Lecture 1: Introduction CS 189 (CDSS offering)
2022/01/19
Course staff
Copyright By PowCoder代写 加微信 powcoder
Enrollment questions:
General course information
Course website: https://www.eecs189.org/
• If and when permitted, this course will be fully in person: lectures, discussions, office hours, exams, …
• Relevant prerequisites:
• Strong background in probability (CS 70, Stat 134, or similar)
• Strong background in vector calculus (e.g., can you take the gradient of a matrix vector product)
• Strong programming skills in Python (e.g., can you learn new libraries quickly)
• Check your background with HW0 (already out!) and Dis 0 next week
Lectures MWF 1-2pm
• Lectures are on Zoom for now, hopefully in Evans 10 later
• Lecture recordings will be available some time after the live lecture
• There may be some guest lectures which may not be recorded
• Professor Shewchuk’s lectures will differ in both content and ordering
• Lectures will be a mix of slides and board work
• Both the slides and notes for board work will be posted after lecture
Discussion sections and office hours
https://eecs189.org/calendar
• You are encouraged to attend any discussion section that you like that has room
• It is very important that you read the office hours policy on Piazza
• You should come to OH prepared and with reasonable expectations
• You should actively look for other students working on the same problems
• You will be limited to a 10 minute window when there is a queue
• Please don’t go to any discussions or office hours for Prof. Shewchuk’s class
Homework assignments DSP students:
• There are seven homework assignments total, released fortnightly
• You will have ~1.5 weeks to complete each homework: released Wed, due Sun
• No late homework will be accepted for any reason, but we will drop your two lowest homework grades
• The remaining five assignments are each worth 10% of your overall grade
• You are encouraged to discuss problems, but the code/writeup must be your own
— infractions will result in (at least) an immediate zero on the assignment
• HW0 is already out on the course website
DSP students:
• There is one midterm exam and one final exam, both in person if permitted
• Midterm (worth 20%): Wednesday, 3/9, 7-9pm (no lecture that day)
• Final exam (worth 30%): Tuesday, 5/10, 8-11am
• Midterm location is Dwinelle 145, final exam location is TBA
• There are no alternate exams: if you miss the midterm, your final is worth 50%; if you miss the final, you receive an incomplete grade
• Exam infractions are serious are will result in (at least) significant points deducted
• Definitely do not go to the exams for Prof. Shewchuk’s class
• This course will be curved at the end after all grades have been computed
• Do not assume that the final grade distribution will necessarily follow historical precedent, e.g., any particular previous semester, previous instructor, …
What is machine learning?
What is machine learning?
• Machine learning has three core components: model, optimization, and data
• The model is a function from inputs to outputs, but it is not programmed by hand
• Instead, the model has parameters that will be optimized (learned)
• The optimization algorithm finds (learns) good parameters — more on this later
• Roughly speaking, parameters are good if they are a good fit for the data • The data typically, but not always, consist of input-output examples
• The goal is for the model to generalize to new data that it wasn’t fit on 10
Types of machine learning problems (non exhaustive)
regression
classification
A model for classification
OxOxtOe0 return O
a model is a parameterized function:
else return x
OTx e o xO
Examples of machine learning problems Are these classification, regression, either, or neither?
• Recognizing digits
• Determining whether or not an email is spam
• Predicting the price of a stock six months from now
• Predicting the rating of a movie by a user
• Determining credit worthiness for a mortgage or a credit card transaction …
A classification example
Hastie, Tibshirani, and Friedman. “The Elements of Statistical Learning”.
A linear classifier
“Nearest neighbors” classifiers “I will judge you by the company you keep”
1 nearest neighbor
15 nearest neighbors
The “Bayes optimal” classifier
Another classification example Digit recognition on the MNIST dataset
• Has a training set of 60000 examples and a test set of 10000 examples
• Each digit is a 28 ! 28 pixel grayscale image
• Test error rates of best systems are under 0.5%
A linear model for digit classification?
• Images: e.g., 28 (height, in pixels) ! 28 (width) = 784 dimensional
• This is actually a tiny image! A somewhat larger image could be, e.g.,
224 ! 224 ! 3 (RGB) = 150528 dimensional 19
Training set vs. test set vs. validation set error
• Training set error: this is what we train the classifier to minimize
• Test set error: this is what we actually care about — how well does the classifier
generalize to new data?
• If the test set error is much greater than the training set error, then the classifier
has overfit to the training set
• Overfitting can be combated by holding out part of the training set and evaluating
the model on this part — this is a validation set
• Validation sets can be used to decide when to stop training and to choose
hyperparameters (e.g., the number of neighbors to use) 20
Bias vs. variance of learning algorithms A classical story we don’t totally believe anymore…
A closing question
On the “classical” view of machine learning
“predict y from x” But what is x?
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com