15-387/86-375/675 Computational Perception

Carnegie Mellon University

Fall 2022

Course Description

In this course, we will study the biological and psychological data of biological perceptual systems, mostly the visual system, in depth, and then apply computational thinking to investigate the principles and mechanisms underlying natural perception. You will learn how to reason scientifically and computationally about problems and issues in perception, how to extract the essential computational properties of those abstract ideas, and finally how to convert these into explicit mathematical models and computational algorithms. The course is targeted to both neuroscience and psychology students who are interested in learning computational thinking, as well as computer science and engineering students who are interested in learning more about the neural and computational basis of perception. Prerequisites: First year college calculus, differential equations, linear algebra, basic probability theory and statistical inference, and programming experience are desirable.

Course Information

• Class location and time: WEH 2302 Monday/Wednesday 1:25 p.m - 2:45 p.m.
• Class recitation and journal club: Class Zoom. Friday 1:25 p.m - 2:45 p.m.
• Website: http://www.ni.cmu.edu/~tai/cp22.html (course info and readings )
• Canvas: Lecture materials and Information would be on Canvas.

Recommended Textbook

• Handouts on Canvas .
• Frisby and Stone Seeing: The computational approach to biological vision . MIT Press, 2010 (recommended).

Classroom Etiquette

• Refrain from using laptop and cell phones during class.

Grading Scheme 15-387

• Total points: 100
• Grading scheme: A: > 88, B: > 75. C: > 65. D > 50.

Grading Scheme 86-375

• Total credit for 86-375: 100
39 points out of top three of five problem sets.
• Flex Requirement: problem set (13), a term project (13) OR term paper (13) OR Journal Club (13)
• Grading scheme: A: > 88, B: > 75. C: > 65. D > 50

Grading Scheme 86-675

• Total credit for 86-675: 100
• Journal Club at least 10 times attendance, 2-3 presentations.
• Term project can be used to replace one or two problem sets, depending on the scale of the project.
• Grading scheme: A > 88, B: > 75. C: > 65

Homework

• There will be 5 homework assignments involving Pytorch (Matlab?). The focus is on performing experiments on existing codes rather than coding algorithms from scratch.
• Each student will have 7 days grace period for late homework. This grace period can be used for one or multiple assignments. Use it wisely and you cannot ask for more. CANVAS or gradescope submission after the starting of class time of due day is considered late by one day.

Term Project

• Term project option is an available for 86-375 and 675 students counts for 25 and 26 points. 15-387 students can do a term project to replace one of their assignment (max: 13 points). A term project must involve some computational experiments, using either downloadable softwares or programs you develop. All the codes should be documented and archived in github and submitted together with a term paper (6-8 pages) materials in a different pdf/doc file and/or matlab zip files. Term project should take about 30 hours to complete. Undergraduate student should work on his/her own project. Graduate students can work on team on a larger scale project but need to be approved in advance. Project proposal is due by midterm. Students are encouraged to discuss project ideas with professor from the very beginning of the semester.

Term Paper

• Term paper option is an available for 86-375 students and counts for a max of 26 points. It should be an extensive in-depth review of a particular topic to be approved by the professor before midterm. The paper will be about 8 pages in NeuRIPS format. Paper proposal is due by midterm. The student is required to give a powerpoint presentation in person or on zoom at the end of the semester.

Journal Club

• Journal club option is an available option for 86-675 students and count for 25/26 points toward the total grade. Each student is expected to present three to four times during the course of the semester and participate in 90% of the journal club discussion.

Examinations

• There will be a midterm (10 points) and a final exam (15 points) to test materials covered in the lectures and homework assignments.
• There would be 12 in-class exercises given out in random time, each worths 1 point. Full credit: 10 points.

