecture) da cache locality	PAC learning, Rademacher complexity da VC-dimension
Tasirin kuɗi	Kasafin compute na sabar da GPU/CPU inference	Ƙarfin RAM da SRAM na cikin chip	Kuɗin tattara bayanai, bayani (annotation) da lakabi
Matsayin ka'ida	Yana auna ingancin lokacin gudanar algorithm	Yana auna ingancin amfani da sarari	Yana ƙayyade generalization error da iyakoki a kan bayanan da ba a gani ba

Ƙashin bayan ka'ida #

Gabatarwar ta faɗi yadda aka gina littafin. Babobi uku ko huɗu na farko suna shimfiɗa tushen ka'ida ga abubuwan da ke biye. Teburin abubuwan ciki yana ba mu damar ganin abin da aka gina wannan tushe da shi, aƙalla ta taken babobi.

Babi na farko a jerin abubuwan ciki yana da take “The PAC Learning Framework”, kuma a cikinsa akwai sassan da ke da take “Guarantees for finite hypothesis sets - consistent case” da “Guarantees for finite hypothesis sets - inconsistent case”. Sunayen sassan suna faɗa mana abubuwa biyu. Ɗaya shi ne cewa tsarin (framework) yana samar da tabbaci, wanda ya dace da girmamawar sashen shimfiɗa. Ɗayan kuma shi ne an raba tabbacin bisa ko lamarin consistent ne ko inconsistent, kuma littafin yana kula da su daban. Ba zan yi hasashe game da abin da lamuran consistent da inconsistent suka ƙunsa ba, domin majiyoyin nan sun ba da taken sassan kawai. Amma haɗa su biyu yana nuna marubutan ba sa gabatar da theorem guda mai tsabta su wuce. Suna ba lamarin da ya fi wuya da rikitarwa sashensa na kansa.

Babi na gaba da aka jera shi ne “Rademacher Complexity and VC-Dimension”. Sassansa sun rufe Rademacher complexity, growth function, VC-dimension da lower bounds. Idan aka karanta shi tare da sashen shimfiɗa, wannan babi shi ne amsar halitta ga tambayar da sashen shimfiɗa ya tayar. Idan tabbaci ya dogara da sarƙaƙiyar concept classes da ake la'akari da su, kana buƙatar kayan aiki don auna wannan sarƙaƙiya. Rademacher complexity, growth function da VC-dimension an jera su a matsayin sassan da ke ɗaukar wannan nauyi. Haɗa lower bounds a jerin sassan shi ma ya cancanci lura. Littafin da ke gabatar da upper bounds kawai yana faɗa maka abin da za a iya cimma. Wanda ya haɗa da lower bounds yana kuma faɗa maka abin da ba zai yiwu ba.

Sannan babi na 4 ya zo, “Model Selection”, tare da sassan da suka haɗa da empirical risk minimization (ERM) da structural risk minimization (SRM). Za mu dawo kan wannan babi idan muka duba bugu na biyu, domin yana ɗaya daga cikin wuraren da sabon bugun ya fi canzawa.

Tsarin an yi shi da gangan. Tushe da farko, sannan injin auna sarƙaƙiya, sannan model selection, sai bayan haka algorithms. Hanya ce ta koyarwa da za a iya kare ta, kuma ta dace da bayanin littafin na yadda yake aiki: babobin farko suna shimfiɗa ka'idar, kuma sauran suna ginuwa a kanta.

Daga ka'ida zuwa algorithms: support vector machines da kernels #

Jerin abubuwan ciki ya kuma ambaci babobin algorithms da ke biye. Ɗaya yana da take “Support Vector Machines”, tare da sassan lamarin separable da na non-separable. Kowane ɗayan waɗannan sassan ya rufe primal optimization problem, support vectors da dual optimization problem. Wani kuma yana da take “Kernel Methods”.

