Syllabus for cmp 464-02, cmp788-02

Schneider

4 hours, 4 credits  We will study neural networks and their applications to machine learning. We will cover the basic mathematical theory and explore neural network techniques. We will use the Python language as well as NumPy and Tensorflow libraries to create custom neural networks that have efficient mathematical algorithms.

We will use online books and tutorials. 
Prerequisites: Linear Algebra (MAT 313);Data Structures and Algorithms I (CMP338)                                          

Location: Gillet 231

Instructor: Robert Schneider

Contact Info: 

• email: robert.schneider@lehman.cuny.edu
• web page: comet.lehman.cuny.edu/schneider/
• office hours-Gillet 200
  
• Tuesday 8:30-9 and 12:30-2:00 
  
• Thursday  3:10PM-4:10 PM
  
•  and by appointment
  

Grading Policy:

• Programming and Theoretical Assignments 60%
• Final Project and Presentation--40%
  
• The assignments and projects may be done in groups of one, two or three but contributions of each individual must be specified and they may have to present their parts.

Course Objectives:

• proficiency in underlying programming paradigms used for neural net work.
• understanding of basic theory
  
• exposure to current applications in AI
  
•  exposure to some future directions
• basis for more advanced courses in neural networks and AI

Materials:
    We will use online materials. We will reference some  specific  materials in  relevant sections. There are many excellent courses and resources on the web on specific and advanced topics.

1. Good overview tutorial. We will program differently.
   
2. Convolutional Neural Networks for Visual Recognition-Spring 2017-Stanford
3. Deep Learning for Natural Language Processing-Stanford
4. Tensorflow for Deep Learning Research-Stanford
5. Deep Reinforcement Learning-Berkeley
1. neural nets to play games, drive cars etc
2. YouTube lectures
6. You tube of industry speakers discussing their applications of neural nets at NYC 2017 Machine Language Conference (MLCONF 2017)
7. KAGGLE-- Competitions with neural nets and public datasets
   

Tentative Course Calendar:


Week1,2,3: Introduction to the theory of neural networks, PC setups, Python and Numpy review/learning

• Setting up your computer and review/learn Python,Numpy

• A guide to setting up your computer
• A simple intro to some of basics of Python, Numpy
• useful later for Matplotlib and SciPy
• A more complete tutorial for Python
• A more complete tutorial for Numpy

• Perceptron and basic idea of going downhill
• basic neural net without hidden nodes
• supervised learning---review in week2
• basic learning rule
  
• if yi*((w,xi)-b)<=0: # here w and xi are vectors and xi is one of the input data. (w,xi) is a notation for the dot product. yi is the correct answer -1 or 1
• w= w+yi*xi
• b=b+yi
• the rule is saying that if the ouput is a different sign (or zero)than what want make it better in all respects for this particular data ( see what happens when yi is 1 vs -1 and you get a wrong sign)
  
• basic picture of deeper neural nets(hidden nodes);
  
• another
• linear algebra setup -- review in week2
  
• convolution as a mask in Photoshop-- prelude to convolutional neural networks
• relation to regression
• week 2 python/numPy program to graph cost functions using different costs--one dimensional features we did in class 9/5
  
• It actually works correctly. I was worried in class for nothing. I have added
  
• You can also get this program at https://github.com/bob509/GitTry2.git if you join GitHub. Supposedly good to be familiar with this for employment. It will be on the Class464Programs branch ( just to get you looking)
  
• Homework and Classwork Next time(9/12) -- We will have a long lab to see how you are doing
  
• HW&LAB(easy) Change the cost function look at graph for program from class of 9/5
  
• HW&LAB(hard if need to graph) Change the input layer in 9/5 program  to have two features for each x ( N inputs where each input  point is in two dimensions)
• should be easy to use minimize in higher dimensions
• hard to graph as in 4 dimensions-- could try scatter plots and color to indicate cost.
• How about 3 or more features. Try without graphing. Is there any way to visualize?
  
• (HW) Assignments to test basic knowledge of Python/Numpy-- in this link
  
• graphs  of some possible cost (some of these are not distances)
  
• as Maple files
• a dataset where images close by the square metric (L2) are mapped close to each other
• an explanation from Stanford with general info of use. You should  search ] for t-SNE to see specific discussion about what is happening in visualization.
  
