1. Introduction to Machine Learning Shao-Chuan Wang Research Center for IT Innovation Multimedia and Machine Learning Lab Academia Sinica 中央研究院資訊科技創新研究中心 多媒體與機器學習實驗室 NTNU 1 Shao-Chuan Wang, Academia Sinica
2. Outline What is involved in intelligence? Why is machine learning important? What can machine learning do? Overview of machine learning applications Challenges of machine learning Future of machine learning 2 Shao-Chuan Wang, Academia Sinica
4. What Is Involved in Intelligence? From Merriam-Webster: “intelligence”: (1) the ability to learn or understand or to deal with new or trying situations. (2) the ability to apply knowledge to manipulate one's environment or to think abstractly as measured by objective criteria Abstraction (finding the common patterns) V.S. Adaptation Learning is dynamic; e.g. a computer chess. 4 Shao-Chuan Wang, Academia Sinica
15. Shao-Chuan Wang, Academia Sinica A Tree Recognition Example (1/2) Suppose that you have never seen trees before, and I give you some “EXAMPLES” Trees examples ‘Not’ Trees examples 15
16. A Tree Recognition Example (2/2) I will ask you if these unseen photos are trees or not. YES Is it a tree? or NO Query Images 16 Shao-Chuan Wang, Academia Sinica (AND) How much confidence?
17. What Is Learning? (Mitchell 2002) Learning is to improve the performance measure P of the task T based on the past experience E. T: To recognize a tree P: Recognition accuracy E: The examples that I gave to you Two key elements of learning: Memorization of past experiences. “Generalization” ability (舉一反三). 17 Shao-Chuan Wang, Academia Sinica
18. Shao-Chuan Wang, Academia Sinica A Simple Algorithm: Nearest Neighbor For a given query image Find the nearest image to the query image in the database Assign the label of the nearest one to the query image. Query Tree! Difference = 13 Difference = 1.5 Difference = 11 Difference = 5.5 Difference = 10 18
19. What Were We Modeling? YES YES YES NO NO NO … TREE Human Concept (exist but unknown) Prediction: NO Infer A Machine (A learning algorithm) Query 19 Shao-Chuan Wang, Academia Sinica Training…
20. What if we do not have label ground truth?? (or labels are very expensive) 20 Shao-Chuan Wang, Academia Sinica
21. Shao-Chuan Wang, Academia Sinica Unsupervised Learning Clustering 21 。Each segment forms a “Cluster”. 。Pattern discover
23. Challenges of Machine Learning How do we model the “difference” between two images? Data Representation Difference Metric What is the “score” or “difference” function? How did we calculate the distance value in the tree example? Learning Does it model well? (can it accurately predict the seen data?) Does it generalize well? (can it be proved?) 23 Shao-Chuan Wang, Academia Sinica OR ?
24. Shao-Chuan Wang, Academia Sinica Example: sea bass or salmon? Suppose that we have only two kinds of fish, and we want a computer system that aids our distinction between sea bass and salmon. Process: 24 Take A Picture Computer Decision
25. Shao-Chuan Wang, Academia Sinica Example: sea bass or salmon? How do we describe a fish? (Data representation) What kinds of information can help us distinguish one from the other? Length, width, size of fins, tail shape, color, etc? How do we measure its distinctness under the chosen data representation? (Difference metric) E.g. if we choose length, than their “distinctness” can be measured using its absolute relative values. 25
26. Shao-Chuan Wang, Academia Sinica Example: sea bass or salmon? Assume that a fisherman (prior domain knowledge) told us that salmon is generally longer than a sea bass. We may use length as a feature to discriminate between them. But how? 26
27. Shao-Chuan Wang, Academia Sinica Example: sea bass or salmon? We use “past experiences” and we calculate a histogram of lengths for two types of fishes. Apply Nearest Neighbor to their average length. 27
28. Shao-Chuan Wang, Academia Sinica Example: sea bass or salmon? The difficulty comes from the ambiguity around the threshold value. Length itself is insufficient to “describe” the fishes. Use more features like width and color, etc. Other manipulation. E.g. use nearest neighbor to “median” of the length; will it be better? Let’s try one more feature: width 28
29. Shao-Chuan Wang, Academia Sinica Example: sea bass or salmon? We can use two features and wrote them down as a vector: Each fish image is represented as a 2-D feature vector: 29 Length : x1 Width : x2
30. Shao-Chuan Wang, Academia Sinica Example: sea bass or salmon? 30 There are still misclassified training examples
31. Shao-Chuan Wang, Academia Sinica Example: sea bass or salmon? Why use Line? We can use complex boundary, but we radically change the boundary just because of some heretics. => may not generalize well. 31
32. Shao-Chuan Wang, Academia Sinica Challenges of Machine Learning Conclusion on this example: We have to incorporate prior knowledge to decide which features we are going to use. At present, there is no universal learning machines. We want a feature that is invariant within certain specie but distinct between different species. There is a trade-off between complexity of decision model s and their “training errors”. 32
33. Shao-Chuan Wang, Academia Sinica The Future of Machine Learning Theoretic foundations of learning Scalability (Parallel) Robustness to dynamic environment 33
35. Thank you for your attention! 35 Shao-Chuan Wang, Academia Sinica
36. Shao-Chuan Wang, Academia Sinica Learning schemes Supervised learning: The tree example is a supervised learning problem. Supervised learning provides label ground truth. Unsupervised learning: Unsupervised learning DOES NOT provide label ground truth. Reinforcement learning: The way you train your pets. 36