Siffar babin support vector machines da kansa tana ba da bayani. Batutuwa uku iri ɗaya suna bayyana a ƙarƙashin kanun separable da non-separable: primal optimization problem, support vectors da dual optimization problem. Wannan maimaitawa tana nuna kulawa iri ɗaya a layi biyu, tambayoyi iri ɗaya da aka yi sau biyu a ƙarƙashin zato daban-daban. Wannan layi biyu yana kwaikwayon rarrabuwar consistent da inconsistent a babin PAC, don haka wani tsari ya bayyana a cikin littafin gaba ɗaya: lamari mai tsabta da farko, sannan lamarin da zato mai tsabta ya kasa.

Zan yi hattara da karanta abu da yawa daga teburin abubuwan ciki. Zai iya faɗa mana abin da littafin ya rufe da yadda ya raba abubuwan ciki. Ba zai iya faɗa mana yadda aka rubuta kowane sashe da kyau ba, kuma babu abin da waɗannan majiyoyi suka faɗa game da hakan. Abin da yake nunawa shi ne cewa harshen optimization a babin support vector machines, primal da dual, yana komawa ga bayanin sashen shimfiɗa na dabarun a matsayin haɗa kimiyyar kwamfuta da ra'ayoyi daga ƙididdiga, yiwuwa da optimization. Zaren optimization na wannan jeri ya bayyana a zahiri a cikin waɗannan taken sassan.

Littafin ya kuma keɓe babi daban ga kowanne daga cikin multi-class classification, ranking da regression. Wannan shawarar tsari ce da ta cancanci kariya. Da ya kasance da sauƙi a ɗauke su a matsayin ƙarin bayani ga binary classification. Ba kowanne babi nasa yana nuna cewa marubutan suna ɗaukar su a matsayin matsaloli daban masu nasu nazari, kuma ya dace da burin da aka faɗa na gabatar da batutuwa da yawa a haɗe maimakon gabatarwar da ke fifita ra'ayi ɗaya.

Abin da littafin ya zaɓa kada ya rufe #

Sashen da ya fi ban sha'awa a gabatarwar ya shafi abin da aka bari. Littafin, a cewar marubutan, gabatarwa ce ta gama-gari ga machine learning da za ta iya zama littafin bincike ga masu bincike da littafin karatu ga ɗalibai. Bayan sun yi wannan babbar da'awa, sun ƙara iyaka: sun zaɓi su bar wasu muhimman batutuwa, ciki har da graphical models da neural networks.

Littattafan karatu kaɗan ne ke faɗin haka kai tsaye, kuma zan iya cewa wannan gaskiyar ƙarfi ce. Mai karatu da ya san tun da wuri abin da babu zai iya tsara kansa a kan haka. Wanda yake buƙatar neural networks ya san ya nemi wani wuri, maimakon ya gano gibin a babi na tara.

An kuma faɗi dalilan. Barin ya kasance saboda taƙaitawa da kuma saboda rashin ƙwaƙƙwaran tabbacin ka'ida a halin yanzu ga wasu hanyoyi. An ba da dalilai biyu. Taƙaitawa iyaka ce ta zahiri ga kowane littafi. Dalili na biyu shi ne wanda ke faɗa maka wani abu game da falsafar littafin.

Haɗa shi da sashen shimfiɗa. Tabbacin koyo na ka'ida shi ne abin da aka tsara wannan littafi a kansa. Tabbacin ya dogara da sarƙaƙiyar concept classes da girman samfurin horo, kuma babobin ka'ida suna zuwa farko domin duk sauran an gina su a kansu. Hanyar da a halin yanzu ba ta da ƙwaƙƙwaran tabbaci tana zama a wuri marar daɗi a littafin da aka gina haka. Don haka barin ya biyo bayan babban alƙawarin littafin. Ya dace da ma'anar batun da marubutan suka zaɓa su buɗe da ita.

Akwai abin kulawa a nan. Rubutun ya ce “some methods”, don haka da'awar ta shafi rashin ƙwaƙƙwaran tabbaci ga wasu hanyoyi, ba ga dukan batutuwan da aka bari ba, kuma ya ce “current”, wanda ke ɗaure hukuncin da lokacin rubutawa. Zan ƙi jarabar taƙaita wannan da cewa “neural networks ba su da ka'ida”. Ba abin da majiyar ta faɗa ke nan ba. Ta ce tabbacin wasu hanyoyi babu shi a halin yanzu, kuma wannan, tare da taƙaitawa, shi ne dalilin da aka bar graphical models da neural networks.

