22 March 2010
10.30am - 6pm
Gatsby Unit Quinquennial Symposium
REGISTRATION IS REQUIRED
TO REGISTER, PLEASE EMAIL: firstname.lastname@example.org before 15 March 2010
We are delighted to announce the 2010 Gatsby Unit Quinquennial Seminar, with talks by distinguished researchers in theoretical neuroscience and machine learning.
The symposium will start at 10:30am on Monday 22nd March in the basement Lecture Theatre, 33 Queen Square, London WCIN 3BG
All are welcome. Lunch and tea will be provided
Daniel Wolpert 10.30 – 11.30
Department of Engineering, University of Cambridge
Probabilistic models of sensorimotor control and decision making
The effortless ease with which humans move our arms, our eyes, even our lips when we speak masks the true complexity of the control processes involved. This is evident when we try to build machines to perform human control tasks. While computers can now beat grandmasters at chess, no computer can yet control a robot to manipulate a chess piece with the dexterity of a six-year-old child. I will review our recent work on how the humans learn to make skilled movements covering probabilistic models of learning, including Bayesian and structural learning, as well as decision making and the revision of decisions in the face of uncertainty.
Israel Nelken 11.30 – 12.30
Dept. of Neurobiology and the ICNC, Hebrew University
The representation of surprise in the auditory system
Neurons in auditory cortex show high sensitivity to rare sounds, a phenomenon often called stimulus-specific adaptation (SSA). I will describe our attempts to find out what do the neurons really respond to, and to what extent SSA can be understood in terms of the simplest possible model, consisting of adaptation in narrow frequency channels. Finally, I will discuss some recent experiments in which we tested the sensitivity of neurons to features of the sound sequence that go beyond the rarity of the rare event, suggesting that neurons in auditory cortex are sensitive to higher-order regularities of the stimulus sequence.
Lunch and posters 12.30 – 14.30
John Hertz 14.30 – 15.30
Niels Bohr Institute, Copenhagen, and NORDITA, Stockholm
The Inverse Ising Model: Why and How
Ising models form a natural framework for modeling the distribution of multi-neuron spike patterns: Of all models that correctly describe the firing rates and pairwise firing correlations, the Ising model is the one of maximum entropy.
The problem at hand here is an inverse one to that we usually encounter. Normally, one has a model with given couplings (J ij) and the task is to compute averages and correlation functions of the variables of the model. Here we are given the averages and correlations and the task is to find the couplings.
In the simplest approach to this problem, one considers only the measured firing rates and equal-time pairwise firing correlations and tries to find the Ising model that has these statistics. In our work we have explored and compared a number of methods for doing this, using data from a realistic model network of spiking neurons. Several of these methods work remarkably well.
This success is tempered, however, by our second set of findings. Using an information-theoretic measure of the overall quality of fit, we find that, while the Ising model is a good description of the distribution of spike patterns for small populations of neurons (~ 10), it does worse and worse for larger and larger populations (for reasons that are not yet understood).
Finally, I will describe some recent work, which extends the Ising approach to describe non-equal-time firing correlations.
Yair Weiss 15.30 – 16.30
School of Computer Science and Engineering, The Hebrew University of Jerusalem
Learning and inference in low-level vision
Low level vision addresses the issues of labeling and organizing image pixels according to scene related properties - such as motion, contrast, depth and reﬂectance. I will describe our attempts to understand low-level vision in humans and machines as optimal inference given the statistics of the world. In particular, I will show how message passing algorithms allow us to solve real-world instances of NP-hard problems and to efficiently learn energy functions despite an exponential number of constraints.
TEA 16.30 -17.00
Marty Banks 17.00 – 18.00
Visual Space Perception Laboratory, UC Berkeley, USA
Perceptual Bases for Rules of Thumb in Photography
Photographers utilize many rules of thumb for creating natural-looking pictures. The explanations for these guidelines are vague and probably incorrect. I will explore two common photographic rules and argue that they are understandable from a consideration of the perceptual mechanisms involved and peoples’ viewing habits.
The first rule of thumb concerns the lens focal length required to produce pictures that are not spatially distorted. Photography textbooks recommend choosing a focal length that is ~3/2 the film width. The textbooks state vaguely that the rule creates ‘‘a field of view that corresponds to that of normal vision’’ (Giancoli, 2000), ‘‘the same perspective as the human eye’’ (Alesse, 1989), or “approximates the impression human vision gives”
(London et al., 2005). There are two phenomena related to this rule. One is perceived spatial distortions in wide-angle (short focal length) pictures. I will argue that the perceived distortions are caused by the perceptual mechanisms people employ to take into account oblique viewing positions. I will present some demonstrations that validate this explanation. The second phenomenon is perceived depth in pictures taken with different focal lengths. The textbooks argue that pictures taken with short focal lengths expand perceived depth and those taken with long focal lengths compress it. I will argue that these effects are due to a combination of the viewing geometry and the way people typically look at pictures. I will present demonstrations to validate this.
The second rule of thumb concerns the camera aperture and depth-of-field blur. Photography textbooks do not describe a quantitative rule and treat the magnitude of depth-of-field blur as arbitrary. I will examine the geometry of apertures, lenses, and image formation. From that analysis, I will argue that there is a natural relationship between depth-of-field blur and the 3D layout of the photographed scene. I will present demonstrations that human viewers are sensitive to this relationship. In particular, depicted scenes are perceived differently depending on the relationship between blur and 3D layout.