Friday, May 15, 2015, 9:00 – 11:30 am, Island Ballroom
Dave Knill was a beloved scientist, teacher, and VSS regular who suddenly passed away in 2014. Dave also served on the VSS Board of Directors from 2002 to 2007. Dave got his Ph.D. from Brown University in 1990 with a thesis about the perception of surface shape and reflectance. He did a postdoc at the University of Minnesota, after which he held faculty positions at the University of Pennsylvania and the University of Rochester, where he was since 1999. Dave left a towering legacy in many areas of vision science and decision-making, from Bayesian modeling to spatial vision to active sensing to multisensory perception to bounded rationality. In this symposium, a few of Dave’s many trainees and collaborators will commemorate his life and work.
Speakers: Dan Kersten, Paul Schrater, Robert Jacobs, Chris Sims, Krystel Huxlin, Wei Ji Ma
By the mid 1980s, computer science had helped to define vision problems to be solved, but had also shown how elusive their solutions could be. Neurophysiology was showing that primate visual processing involved significantly more cortex than had been thought. Marr’s book had just been published and understanding human vision was starting to look like a bigger and more interesting challenge. Around the same time, advances in digital signal processing were providing the means to create, filter and manipulate images. 3D computer graphics was making it possible to generate images from models of objects and scenes. Signal detection theory had been widely used for several decades, but most applications to studies of human vision had involved image patterns as the signals. This was the state of affairs when David Knill began his graduate work at Brown University. In my talk, I will describe the enormous role David played over the subsequent ten years in developing our understanding of objects and scenes as signals, images as their causal results, and from there, perception as Bayesian inference.
Cue integration formed a critical problem when I was Dave’s graduate student and formed one of Dave’s research foci throughout his career. I will describe Dave’s key contributions to the conceptualization of cues, Bayesian approaches and methods for elicitation. As we both moved to visuomotor control, we began to reconceptualize cue integration from a control perspective. I will trace that history and describe its deep influence on more recent work where we challenge the idea of cues as information towards privileged variables like object shape, size or location, and instead develop the idea that integrating perceptual information should subserve the goals of action. In effect, what you are doing determines what information is relevant, which variables should be estimated, and how perceptual input relates to the variables needed to make control decisions. I’ll review Dave’s innovative approaches to assessing cue integration, from slant from texture to visual signals to hand location. I’ll also describe a cue integration experiment where subjects successfully learned to integrate visual and auditory cues in non-standard ways in order to control an object. Throughout, Dave’s pioneering use of probabilistic modeling for conceptual development, stimulus design and data analysis will be highlighted.
I will start by describing research that was generated by Dave’s scientific creativity, rigor, and passion. This research, conducted by Joseph Atkins, Dave, and me, examined how inconsistent sensory signals in a multisensory (visual-haptic) environment can lead people to recalibrate how they combine depth information from multiple visual cues. I will then review subsequent research from my lab on crossmodal transfer of object shape knowledge across visual and haptic modalities, as well as work on transfer of knowledge from a perception task to a motor production task. Dave and I were both interested in sensory integration, multisensory perception, and possible relationships between perception and motor production. Talking with Dave about all of these topics was great fun and often insightful.
Dave Knill wasn’t content to observe or measure human behavior, he wanted to explain it. To Dave, an explanation for behavior and brain function almost always consisted of its elegant and parsimonious restatement as the solution to a computational problem. During the time I spent in his lab, I focused on two projects: Understanding the adaptive allocation of visual gaze in complex tasks (an offshoot of his work in visual-motor control with Jeff Saunders), and redefining visual working memory as the problem of minimizing behavioral costs under a capacity constraint (building on his work with Anne-Marie Brouwer). This latter project advanced information theory as a principled approach to defining the limits of visual working memory. In hindsight, both projects are really about bounded rationality—computational explorations of the idea that the brain can be highly limited, and yet simultaneously efficient. Dave was selfless as a mentor, and I am honored to have had the chance to grow as a scientist working in his laboratory.
Understanding this was the goal of the research Dave and I were pursuing with our graduate student, Laurel Issen. Dave passed away before this goal could be fully realized but it illustrates his openness and willingness to apply rigorous approaches to the study of clinical problems. I will start by explaining why our original question was (and still remains) of interest in the context of cortically blind people. My lab studies how visual training restores some of the vision lost in this patient population. It was of great interest for Dave and me to better understand why the cortically blind, who retain at least one intact hemifield of vision in both eyes, have such trouble navigating and orienting in their environment. Answering this question represents a first step in assessing whether restoring some of the vision lost is likely to impact visually-guided functions and ultimately, quality of life in this patient population.
Wei Ji Ma
Although Dave was one of the humblest people you would ever meet, the theoretical and empirical contributions he made to perception research are second to none. I will highlight mixture priors and causal inference, two intertwined computational concepts related to the inference of hidden causes. In a seminal 2003 paper on depth perception, he introduced the concept of mixture priors in vision. He later studied visuo-memory cue combination in a naturalistic reaching task with Anne-Marie Brouwer. In 2008, I worked on an extension of this study with him, in which we explored causal inference in the same reaching task. Very recently, Dave and Oh-Sang Kwon used the same notion of causal inference to explain how the brain resolves conflicts between local and global motion cues. Dave was a dear friend, an amazingly selfless and patient mentor, and one of the best scientists I have known.