Chapter 4 (Part 3):Artificial Neural NetworksCS 536: Machine LearningLittman (Wu, TA)Artificial Neural Networks[Read Ch. 4][Review exercises 4.1, 4.2, 4.5, 4.9, 4.11]• Threshold units [ok]• Gradient descent [ok]• Multilayer networks [ok]• Backpropagation [ok]• Hidden layer representations [ok]• Example: Face Recognition [today]• Advanced topics [today]Overfitting in ANNs (Ex. 1)Overfitting in ANNs (Ex. 2)When do you stop training?Face RecognitionTypical input images• 90% accurate learning head pose,and recognizing 1-of-20 facesLearned Weightshttp://www.cs.cmu.edu/tom/faces.htmlw0 w1 w2 w3For left, strt, rght, upAlternative Error FunctionsPenalize large weights:E(w) ! 1/2 "d in D "k in outputs (tkd-okd)2 + # "ij wji2Train on target slopes as well as values:Tie together weights:• e.g., in phoneme recognition networkRecurrent NetworksUnfolding: BPTTLearning Problems• Need to run a classifier on the ebaydata.– football vs. all others– football vs. woodwind• Also, describe 3 learning problems– what’s the data (features, instances)?– what’s the target?– how much data is available fortraining?Project MethodologyLots of good ideas for algorithms anddomains.The hard question is: “How will youevaluate it?”Ultimately, you need to present more thanone algorithm (and perhaps more thanone problem) and you’ll need some wayof saying what worked better.What’s the gold standard?Simple Project IdeaTake an interesting datasetCompare several learning approachesfor prediction• decision trees• ANNs• instance-based methods• SVMs• boostingTraining/Testing DataGetting labeled data is hard. Ideas:• association: use some features to predictothers. Example: spelling correction.• prediction: use past features to predictfuture.• existing data: UCI repository• natural labels: newswire categories• automatic checker: answer is tricky, butknow if you’re right.• artificial data: generator has a known ruleDatasetsWeather prediction dataSensor data: http://www.greatduckisland.net/Don Smith has log data on dialup usage(use weather, day, time, …)Other log data: network usage, CPU usageGPS data (campus buses, my car)Vision data (score a goal)Bio dataText ClassificationEasy to get lots of text: web, TRECdata, emailPredict topic, authorship, sentiment,style, affect, attitude.Reinforcement LearningPersonal favoriteCan be challenging to do well: generatedata or direct control• Optimize gas well production• Object tracking• “Tag” grid world• Rover control• animal behavior experimentsCompare approaches (direct policy search,value function learning, model-based, …)Active LearningTake a classic dataset.Explore the tradeoff between size oftraining set and generalization.Devise schemes for choosing items to“pay” for labels to maximizeaccuracy with minimum cost.Theory StuffAny interest? I’ll introduce instance-based methods next (Ch.
View Full Document