Syllabus

Reading (relevant, but optional reading)

Week 1 (Lectures 1 and 2) Observations, Theories and Computational Philosophy

• What is the purpose of perception? What do we see what we see? What does seeing mean? How do we formulate theories and computational theories of perception? These are the questions we will explore and think about. We will read some classic papers by David Marr and Gibson. This week's journal club, Sunreeta Bhattacharya presented the classic Science and Royal Society papers by David Marr.
• Marr, D. (1982) Chapter 1: The Philosophy and the Approach Vision 8-38 .
• Gibson's ch 8 of The Ecological Approach to Visual Perception Houghton Mifflin Press. 1979

Week 2,3 (Lectures 3, 4, 5) Retina, Pyramid and Neural Network

• Visual perception starts with the eyes and the photoreceptors. However, there is already sophisticated computation in the retina. We will read some classic and the modern papers on retinal processing, cover some basic background on frequency analysis, pyramid representation, as well as the current computational approach (via deep learning) for modeling retinal processing. We will do a problem set on retinal processing, and explore its relationship to some visual illusion and perception. This week, Jingkai Wen will present Gollisch and Meister (2010) and Joshua Kosnoff will present Maheswarantha ...Ganguli and Baccus (2018).
• Lettvin, Maturana, McCulloch and Pitt. (1959) What the frog's eye tells the frog's brain Proceedings of the IRE 1940-1959 .
• Gollisch and Meister (2010) Eye Smarter than Scientists Believed: Neural Computations in Circuits of the Retina Neuron 65: 151-164.
• Maheswaranathan, Kastner, Baccus and Ganguli (218) Inferring hidden struture in multilayered neural circuits. PLOS Computaitonal Biology 14(8):e1006291
• Maheswaranthan, .... Ganguli and Baccus (2018) Deep learning models reveal internal structure and diverse computations in the retina under natural scenes bioRxiv, June 8, 2018.
• McIntosh, Maheswaranathan, Nayebi, Ganguli and Baccus (2016) Deep Learning Models of the Retinal Response to Natural Scenes NIPS
• Perdreau, F. & Cavanagh, P. (2011). Do artists see their retinas? Frontiers in Human Neuroscience, 5:171

Week 4 (Lecture 6,7,8) Lightness perception and Intrinsic Images

• Our perception of brightness (or lightness) and color is not determined what are sensed by the retina, but in fact an interpretation of the ligthness and color properties of the object surfaces in the world. We will explore the classic theory of retinex as well as modern computational theory of intrinsic images for understanding lightness and color perception, culminating in a problem set on these issues. Shaurjya Madal will present Janner-Tenenbaum's (2017) NeurIPS paper, Madeline Davis will present Ma-Torralba 2018 ECCV paper.
• Adelson, Ed, (2000) Lightness Perception and Lightness Illusion The New Cognitive Neuroscience, Gazzaniga ed. MIT Press. (2000).
• Land, E, (1977) The retinex theory of color vision Scientific America 1977
• Horn, B, (1974) Determininng lightness from an image Computer Graphics and Image Procwssing. 1974
• Morel JM, Petro AB, Sbert C. (2010) IEEE Trans Image Process. 19(11) 2825-37.
• Tappen, Freeman and Adelson (2005) Recovering intrinsic images from a single image IEEE PAMI. 27(9): 1459-1472.
• Michael Janner, Jiajun Wu, Tejas D. Kulkarni, Ilker Yildirim, Joshua B. Tenenbaum (2017) Self-Supervised Intrinsic Image Decomposition. NeurIPS.
• Wei-Chiu Ma, Hang Chu, Bolei Zhou, Raquel Urtasun and Antonio Torralba1 (2018) Single Image Intrinsic Decomposition wihtout a Single Intrinsic Image. ECCV

Week 5 (Lecture 9,10). Perception of 3D shapes

• Shading as well as many other cues allow us to infer 3D shapes. Intrinsic images of shading is a consequence of illumination on 3D shape, suggesting illumination and 3D shapes are further decompositon of shading images. We wonder whether how 3D shape as an intrinsic image or objects is represented in the brain. Is it 2D images, 2.5D sketches or 3D solids. Emily Lopez explores whether the brain might encode some intrinsic images by reading Ramanujan Srinath et al's paper on solid shape coding. Rachel Hagani will read the classic paper by H H Bülthoff , S Y Edelman, M J Tarr on "How are three-dimensional objects represented in the brain?" while Yihan Zhang will read a deep learning paper by Jiayun Wu that might have implications on our understanding of the brain.
• Ramaujan Srinath, A. Emonds, O. Wang, A. Lempel, E. Dunn-Weiss, CE, Connor, K. Nielsen Early Emergence of Solid Shape Coding in Natural and Deep Network Vision. Current Biology, 31, 51-65. 2021.
• S Y Edelman, M J Tarr on "How are three-dimensional objects represented in the brain?" Cerebral Cortex 1995 May-Jun;5(3):247-60.
• Zhang, X, Zhang, Z, Zhang C, Tenenbaum J, Freeman W, Wu, J. Learning to Reconstruct Shapes from Unseen Classes NeurIPS 2018