• HW&LAB - do these cost functions in python/NumPy and just get values
• (Harder) graph them
  
• Tuesday 9/26
• HW due by email WE  WILL WORK ON THIS --no graph necessary-- Note you will use matrix multiplication in NumPy. Look up numpy.dot
  
• Change the input layer in 9/5 program  to have two features for each x ( N inputs where each input  point is in two dimensions)
• should be easy to use minimize in higher dimensions
• Try to learn the OR truth table. How about And. How about XOR 
• Can you add an extra bit so you have 3 feature inputs and then learn the XOR. Discuss what has happened.
• OCTOBER 3 notes
  
• as pdf
• as word
• October 10 notes
• Do a final version of the homework in red above by 10/17. We worked on the following program in class on October 10
  
• Annotated program as a Python File
• You should finish this program. For example for the OR gate if the input is (0,0) you should get a 0. If input is (0,1) you should get a 1 etc. Same for the AND gate. You should try to do this for the XOR but it wont work. You should get a number output using the weights  W and bias that the scipy minimize function returns (res.x[0]=w0,res.x[1]=w1,res.x[2]=bias). Calculate the outputs!!! (note the plural) as np.dot(X,W)+bias (where W[0,0]=res.x[0],W[1,0] = res.x[1] and bias =res.x[2]). You could now use a loop and if to find out which of 0 or 1 your answers are closest too. You can then decide that indeed you have learned the gate. I would like you to do this in numpy without using loops but with matrix manipulations.
  
• example of turning matrix into T or F in numPy without loops
  
• Try to make inputs with 3 features that can learn the XOR of the first two features. Note you will need 3 weights and one bias. You will have to change all of your routines to do this. You will have to give four initial values to the scipy minimize function.
• Please email me with difficulties as you progress

• October 24 Notes
• XOR with extra feature program
• Class notes on append and other  matrix operations
•  HW due October 31.
• Use cross-entropy for the cost in your program to try to  learn OR,AND and XOR
• At this point you can use my XOR type program if you want.
  

• October 31 notes:
• Lab to get cross entropy into cost
• November 7-- Gradient descent
  
• Maple program to illustrate how derivatives help you minimize the cost function by iterating
  
• one dimensional derivative to get direction of downhill
  
• but even with one weight have 2 variable minimization (w,b)
  
• maple program to illustrate
• pdf
  
• How do we figure best direction to go
• what is a direction-- line in a vector direction from point your at
• same maple program looks at this
• How do we make calculations to figure out the best direction and derivative in that direction
• chain rule on line from (w0,w1,w2,...,b)
• How do we even figure out partial derivative of our cost with respect to a weight or b.
• chain rule and in general it is called backpropagation -see chapter 2
• Homework: Create a NumPy program to implement gradient descent on functions. Model it from the Maple loop we have
  
• Try it on functions you know the minimum for
  
• make more elaborate functions to use
• Due November 14
  
• 
  
• November 14:Tensorflow
• Look at first two slides in Stanford course and several programs
• Several beginning programs:
• beg1
• beg1A
• beg2
• beg2a
• beg2b
• logistic regression for truth tables tensorflow without tensorboard hooks
  
• logistic regression for truth tables tensorflow with tensorboard hooks
  
• Homework -Due after thanksgiving: Multiply  4 matrices of dimensions (16,8),(8,4),(4,2),(2,1) in Tensorflow
• Part 2: You are successively going to multiply the original (16,8) matrix on the right hand side. Take the (16,4) outputs from the multiplication of the (16,8) and (8,4) and Apply a Logistics function on all  outputs. Then multiply this new matrix of shape (16,4) by the (4,2) matrix. Take these (16,2) outputs and apply a Logistics function on them. Now multiply this new (16,2) matrix by the (2,1) matrix.
  
• You should do all of this with matrix manipulations.
• How does this relate to a neural network. Can you make a graph that corresponds to these calculations.
• November 28th
• Continue tensorflow with logistic regression
• Udacity program that downloads letter image data and puts into form for neural nets.
• Pickled Data File if you cant afford memory
  
• Download the
  Zipped  Udacity Tutorial- From Machine Learning to Deep Learning
• Start watching the videos from these zipped tutorials.
  
• Lab for homework due
  
• December 5
• Program to recognize 10 letters A-J from Udacity with comments
• Use maximum likelihood with softmax function and one hot encoding
• illustration of tricky python to get one hot encoding of labels
• Final Homeworks- include regularization- due final time Dec 19 at 1:00PM
  
• Make a 2 layer neural net to classify the 10 letters: You should tune your parameters and number of nodes in the added layer. You should show your output and you should do better than with 1 layer. Write a little synopsis to tell me what tuning you did. See our online text and the Udacity tutorials for tuning.
  
• Make a 2 layer neural net to learn XOR
• Make an appointment or see me in class to test your knowledge of your work ( no grade without approval and will not be in much next semester)
  

Week 5-?:Zipped  Udacity Tutorial- From Machine Learning to Deep Learning

• 5 Assignments with multiple parts. Below are direct links but they are also explained in the video
  
• Assignment  1
• Assignment 2
  
• Assignment 3
• Assignment 4
• Assignment 5
• Assignment 6 (if have time)
  

Work on projects