Ga masu karatu na fasaha da ke zaɓar abin da za su karanta, sakamakon aiki mai sauƙi ne. Wannan littafi yana ba ka kalmomin tabbaci, kuma yana faɗa maka a sarari waɗanne fannoni bai rufe ba.

Abin da bugu na biyu ya canza #

Bugu na biyu ya fi sake bugawa da gyare-gyare kawai. Gabatarwar ta jera ƙarin abubuwa, kuma kowanne yana faɗa mana inda marubutan suka ga bugu na farko ya yi siriri.

Babban canjin shi ne cikakken babi kan model selection, babi na 4, a kan batun da marubutan suka bayyana a matsayin mai muhimmanci kuma wanda aka tattauna a taƙaice kawai a bugun da ya gabata. Harshen kai tsaye ne. Ba su da'awar cewa bugu na farko ya rufe shi da kyau ba. Sun ce an tattauna shi a taƙaice kawai kuma yanzu sun ba shi cikakken babi. Sassan sun haɗa da empirical risk minimization (ERM) da structural risk minimization (SRM), kamar yadda jerin abubuwan ciki ya nuna.

Me ya sa model selection ya cancanci wannan ɗagawa? Majiyoyin ba su faɗa ba, kuma ba zan ƙirƙiri dalili a madadin marubutan ba. Amma matsayin yana da ma'ana. Babi na 4 yana biye kai tsaye bayan babobin PAC framework da na Rademacher complexity da VC-dimension, wanda ya sanya shi cikin abin da gabatarwar ta kira babobi uku ko huɗu na farko da ke shimfiɗa tushen ka'ida. Don haka marubutan ba su liƙa model selection a ƙarshe a matsayin tunani na baya ba. Sun gina shi cikin tushen ka'ida.

Ƙarin da ke biye su ne sabon babi kan Maximum Entropy models, babi na 12, da sabon babi kan Conditional Maximum Entropy models, babi na 13. Sabbin babobi biyu gaba ɗaya faɗaɗawa ce mai girma, kuma yanzu littafin ya keɓe babobi masu bin juna ga waɗannan biyu.

Ƙarin bayanan (appendices) ma sun canza. Marubutan sun ƙara cikakken sashe kan Fenchel duality a appendix B kan convex optimization, kuma sun ƙara appendix E kan information theory. Littafin ya kuma haɗa da wani faɗaɗɗen appendix da ke gabatar da taƙaitaccen bita na linear algebra. Waɗannan canje-canje suna nuna hanya ɗaya: lissafin da ke tallafawa ya girma tare da babobin. Fenchel duality ya shiga appendix na convex optimization, kuma information theory ya samu appendix nasa. Zan ƙi cewa information theory yana nan musamman don tallafa wa babobin Maximum Entropy, tun da majiyoyin ba su faɗi irin wannan alaƙa ba, kuma ƙarin biyun sun bayyana a gabatarwar a matsayin abubuwa daban. Zuwansu a bugu ɗaya wani daidaituwa ce da zan bar ta haka.

Ga wa littafin yake, da yadda yake tsammanin a karanta shi #

Littafin ya faɗi masu karatunsa: an yi shi ne ga ɗalibai da masu bincike a machine learning, ƙididdiga da sauran fannoni masu alaƙa. Wannan jeri ya haɗa da ƙididdiga da sunanta, abin da ke maimaita batun sashen shimfiɗa cewa machine learning yana da alaƙa ta asali da nazarin bayanai da ƙididdiga.