Week 5 (Lecture 9,10). Source Separation and Representation Learning

• Lightness perception tells us that we perceive the representations of the properties of the world that we care about. But how does a brain decide what properties we should care about and how these properties are represented in the brain? We will explore the principles of Bayesian inference, blind source separation, efficient coding, autoencoding and and statistics of natural scene structures to understand these issues.
• Olshausen and Field (1996) Emergence of simple-cell receptive field properties by learning a sparse code for natural images Nature 381: 607-609.
• Olshausen and Field (2004) Sparse coding of sensory inputs Current Opinions in Neurobiology 14:481-487.
• Barlow (1954) Possible principles underying the transformation of sensory messages Sensory Communication MIT Press, p217-234.
• Barlow (2001) Redundancy reduction revisited Network: Computation in Neural Systems 12: 241-253.
• Torralba and Oliva (2003) Statistics of natural image categories Network: Comptuations in Neural Systems 14: 391-412.
• Hinton, GE and RR Salakhutdinov (2006) Reducing the dimensionlity of data with Neural Networks Science 313, July 28, 504-507.

Week 6 (Lectures 11, 12) Perceptual inference: contours, depth, surfaces

• Our visual system is not just trying to "represent" the outside world, but to make inference. Representations are created to facilitate inference. In this segment of the course, we will discuss perceptual inference in the context of texture and surface perception, and later on, in later lectures, to motion and cue-integration.
• Marr and Poggio (1976) Cooperative Computation of stereo disparity. Science 194. 283-287.
• Belhumeur and Mumford (1992) A Bayesian treatment of stereo correspondence problem using half-occluded regions. CVPR
• Zhang et al. (2016) Relating functional connectivity in V1 neural circuits and 3D antural scenes using Boltzmann machines Vision Research 120: 121-131.
• Weiss, Simoncelli and Adelson (2002) Motion Illusion as optimal percepts Nature Neuroscience 5(6): 598-.
• Tom Albright. (2012) On the Perception of Probable Things: Neural Substrates of Associative Memory, Imagery, and Perception Neuron 74. 227-245.

Weeks 7 and 8 (Lectures 13, 14, 15, 16) Perceptual Organization: Grouping, Gestalt, Content and Style

• In addition to inferring "visible" physical properties of the world, the brain also tries to organize the sensory information into parsimonious descriptions to infer more abstract and global properties or summary statistics of the world such as boundary, texture, style, and contents. In this segment of the course, we will study Gestalt school of thoughts and models for extracting these properties.
• Bela Julez (1981) Textons, the elements of texture perception and their interaction. Nature 290. 91-97.
• Heeger and Bergen (1995) Pyramid-based texture analysis/synthesis SIGGRAPH 1995
• Portilla and Simoncelli (2000) A parametric texture model based on joint statsitics of complex wavelet coefficients Internal journal of computer vision 40(1), 49-71.
• L. A. Gatys, A. S. Ecker, and M. Bethge (2015) Texture Synthesis Using Convolutional Neural Networks NIPS 28
• L. A. Gatys, A. S. Ecker, and M. Bethge (2017) Texture and art with deep neural networks Current Opinions in Neurbiology 46, 178-186.
• L. A. Gatys, A. S. Ecker, and M. Bethge (2016) Image Style Transfer Using Convolutional Neural Networks CVPR 2016
• Freeman J, Simoncelli EP. (2011) Metamers of the ventral stream. Nature Neuroscience.
• Kovacs, I., Papathomas, T., Yang, M. and Feher, A. (1996). When the brain changes its mind: Interocular grouping during binocular rivalry. Proceedings of the National Academy of Sciences, 93(26), pp.15508-15511.
• Bilge Sayim and Patrick Cavangah (2011) What line drawings reveal about the visual brain.

