15-874 Machine Learning from Neural Cortical Circuits
Carnegie Mellon University
Spring 2016.
Course Description
In the last few years, deep learning methods based on convolutional neural networks have produced state-of-the-art performance in object and speech recognition. These neural networks have also been found to provide a reasonable approximation for the neural representations in the primate visual systems. Yet, real biological neural networks are far more intricate and complex than the current neural networks. For example, only 5% of the synapses of a neuron in the real network in the visual cortex listen to bottom-up input signals yet current neural networks are primarily concerned with feed-forward computation. What are the functions of the other 95% of the synapses of a neuron? What could be the computational roles of the recurrent connections in the real biological circuits? What other learning rules are known or implementable in the real circuits? Can we develop new computational vision models and machine learning techniques from our knowledge of the cortical neural circuits? We will study current relevant machine learning, computer vision and biological papers to explore the answers to these questions. Students of all levels, from undergraduates to Ph.D. students are welcome, though priority will be given to more senior students. The course will involve paper presentation and discussion, research term projects by students, and lecture presentation by professors.
NOTE: Although we have not imposed prerequisites, the course is intended for students with some background in machine learning and neural computation, who are doing or are interested in doing related research. You should contact the instructor to see whether it is appropriate for you to enroll, or we can decide at the beginning of the course.
Course Information
Instructors
Office (Office hours)
Email (Phone)
Tai Sing Lee (Professor)
Mellon Inst. Rm 115 (office hour anytime or by appt)
tai@cnbc.cmu.edu (412-268-1060)
Class location and time: Friday 1:00-3:00 p.m. Mellon Institute Rm 115 Conference Room
Small group meeting: We will set up time for individual and small group meetings at MI during the week.
Blackboard: http://www.cmu.edu/blackboard/ (Students in all sections should use the same BB for access of course materials, grade center and announcements.
Paper Presentation (60 percent of the grade)
Presentation:
Each regular class in the course will be a series of papers on a particular topic presented by two students. As presenter, you will be expected to have read the papers listed for the day (including any available code) that are assigned to you, present them, and lead a discussion. Discuss with at least one week before the presentation and then by Wed the week of presentation.
Paper Critic:
Each regular class in the course will also have one or two designated "critics." As a critic, your job is to read the chosen paper and review it: list strengths and weaknesses, and start the discussion on the blog. You will post your review on the blog (24 hrs before the class) and then raise points from your review during discussion. The blog is currently a blackboard forum.
Summaries and Blog
For each regular class, everyone is required to make at least one post on the blog. For the blog, don't just summarize the paper; point out something others might have missed, give a critique, or respond to someone else's post. Alternatively, you can act as a critic, pose a question for the presenter 24 hours before the class. You can miss up to three blog postings in a semester.
Term research project (40 percent of the grade)
You are required to develop a project related to the MICrONS issues with me or by yourselves and discussed with and approved by me. You can work in a team of 2 or 3 students.
A project does not have to "work" in order to get a good grade.
But it requires a significant amount of work (remember this is a 12-unit class...)
You will need to produce a polished report/presentation, the sort that you would be willing to publish/present at a conference.
Feb 12, 11:59PM: One page Project Proposal Due
Mar 25: Mid-term Presentations
May 5-10: Final Project Presentations
Week 1 (Jan 22) Vision and the Visual System
We will give an overview of the visual system, particularly that of the primary visual cortex of cats and monkeys. We will also give an overview of the key computaitonal ideas and philosophy. Recommended reading to get you started:
We will discuss the basic theories of sparse coding and predictive coding. Later on, we will read more papers on the neural mechanisms and computaitonal models related to these two principles.
Cadieu, C. F., Hong, H., Yamins, D. L., Pinto, N., Ardila, D., Solomon, E. A., et al. (2014). Deep Neural Networks Rival the Representation of Primate IT Cortex for Core Visual Object Recognition. PLoS Comput Biol, 10(12), e1003963. http://doi.org/10.1371/journal.pcbi.1003963.s006
Cadieu, C. F., Hong, H., Yamins, D. L., Pinto, N., Majaj, N. J., & DiCarlo, J. J. (2013). The Neural Representation Benchmark and its Evaluation on Brain and Machine. Presented at the International Conference on Learning Representations 2013.
Khaligh-Razavi, S.-M., & Kriegeskorte, N. (2014). Deep Supervised, but Not Unsupervised, Models May Explain IT Cortical Representation. PLoS Comput Biol, 10(11), e1003915. http://doi.org/10.1371/journal.pcbi.1003915
Kriegeskorte, N., Mur, M., & Bandettini, P. A. (2008a). Representational similarity analysis - connecting the branches of systems neuroscience. Frontiers in Systems Neuroscience, 2(4). http://doi.org/10.3389/neuro.06.004.2008
Kriegeskorte, N., Mur, M., Ruff, D. A., Kiani, R., Bodurka, J., Esteky, H., et al. (2008b). Matching Categorical Object Representations in Inferior Temporal Cortex of Man and Monkey. Neuron, 60(6), 1126–1141. http://doi.org/10.1016/j.neuron.2008.10.043
Yamins, D. L., Hong, H., Cadieu, C. F., & DiCarlo, J. J. (2013). Hierarchical Modular Optimization of Convolutional Networks Achieves Representations Similar to Macaque IT and Human Ventral Stream. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, & K. Q. Weinberger (Eds.), Advances in Neural Information Processing Systems 26 (pp. 3093–3101).
Yamins, D. L., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D., & DiCarlo, J. J. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the National Academy of Sciences, 111(23), 8619–8624. http://doi.org/10.1073/pnas.1403112111
Week 12 (April 22) Sparse HMAX and Deep Residue Networks
One-shot learning with Bayesian Networks, by Maas and Kemp: http://repository.cmu.edu/psychology/972/
A Bayesian approach to unsupervised one-shot learning of object categories, by Fei-fei Li: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1238476