Ya kuma faɗi abin da yake ɗauka a matsayin sananne. Ana tsammanin mai karatu ya saba da muhimman ra'ayoyin linear algebra, yiwuwa (probability) da nazarin algorithms. Wannan gajeren jeri ne, kuma ya dace da jerin sinadarai na sashen shimfiɗa, ko da yake ba daidai ba. Sashen shimfiɗa ya ambaci kimiyyar kwamfuta, ƙididdiga, yiwuwa da optimization a matsayin tushen dabarun. Abubuwan da ake buƙata kafin karatu sun ambaci linear algebra, yiwuwa da nazarin algorithms. Optimization ba ya cikin abubuwan da ake buƙata, kuma littafin ya magance hakan daga ɗaya ɓangaren: appendix B ya rufe convex optimization, tare da sabon sashen Fenchel duality. Ana tsammanin linear algebra a matsayin ilimin baya, kuma duk da haka littafin ya samar da faɗaɗɗen appendix da ke bita a kansa. Don haka abubuwan da ake buƙata da appendices sun haɗu da gangan, kuma mai karatu da linear algebra ɗinsa ta yi tsatsa yana da taƙaitaccen bita a cikin littafin ɗaya.

An kuma faɗi hanyar koyarwa a sarari. Kowane babi yana ƙarewa da jerin motsa jiki, tare da cikakkun amsoshi da aka gabatar daban. Wannan tsari ne mai ma'ana ga littafin da ke yi wa rukunin masu karatu biyu hidima. Ɗalibi zai iya gwada motsa jikin ya kwatanta da amsoshin, kuma mai bincike da ke amfani da littafin a matsayin madogara ba a tilasta masa wuce amsoshin da aka warware a kowane shafi ba. Ko amsoshin sun zo a matsayin wani juzu'i daban ko wata takarda, majiyar ba ta faɗa ba, sai dai cewa an gabatar da su daban, don haka zan bar shi a nan.

Idan aka haɗa su, masu karatu, abubuwan da ake buƙata, motsa jiki da appendices suna bayyana littafin da aka tsara don a yi aiki da shi daga farko har ƙarshe. Shekaru goma sha huɗu na koyarwa a Courant Institute suna bayan wannan tsari. Kwas ɗin da aka koyar akai-akai yana yawan samar da irin wannan kayan aiki, ko da yake gabatarwar ba ta da'awar wannan alaƙa a fili ba, kuma ina ba da shi ne kawai a matsayin karatun shaida, ba a matsayin binciken da aka tabbatar ba.

Amfani yana da faɗi, kuma littafin ya faɗi haka #

Littafin da ya kasance na ka'ida haka za a iya kuskure a ɗauka cewa yana nesa da aiki. Sashen shimfiɗa ya hana hakan da magana mai sauƙi: machine learning yana da faɗaɗɗen jerin amfani na zahiri. Wanda ya ambata shi ne text ko document classification, wanda ya haɗa da gano spam.

Gano spam misali ne mai kyau saboda dalilin da tsarin marubutan da kansa ya bayyana. Matsala ce ta document classification daidai a ma'anar misalin da ya gabata: takardu masu lakabi suna shiga, hasashe game da takardun da ba a gani ba suna fita. Abubuwan girman samfuri da ingancin lakabi sun shafe shi kai tsaye. Samfuri mafi girma yana sauƙaƙa aikin, kuma saƙonnin da aka yi wa lakabi ba daidai ba suna wahalar da shi. Yawan rukunan da ka iya kasancewa ƙanana ne, abin da, bisa dabarar sashen da ya gabata, shi ma yana shafar wahala. Ina amfani da abubuwan da littafin ya faɗa a kan misalin da ya faɗa, kuma ba na da'awar komai game da matatun spam a duniyar zahiri.

Akwai kuma hujja a ɓoye a matsayin wannan sashe. Marubutan sun gabatar da document classification a matsayin aikin koyo sannan nan take suka lura cewa ba shi kaɗai ba ne. Amfani yana da faɗi domin ma'anar da ke ƙarƙashinsa tana da faɗi. Duk abin da ya dace da “hanyoyin lissafin kwamfuta da ke amfani da gogewa don inganta aiki ko yin hasashe daidai” ya cancanta. Wannan gama-garin kuma shi ne dalilin da littafin ya damu da ka'ida. Tsarin da aka yi niyyar ya rufe faɗaɗɗen jerin amfani ba za a iya ɗaure shi da cikakkun bayanan ɗaya ba, kuma tabbacin da aka faɗa bisa sarƙaƙiyar concept class da girman samfuri su ne daidai irin waɗanda ke tafiya tsakanin matsaloli.