Week 9 and 10. (Lecture 17, 18, 19, 20) Analysis by Synthesis and Predictive Coding

• Early computer vision approach emphasized on analysis by synthesis. This framework can be generalized to conceptualize the recurrent interaction in the hierarchical visual system and perception as inverse graphics. We will study these theories and their neural foundation, as well as to see how these principles can be extended to self-supervised learning based on prediction principles.
• Van Essen, Anderson and Felleman (1992) Information processing in primate visual systems: an integrated approach Science 5043: 419-423.
• Fellman and Van Essen ( 1991) Distributed Hierarchical Processing the the Primate Cerebral Cortex Cerebral Cortex 1-47.
• Mumford, D (1992) On the computational architecture of the neocortex Biological Cybern. 66: 241-251.
• Rao and Ballard (1998) Predictive coding in the visual cortex: a functional interpretation of some oextra-classical receptive field effects Nature Neuroscience 2(1), 79-87.
• Lee and Mumford (2003) Hierarchical Bayesian inference in the visual system J. Optical Society of America 20(7), 1434-1448.
• Lee, T.S. (2015) The Visual System's Internal Models of the World Proceedings of the IEEE Vol 103, issue 8, 1359-1378.
• Lotter, Krieman and Cox (2020) A neural network trained for prediction mimics diverse features of bioloigcal neurons and perception Nature Machine intelligence vol 2, 210-219.
• Ilker Yildirim, Mario Belledonne, Winrich Freiwald, Josh Tenenbaum (2020) Efficient inverse graphics in biological face processing Sci. Adv. 2020; 6 : eaax5979 4 March 2020
• T. D. Kulkarni, W. F. Whitney, P. kohli, J. Tenenbaum, Deep convolutional inverse graphics network, in Proceeding of the Advances in Neural Information Processing Systems (NIPS, 2015), pp. 2539–2547
• Kar K, Kubilius J, Schmidt K, Issa EB, DiCarlo JJ. Evidence that recurrent circuits are critical to the ventral stream’s execution of core object recognition behavior. Nature Neuroscience. 2019. doi: 10.1038/s41593-019-0392-5.

Week 11 (Lecture 21, 22) Compositional Theory, Scenes, Objects and Parts

• Compositionary Theory argues the brain or the visual system is organized in a recursive nearly decompositional system that best models the compositional nature of parts, objects and scenes in the natural world, their flexible combination and deformation. We will explore some classical and modern theories on composition, exploring the linkage between modern deep neural networks and symbolic AI from this perspective, and the connection between language and vision.
• E. Bienenstock, S. Geman, and D. Potter. Compositionality, MDL Priors, and Object Recognition NIPS Advances in Neural Information Processing Systems 9. 1998.
• S. Geman, Hierarchy in machine and natural vision. Proceedings of the 11th Scandinavian Conference on Image Analysis, 1999.
• Yuille. Towards a Theory of Compositional Learning and Encoding of Objects 1st IEEE Workshop in Information Theory in Computer Vision and Pattern Recognition. ICCV 2011.
• S.C. Zhu and D. Mumford (2006) A Stochastic Grammar of Images Foundations and Trends in Computer Graphics and Vision 2(4): 259-362.

Week 12 (Lecture 23, 24) Attention and Awareness

• Information processing bandwidths of the brain is limited. To deal with the complexity of the world, perception is made to be dynamics and selective. We will consider both biological and computational aspects of attention, and their connections to awareness and consciousness on the one hand, and to attention and self-attention in neural networks.
• Grace Lindsay (2020) Attention in psychology, neuroscience and machine learning Frotnier Computational Neuroscience April 2020.
• Luo and Maunsell (2019) Attention can be subdivided into neurobiological components corresponding to distinct behavioral effects PNAS 116(52) 26187-26194.
• Eric Knudsen (2018) Fundamental components of attention. Annual Review of Neuroscience
• Olshausen, Anderson and Van Essen (1995) Mutliscale dynamic routing circuit for forming size- and position-invariant object recognition J. Neuroscience 2:45-62.
• Sabour, S., Frosst, N. and G. Hinton (2017) Dynamic routing between capsules NIPS

Week 13 (Lecture 25, 26) Perception, Art and Beauty and Review

What is beauty? Is it just something in the eyes of the beholders, or is it something universal? In this final week of the class, we hope to explore the concepts and computational theories of beauty and art and how they might be related to art, and the computational principles underlying perception that we have studied in the course.
• Cavanagh, P. (2005) The artist as neuroscientist. Nature, 434, 301-307.
• Schmidhuber Jurgen. (1997) Low complexity art.
• Schmidhuber Jurgen. (2008) Driven by Compression Progress: A simple principle explains essential aspects of subjective beauty, novelty, surprise, interestingness, attention, curiosity, creativity, art, science, music, and jokes!