D Tnpak past ble or training als mals Me ML mmncbel Ape that ML modal duarre firm Whe dale a. cisnlés jes) modidly @ when you impat mew fale 15 me omodal, it fz I accordiry Jz Itz Sepsent mmodly pt Scanned with CamScannerBeer ee ole ee nerealels > Rader, herman adore > veduen Lk and > Im cuaser Work abeoneg Advar Ww © Lew errer alt 9 Treille cladly nccrvacy and spud a sa J me ie © Aulsmatzon —> Reduces ke mea oO a pa the fe | O wih Kamae aplicalion. E | Wealtheare ) G-Commurce , 1 Bambong @Frenhs and pattern enhix eclien Stock market und emodile bor peubecbong © conk mows improve went Disadvantages qm | Ocostl, 5 expeneive ood Lime —consumniny Le bald sebuild aod oper : Lintted Mies acd ae wuoures 5 Need a ot blz, y good: apelity dale dé duann sfieentl storages acd cloud sowie ae mot Aufprewnt ment —> Machinse aeplecey fava na Scanned with CamScannercsotptnviced pein femal ran a on ead i at = anamely dleclion sil he Bee labeled ie silken Geel Machine 25 ah q idanee es. ‘ : tcc Le? ES as} fe er PY ei leatrix ap PSE es 1) te Tr ade: ate | cupurviced Joan ™ Wades a “te oO ae @® Rareciey Gpotegeiced Jake Vip Ndmduicnl Ali opphicelions + pond ¢ Tiller , jatar @) Uns vieed Learning be ss i Machine ack nen dalz Ahab i, mith clarified er © en : io sande benana ds Me mlodall , ohreh machine Bras ono clan abt De ie le ile i ras 1a et Scanned with CamScanneri ret We past aap Bi oa ee," tra span eee orl danse form he yard pt mr agi fe vai same irmege Covreck °| unknown te “ey a dug A" Reimporearrraent learn 4 woe * '® Negalive a ® Positive ] ats oh alee ids. much ee ) each, aS ghd hay 1G) Semi Supsrviced Learning ft Th na sof Q om algorithm Bes an poe tae tobe hat He ifees J dalz a Unlabeled Saber fe Jehaack neces: Toke | Scanned with CamScanneray 4, ogy mu tee a oo DS modal Salhiny UE enbine dale ie. enki ddlas, ! acd fe Jiaining ‘wostead waigy peal, boy peo se imesmnentd focinig Hah ee dew tae dle ve bs be Anaimed 7 sence ab voould be Cath, bane ce Ux batch ceaininy vohue we Karn UX model Of fess faking huge dabx | 9 ona moda as Raimed we sila ‘yl Tt onls Lhe EET in ba mo eee les [model m tue dave Adz placed won modell , smnoda! get bese Ona DE mode i. toaind, we Dut the amodel jand once u one gu Sil ur model Se tuted papel) ec pub af on tke ene odel wll tury vohenever User we On thot saver our ME Im ‘ Bde ood gC Die guudliclian ¥ Cr nue oes pe elie veh Kap Ee Osurqut movies or Netples Js et io atl He a te a. Aaktar (aegis tohich ao ic. ig ead Ne machune ‘ » eat aul bh oat “ali 4 nol Js aly i she) ates Scanned with CamScanner“Yt ee (au boy oy (Vmalh) =~ pera updated dele alony sath old modal - Dicadvantager J or ® tok 4 dk = Umabl JS paecus a vt ® Hordoarne Lintkalsn > © Avatlabattle a Orkin Learning * ts don wally Se peat ly oe bah batch - venain red dela sn Pe ee ae ae hig de Bele, dae Ay wren dale i seek riety power dou a thonge Aiffuince He Scanned with CamScannerJa a teain th ae yohuia in omline daaxning , We can Apamieally m Sewer Teel | & H vm dztiprah J svmulalz conventabes ee Oy, ow ae wert Duman Ture ut mnoe , abaaa , Sit » goo } ' a E pi d on sotver , Wy au pedecbenry Cobb) an welll ar Jeawniny @ Noukube » Every tee we opr Youtule, vecteh = Video omy Paes back | O47 peck qt charged baud on purvions een Bice luc om call wrren Fae a ak 4 ut % CH) Aa tm) er le os hauliad intr Scanned with CamScanner© ‘ , : te ee OC saqunited Learning @ semi dupavised Learning, batler junebzon ® Reinbowes runt “Leatening ( Wwe have an a. 4 tohich adi im I envieon ment ah agent can take achitn pand his adion can immpa ck Le eovineona, Tn 2 police, Alege, UT agent fakes an aden ord Ue envimonmafe ged Le a rus state arch ge tore sawoaned Le Ue agent Heke ra0004 du, be a +ve seccarel | Gin be a ~Ve recon d ie oF can be molhing at that parlicstey Lire oh Bub the ap at ds continually acing im Dhue some Scanned with CamScannerSim ove given : a Predict Is values for a Lan fo ir = dessip Lise when Y as disculi eran when Y ds conlimous Value jo x, Wa, xwa, ab tact imslanee, 2 ite Covrespondiiny ole 4 Yy Scanned with CamScannerpnts dab yous wean UE Come mS A Lier which, gwen De ablebuls fa mes person atl prudict Whelkin that pasion 4 Sigh cel or foc mesle, And tha Gan evarmph Qo cule that you ave found. if imurme greta than ©, and Soviege ob qe Ahan ©, , [hin the person Ja | low ~iske die tke prsten de Ligh wick. (oe an cet po aa Example rene) Enample: Price fa ud car L, ay H cor atlribule Y awe +e, Yi pri a » = que) JO modu © paramalay a2 writes Leb us say that h y | “| pee hove only Looking ak Dts ami Dh P kage ke cy u on wmileage toe ane pa dich ny UE placed i} bared So given UR mileage | poice YY several eau yor Dave Ue Sle You con Come up wilh a furan , £0 that Men a mews cay Whee, orange Mo geen you cnn predict pre | So im ateration , you com up PE lands, which Sle OK 1% lb inet, and tk peramdiay, We moda \ s Y points 2. ee ter thal tmerdly US Laterihe an im =, liu Hag _— ] © kaa Scanned with CamScannerExample Dale Vraininy € xamle : = Ache, Auther le: A\B, AB oO rr ( ond emal (eq lange, rmellasan , small) ® ce Valued Cee: ne 5 woords im a tent), @® veal-valuad (eg hota ht) Hood ar“p) ‘ TThoead’ _Lingth Se f er | Skips known mes tery Home a ae s Unknown mess chest ee st Baise un old lorg ae Y < eu [Sek old long