Cacar da littafin ya yi #

Haɗa zarurrukan tare sai matsayi mai daidaito ya bayyana, wanda ya fi abin da teburin abubuwan ciki na yawancin littattafan karatu za su iya da'awa. Littafin yana ma'anar koyo bisa gogewa da hasashe. Ya ambaci sarƙaƙiya uku, time, space da sample, a matsayin ma'aunan inganci. Yana ɗaure tabbaci da sarƙaƙiyar concept class da girman samfuri, kuma yana sanya ka'idar da ke samar da waɗannan tabbaci a gaba. Sannan yana gina babobin algorithms a kan wannan tushe kuma yana faɗa a fili waɗanne shahararrun fannoni yake barin su saboda tabbacin wasu hanyoyi, a yanzu, ba ƙwaƙƙwara ba ne.

Wannan caca ce a kan ka'ida a matsayin abin da ke dawwama. Na fahimci dalilin da ƙungiyar marubuta da suka shafe shekaru suna koyar da kwas na digiri na biyu za su yi ta. Cacar tana da farashi, kuma marubutan sun ambace shi: barin graphical models da neural networks shi ne farashin dagewa a kan ƙwaƙƙwaran tabbaci. Mai karatu ya kamata ya auna wannan farashi da gaskiya. Idan aikinka ya mamaye fannonin da aka bari, wannan littafi tushe ne ba jagorar aiki ba. Idan aikinka yana buƙatar tunani a kan ko mai koyo zai iya generalise daga samfuri mai iyaka, kalmomin littafin su ne manufarsa.

Ga masu kafa kamfanoni, darasi mai amfani shi ne mafi ƙarancin haske. Inganci da yawan bayanai suna da matuƙar muhimmanci ga nasarar hasashe, kuma wahalar aiki ta dogara da ingancin lakabi da yawan batutuwan da ka iya kasancewa. Waɗannan maganganu ne game da shigarwar da sau da yawa za ka iya rinjaya. Littafin bai yi alƙawarin cewa wani algorithm na musamman zai ceci samfuri siriri ko mai lakabin kuskure ba, kuma ya yi daidai da bai yi ba.

Ga masu karatu na fasaha, tambaya mafi ban sha'awa ita ce wadda gabatarwar ta bar ta a buɗe. Marubutan sun bar wasu hanyoyi saboda rashin ƙwaƙƙwaran tabbacin ka'ida a halin yanzu. Kalmar “current” tana tsara hakan a matsayin halin da ake ciki a lokacin rubutawa, kuma majiyoyin ba su faɗi ko zai canza, ko ta yaya ba. Me zai ɗauka kafin waɗannan tabbaci su zama ƙwaƙƙwara isa su cancanci babi? Littafin ba zai iya amsa hakan ba. Ta hanyar zama daidai game da abin da ya bari da dalili, aƙalla yana nuna inda za a duba.

Abin da Wannan ke Nufi a Kowane Fanni #

Dagewa a kan tabbacin statistical learning da sample complexity yana samar da iyakoki da za a iya aiki da su a fannonin software da bincike da dama.

Machine learning na ilimi da tsara manhaja #

Shirye-shiryen koyarwa dole ne su ƙi matsin lambar ɗaukar machine learning a matsayin prompt engineering na gwaji ko aikin haɗa frameworks. Kafa ɗalibai a kan PAC framework, Rademacher complexity da structural risk minimization yana gina ɗabi'un lissafi masu ɗorewa da ke tsira daga sauye-sauyen gine-ginen deep learning na ɗan lokaci.

Injiniyancin AI na kamfanoni da kimanta darajar bayanai #

Ƙungiyoyin injiniya da ke tura tsarin hasashe sau da yawa suna ware jari fiye da kima ga compute (girman model) yayin da suke watsi da sample complexity da amincin lakabi. Tun da generalization error yana ɗaure sosai da sarƙaƙiyar concept class idan aka kwatanta da girman samfuri, ƙungiyoyi suna samun ROI mafi girma ta hanyar saka jari a bayani (annotation) mai tsabta da binciken samfuri maimakon manyan models marasa regularization.

