Kick-off Sweet Home Country Grammar
Mei-Lwun Announcements Go over names Go over assignment 01 Survey New Material Talked about a single state markov model Talked about probabilities and calculating probabilities Smoothing Training a model is based on generalizing over a set of training instances memorizing examples is consistent but doesn't generalize Occam's razor "all other things being equal, the simplest consistent explanation is best" P(feature|class) = n_featureandclass/n_class Since you rarely have all the training instances you must account for those you do not have a rare feature avoid zeroes in the probability distributions Smoothing is one way to do this. LaPlace smoothing using an m-esitmate assumes that each feature is given a prior probability,p, that is assumed to have been previously observed in a "virtual" sample of size m. P(feature|class) = n_featureandclass + mp /( n_class + m) For binary features, p is assumed to by 0.5. This is equivalent of seeing each word in a category once. Do an example Using a Markov model generatively Start with probability estimates and a good random number generator walk forward through the world Assignment 02 Data set is now available only 2 overlaps ./ACTION/Leon_JiShengyue.txt./ACTION/PirateOfTheCaribbean_SinhaPinaki.txt
./ACTION/PulpFiction_NasrRamzi.txt
./ACTION/PulpFiction_SutterNathan.txt
./ACTION/Terminator_BichutskiyVadim.txt
./ACTION/TheBourneIdentity_JiShengyue.txt
./COMEDY/AmericanSplendor_VernicaRares.txt
./COMEDY/DogDayAfternoon_VernicaRares.txt
./COMEDY/DumbAndDumber_DesaiChaitanya.txt
./COMEDY/DumbAndDumber_PirzadehPouria.txt
./COMEDY/Election_TikuZubin.txt
./COMEDY/Friends1_PirzadehPouria.txt
./COMEDY/Friends2_PirzadehPouria.txt
./COMEDY/Friends3_PirzadehPouria.txt
./COMEDY/Friends4_PirzadehPouria.txt
./COMEDY/TheresSomethingAboutMary_SaprooSameer.txt
./DRAMA/Braveheart_PartidaAugusto.txt
./DRAMA/Brick_JavanmardiSara.txt
./DRAMA/FightClub_NasrRamzi.txt
./DRAMA/GodFather_HabibiAmir.txt
./DRAMA/GodFatherII_HabibiAmir.txt
./DRAMA/Rocky_LinsteadEric.txt
./DRAMA/SixthSense_AlmishariMishari.txt
./DRAMA/TheQueen_JavanmardiSara.txt
./DRAMA/TheSting_LinsteadEric.txt
./DRAMA/TrainingDay_AlmishariMishari.txt
./FAMILY/FatherOfTheBride_SinhaPinaki.txt
./FAMILY/Shrek_BauTien.txt
./FANTASY/AI_SaprooSameer.txt
./FANTASY/Dune_TikuZubin.txt
./FANTASY/LOTR_ROTK_PartidaAugusto.txt
./FANTASY/ReturnOfTheJedi_BauTien.txt
./MISC/Pi_SutterNathan.txt
./MISC/PrettyWoman_BichutskiyVadim.txt
./MISC/TheMummy_DesaiChaitanya.txt
Break Face transformer More new material Classifying a sequence Probability calculation lots of multiplications of small number goes under zero Underflow prevention log(P(a)*P(b)) =log(Pa)+log(Pb) class with highest final unnormalized log probability score is the most probable Bring this back to classifying a movie genre