CS 678 - Advanced Machine Learning and Neural Networks
Assignments – Winter 2014
Note: All written assignments (except homework) are to be done with a word processor and be neat and professional. You will hand in a hard copy. Good writing, grammar, punctuation, organization, etc. are important and will affect your grade.
Do a write-up of approximately ½ single spaced page each on three papers of your choosing in the area of machine learning which you will hand in at the beginning of class on the due date. Include the full reference for each paper. Also e-mail me PDF files for each paper. These papers should be ones you have not read before the semester (and areas we will not directly cover this semester) and the papers should cover three different areas in the field. This is a chance for us to explore. The majority of the papers should focus on innovative technical aspects of machine learning and not just applications of machine learning. As much as possible, choose papers which are relatively recent. In your half page, you should a) briefly describe the main approach and contribution of the paper and b) give your opinion on the specific strengths, weaknesses, and contributions of the approach. Grading will be based on the perceived effort and insights you have gained.
You will also give me a priority order on the three papers regarding which you would most like to present. I will let you know which paper you will present and will most likely have you present the highest priority paper, but may choose another if more appropriate. You will give a 15 minute presentation on the paper to the class including an explanation of the approach or concept proposed in the paper and also the most interesting aspects and insights of the paper. The goal is to educate the audience on the most interesting aspects of the research in the limited time you have. Be prepared to answer questions. We will do one presentation at the beginning of each class starting early February. The quality of the presentation will also be part of the grading. You are encouraged to use slides or any other appropriate teaching aids. I especially like when in addition to reviewing the main points of the paper, you also go beyond and add some or your own insights into interesting aspects of the approach.
Goals of this assignment include a) an opportunity to better learn how to read and analyze research papers, b) an opportunity to practice your oral presentation skills, and c) an opportunity through the presentations for all of us to gain new and diverse insights into machine learning areas.
Some good sources for potential papers include Proceedings of the International Conference on Machine Learning, Proceedings of Neural Information Processing Systems, Machine Learning Journal, Journal of Artificial Intelligence Research, Journal of Machine Learning Research, Neural Networks, IEEE Transactions on Neural Networks, Neural Computation, among others.
Projects: The three following projects are due as a hardcopy at the beginning of class on the due date. Do not attach your source code. The write-up would typically be about 4-5 pages.
Recurrent Neural Networks Project - BPTT
Due: Feb. 20
1) Implement BPTT on the platform of your choice.
2) Test your BPTT on the Delayed Parity Synthetic Task. This task has a single time series input of random bits. The output label is the parity (even) of n arbitrarily delayed (but consistent) previous inputs. For example, for DParity(0,2,3) the label of each instance would be set to the parity of the current input, the input 2 steps back, and the input 3 steps back. Make sure your BPTT first works for DParity (0,1) which is just the exclusive-or of the current and previous input. Then experiment with some different DParity(x,y,z) tasks, k values, number of hidden nodes, and other learning parameters, and analyze and discuss how it does. Begin with the parameters suggested in the slides, and alter those a bit. Use tables/graphs or whatever best explains your findings. What happens if k < z, equal z, or greater than z? Of course in an actual application you would not know z apriori.
3) Test BPTT on another time-series/recurrent task of your choice. Explain your exact network structure (i.e. a) # of input, output, and hidden nodes and b) exactly how you represented the inputs and outputs). Analyze and discuss your results with appropriate graphs and tables. Note that a number of time series problems can be found in the UC Irvine Database. Try to choose one which really needs recurrence to do well (not all do). This part can be challenging, but do your best to get reasonable results. You may also work with a fellow classmate(s) on this part if you wish.
Deep Learning Project
Due: March 27
1) You will solve the MNIST task using deep learning. We have the data in ARFF format here. There are lots of papers regarding MNIST which you can review if you would like. For consistency and comparison with each other, we will just use the given MNIST input features without initial pre-processing, though you are welcome to also try some pre-processing of the features to see what kind of results that allows. We will just use their given training set and test set for training, testing, and comparison.
2) Train on a standard single layer MLP with 200 hidden nodes and report test results as a baseline to compare with.
3) Implement an initial unsupervised deep network (Deep belief net or stacked sparse autoencoders) for preprocessing. If using attacked autoencoders make sure you include mechanisms to encourage sparsity. Use those features as an input to an MLP. (Optionally you can also try to refine the weights of the entire network with Backpropagation after your initial learning). Test the MLP with 200 hidden nodes, but you could also try other variations, based on the number of features you have, etc.). Try some variations on your deep network and MLP to get the best accuracy possible.
4) Analyze and discuss your results.
Model of your choice Project and Presentation
Due: At the beginning of class on the day you give your oral presentation of the model. DonŐt wait until you are done with Deep learning to start this. This should be started earlier in the semester.
1) Implement an ML model of your choice. This will be a model not previously implemented for another class. You may choose one of the models discussed in the class, or one that you have an interest in. I will have you propose a ranked list of 3 possible models and e-mail it to me no later than Feb. 4 (this list can overlap with the models you choose for the paper presentation above). You can give it to me sooner and those who get it in soonest will have highest priority if there are multiple people wanting to do the same model. I will then give you the OK to proceed forward. This will allow me to give you feedback on your choice, and allow some more diversity on the models we consider.
2) Test the model on at least 2 different data sets and for a reasonable spread of the learning algorithm parameters. Analyze and discuss your results.
3) Do a creative experiment with the model and discuss your findings.
4) In addition to your written report, you will prepare a slide presentation to be given to the class at the end of the semester. The talk should motivate your model, review how it works, discuss results on your tasks, and give your overall analysis and recommendation for potential improvements of the model.