Machine Learning ICS 178.

Uploaded on:
Category: Business / Finance
Chess Playing is an exemplary AI issue. all around characterized issue. exceptionally intricate: troublesome for ... Netflix needs to prescribe concealed films to clients taking into account motion pictures he/she ...
Slide 1

Machine Learning ICS 178 Instructor: Max Welling

Slide 2

What is Expected? Class Homework/Projects (40%) Midterm (20%) Final (40%) For the tasks, understudies ought to make groups. This class needs your dynamic interest: please make inquiries and take part in discourses (there is no such thing as an idiotic inquiry).

Slide 3

Syllabus 1: Introduction: outline, illustrations, objectives, likelihood, contingent autonomy, grids, eigenvalue deteriorations 2: Optimization and Data Visualization: Stochastic inclination plummet, coordinate drop, focusing, sphering, histograms, disseminate plots. 3: Classification I: emprirical Risk Minimization, k-closest neighbors, choice stumps, choice tree. 4: Classification II: irregular timberlands, boosting. 5: Neural systems: perceptron, logistic relapse, multi-layer systems, back-spread. 6: Regression: Least squares relapse. 7: Clustering: k-implies, single linkage, agglomorative bunching, MDL punishment. 8: Dimesionality decrease: main segments examination, Fisher direct discriminant investigation. 9: Reinforcement learning: MDPs, TD-and Q-learning, esteem cycle. 10: Bayesian strategies: Bayes guideline, generative models, innocent Bayes classifier.

Slide 4

Machine Learning as indicated by The capacity of a machine to enhance its execution in view of past results. The procedure by which PC frameworks can be coordinated to enhance their execution after some time. Illustrations are neural systems and hereditary calculations. Subspecialty of computerized reasoning worried with creating strategies for programming to gain for a fact or concentrate information from case in a database. The capacity of a system to gain as a matter of fact — that is, to alter its execution on the premise of recently obtained data. Machine learning is a region of manmade brainpower worried with the advancement of systems which permit PCs to "learn". All the more particularly, machine learning is a strategy for making PC programs by the examination of information sets. Machine learning covers intensely with measurements, since both fields ponder the examination of information, yet not at all like insights, machine learning is worried with the algorithmic multifaceted nature of computational usage. ...

Slide 5

Some Examples ZIP code acknowledgment Loan application characterization Signature acknowledgment Voice acknowledgment over telephone Credit card misrepresentation recognition Spam channel Suggesting different items at Marketing Stock business sector expectation Expert level chess and checkers frameworks biometric recognizable proof (fingerprints, DNA, iris filter, face) machine interpretation web-look record & data recovery camera reconnaissance robosoccer thus on et cetera...

Slide 6

Can Computers play Humans at Chess? Chess Playing is a great AI issue all around characterized issue extremely mind boggling: troublesome for people to play well Conclusion: YES: today\'s PCs can beat even the best human Garry Kasparov (momentum World Champion ) Deep Blue Deep Thought Points Ratings

Slide 7

2005 DARPA Grand Challenge The Grand Challenge is a rough terrain robot rivalry contrived by DARPA (Defense Advanced Research Projects Agency) to advance exploration in the territory of self-ruling vehicles. The test comprises of building a robot equipped for exploring 175 miles through desert territory in under 10 hours, with no human intercession.

Slide 8

2007 Darpa Challenge

Slide 9

Netflix Challenge Netflix honors $1M for the individual who enhances their framework by 10%. The applicable machine learning issue goes under then name: "client proposal framework" or "community oriented separating". When you shop online at they suggest books taking into account what joins you are clicking. For netflix the applicable issue is foreseeing motion picture rating values for clients. aggregate of +/ - 400,000,000 nonzero sections (99% inadequate) films (+/ - 17,770) clients (+/ - 240,000)

Slide 10

source: Netflix Challenge # motion pictures # motion pictures with that mean # evaluations mean motion picture rating esteem # clients with that mean # clients # appraisals mean client rating esteem

Slide 11

The Task The client motion picture framework has numerous missing passages: Joe did not happen to rate "ET". Netflix needs to prescribe concealed motion pictures to clients in light of motion pictures he/she has seen (and evaluated!) before. To prescribe motion pictures we are being requested that fill in the missing passages for Joe with anticipated evaluations and pick the motion pictures with the most astounding anticipated appraisals. Where does the data originate from? Let\'s assume we need to foresee the rating for Joe and ET. I: Mary has evaluated all motion pictures that Joe has found in the past comparatively. She has additionally seen ET and evaluated it with a 5. What might you foresee for Joe? II: StarTrek that has gotten fundamentally the same as appraisals as ET from all clients. StarTrek was evaluated 4 by Joe. What might you anticipate for ET?

Slide 12

Your Homework & Project You will collaborate with 1 or more accomplices and execute calculations that we examine in class on the netflix issue. We will likely get high up on the leaderboard This includes both experimenting with different learning methods (machine learning) and in addition managing the vast size of the (information mining). Towards the end we will consolidate every one of our calculations to get a last score. Each class (beginning one week from now) we will have a presentation by 1 group to give an account of their advancement and to share experience. Perused this article on how great these frameworks can be:

Slide 13

Text Data Text corpora are broadly accessible in computerized structure nowadays (examined diaries, filtered daily papers, blogs,...). We can mine this content and find fascinating examples: what subjects are available in this article, what is the most comparable/pertinent article/website page in the corpus. Here the information has a fundamentally the same as organization: 99% inadequate word-tokens (+/ - 20,000) archives (up to 1000,000)

Slide 14

Text Data Each record is spoken to as a check vector for each of the words in the vocabulary: [20,5,3,0,1,0,2,0,0,0,5,0,...]. Things being what they are, in the article "president" showed up 5 times (would you be able to figure a subject?). Presently, we would prefer not to fill in missing sections (inadequate signifies "0", not missing). Our assignment is to discover for occasion which records are most comparable (archive recovery). Numerous more information networks have the same configuration: for occurrence quality expression information is a grid of qualities versus tests where the qualities speak to the "movement level" of the quality in that trial. Could we distinguish infections? "president" "the"

Slide 15

Why is this cool/vital? Cutting edge advancements create information at an uncommon scale. The measure of information duplicates each year. "One petabyte is comparable to the content in one billion books, yet numerous logical instruments, including the Large Synoptic Survey Telescope, will soon be producing a few petabytes every year". ( 2020 Computing: Science in an exponential world: Nature Published online: 22 March 2006) Computers overwhelm our day by day lives Science, industry, armed force, our social cooperations and so forth. We can no more "eyeball" the pictures caught by some satellite for fascinating occasions, or check each website page for some point. We have to trust PCs to take every necessary step for us.

View more...