Wome er | reads mes Short home ee | Sik old orp bon News example, a\a x rs - e\% ov S. = a foe awe os adit olioe @) wee etn etapa meal Scanned with CamScannera | ee a—input: a eet J imsbances d,---dy ban iealane Ae a tek 4. feabioes aol / i Veclar da m,--2, 8 wh on imslane as a Scanned with CamScanner+ ae Oa Beha eS ir #e I Ht In this post, we will come throush some of the major challenges that you night d 1g your machine learning model. Assuming that you know what machine really about, why do people use it, what are the different categories of machin: : lear how the overall workflow of development takes place. \ What can possibly go wrong during the development and prevent you from gettin ! | A cual ceurst predictions? i 4 1 \ Hon Gh | So let’s get started, during the development phase our focus is 10 select a learning alo train it on some data, the two things that might be a problem are a bad algorithm or his” data, or perhaps both of them, | 1% Table of Content : \ Not enough training data. Poor Quality of data. Irrelevant features. 5 Nonrepresentative training data, $ Overfiting and Underfitting 1. Not enough training data Let's say for a child, to make him learn what an apple is. all it and say apple repeatedly. Now the child can recognize all sorts of apples. Well, machine learning is still not up to that level yet: it takes a lot of data tor mdst of hy algorithms to function properly. For a simple task. it needs thousands of examples) to m Something out of it, and for advanced tasks like image or speech recognition. it may lakhs(millions) of examples. t \ | \ 2. Poor Quality of data: i | Obviously, if your training data has lots of errors, outliers, and noise, it will make it i os: for your machine learning mode! to detect a proper underlying pattern, Henee,it will nol per well. it in every ounce of effort in cleaning up your training data. No inatter how good g and hyper tuning the model, this part plays a major role ih helping us make Scanned with CamScannerTwice) Dulin we image, we can sec that even if our model is “AWESOME” and we feed it with fata, the result will also be garbage(output). Our training data must always contain more vantiand less to none irrelevant features. erggit for a successful machine learning project goes to coming up with a good set of ures pn which it has been trained (often referred to as feature engineering ), which includes ture jselection, extraction. and creating new features which are other interesting topics to be Wered jn upcoming blogs. —— ae | a0 ‘onfgpresentative training data: 5 make, sure that our model generalizes well, we have to make sure that our trai : entative of the new cases that we want to generalize to. ‘fae MB ; . Pet {rain our mode! by using a nonrepresentative training set, it won't be accurate in predictions it jased against one class or a group. ing data should Let us say you are trying to build a model that recognizes the genre of music. One way ‘our training set is to search it on youtube and use the resulting data, Here we assume be’s search engine is providing representative data but in reality, the search will be rds popular artists and maybe even the artists that are popular in your location(if you ia you will be petting the music of Arijit Singh, Sonu Nigam or etc), Scanned with CamScannerample, say one day you are walking down a sirvet o b ‘you offer him something to eat but instead of cating he ‘but somehow you are safe, After this particular incident, you mig snot worth treating nicely. } ‘So this overgeneralization is what we humans do most of the time, and unfortunately Jearning model also does the same if not paid attention, In machine learnin, ‘overfitting jie model performs well on training data but fails to generalize well ‘Overfitting happens when our mode! is too complex. ‘Things which we can do to overcome this problem 1. Simplify the model by selecting one with fewer parameters, 2. By reducing the number of attributes in training data | 3. Constraining the model. 4. Gather more training data 5. Reduce the noise. What is underfitting? Yes. you guessed it right underfitting is the opposite of overfitting. It happens wheh our model fs 100 simple to learn something from the data. For E.G.. you use a linear model -on a Set with ; multi-collincarity it will for sure underfit and the predictions are bound to be inaccurate on tte training set 100. | Things which we can do to overcome this problem: it 1. Select a more advanced model, one with more parameters. 2. Train on better and relevant features. i 3. Reduce the const Conclusion : t i Machine Learning is al! about making machines better by using data so that we don’t need to code them explicitly. The model will not perform well iF training data is small. or ndisy with ‘errors and outliers, or if the data is not represemative(resulls in biased), consists oF ferelevaitt features(garbaye in, yarbage out), and lastly neither Wo simple(results in undeetitting) ‘nor fol complen(resulls in overfiting), Afier you have trained @ model by keeping the above pakamete ‘in mind, don’t expect that your model would simply generalize well 10 new eases You dkxy a ‘to evaluate and fine-tune it, how 0 do that? Stay {uned this is « topic that will be eave Scanned with CamScanner