Binciken AI na sahun gaba da binciken models #

Yayin da tsarin dokoki kamar EU AI Act ke umartar iyakoki masu tabbas a kan tsarin yanke shawara na atomatik, tabbacin statistical learning da aka tsara a Foundations of Machine Learning ya zama dole don bin doka. Masu tsara dokoki da masu binciken tsaro suna buƙatar generalization bounds da za a iya tabbatarwa da tsayayyun ka'idojin model selection maimakon makin validation na ƙiyasi.

Kammalawa #

Foundations of Machine Learning ya kasance anga da ba za a iya maye gurbinsa ba a fannin da sau da yawa empiricism ɗanye ke jarabtar sa. Ta hanyar gina dukan manhaja a kan PAC learning, Rademacher complexity da VC-dimension, Mehryar Mohri, Afshin Rostamizadeh da Ameet Talwalkar suna tunatar da masu bincike cewa kyawun algorithm ba zai iya maye gurbin tabbacin lissafi ba. Barin da marubutan suka yi na neural networks da graphical models da gaskiya saboda rashin tabbaci na closed-form ba aibi ba ne - shi ne ƙarfin da ke ayyana littafin, yana raba abin da fannin zai iya tabbatarwa da tsauri da abin da yake lura da shi kawai.

Tambayoyin da Ake Yawan Yi #

Me ya sa marubutan suka bar neural networks da graphical models a bugu na biyu?

Marubutan sun bar neural networks da graphical models da gangan saboda taƙaitawa da kuma saboda rashin ƙwaƙƙwaran tabbacin ka'ida a halin yanzu ga da yawa daga cikin waɗannan hanyoyi. An tsara littafin sosai a kan algorithms da ke da tsayayyun generalization bounds.

Me ya bambanta sample complexity da time da space complexity?

Yayin da time complexity ke kimanta saurin gudanarwa kuma space complexity ke auna amfani da ƙwaƙwalwa, sample complexity yana ƙididdige yawan bayanan horo a lissafi da algorithm ke buƙata don koyon iyalin ra'ayoyi a cikin ƙayyadaddun iyakokin kuskure da amincewa.

Menene babban bambanci tsakanin empirical risk minimization da structural risk minimization?

Empirical risk minimization (ERM) yana zaɓar hypothesis da ke rage kuskuren horo kai tsaye, yana fuskantar haɗarin overfitting idan sarƙaƙiyar hypothesis ta yi yawa. Structural risk minimization (SRM) yana daidaita kuskuren horo da sarƙaƙiyar hypothesis (kamar VC-dimension ko Rademacher complexity), yana samar da iyakoki na yau da kullum a kan haɗarin generalization.

Manazarta #

Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar, (2018). Foundations of Machine Learning, Bugu na Biyu ⧉. The MIT Press.
Leslie G. Valiant, (1984). Ka'idar Abin da Za a Iya Koya ⧉. Communications of the ACM.
Vladimir N. Vapnik, (1998). Ka'idar Statistical Learning ⧉. John Wiley & Sons.

Bita ta ƙarshe 2026-10-10.

Bita ta ƙarshe 2026-10-27.

Sake buga wannan labari

Teburin abubuwan ciki na littafin machine learning: abin da yake faɗi — Sebastien Rousseau

Foundations of Machine Learning (MIT Press, 2018) yana ɗaukar machine learning a matsayin tsara algorithms na hasashe, bisa sample complexity da tabbacin ka'ida.

An lasisin wannan labari a karkashin Creative Commons Attribution 4.0 International. Sake bugawa na bukatar nuna asalin URL na asali.

Teburin abubuwan ciki na littafin machine learning: abin da yake faɗi — Sebastien Rousseau

Foundations of Machine Learning (MIT Press, 2018) yana ɗaukar machine learning a matsayin tsara algorithms na hasashe, bisa sample complexity da tabbacin ka'ida.

Originally published at https://sebastienrousseau.com/ha/2026-10-10-teburin-abubuwan-ciki-littafin-machine-learning/ by Sebastien Rousseau.
Licensed under CC-BY-4.0.

SEBASTIEN ROUSSEAU FOUNDER · ENGINEER