Seminar: University Seminary on Cognitive and Behavioral Neuroscience (#603)
Title: Coding of eye movements by neural activity
Speaker: F.A. Miles, Laboratory of Sensorimotor Research, National Eye Institute
Attendees: Herb Terrace, Co-Chair, Psychology Department, Columbia University
Jacqui Rick, Psychology Department, Columbia University
Yaakov Stern, Co-Chair, Taub Institute, Columbia University
Michael Goldberg, Neurology Department, Columbia University
Len Matin, Psychology Department, Columbia University
Ning Qian, Neurobiology & Behavior Center, Columbia University
Aniruddha Das, Neurobiology & Behavior Center, Columbia University
Daniel Salzman, Neurobiology & Behavior Center, Columbia University
Wenxun Lee, Psychology Department, Columbia University
Jon Horvitz, Psychology Department, Columbia University
Tammy Moscrip, Psychology Department, Columbia University
Roee Holtzer, Sergievsky Center, Columbia University
Chris Habeck, Taub Institute, Columbia University
Josh Wallman, Biology Department, City College
Jay Edelman, Biology Department, City College
Robert L. Thompson, Psychology Department, Hunter College
Rapporteur: Michael R. Drew
Summary:
Dr. Miles talked about coding of short-latency vergence eye movements in the medial superior temporal (MST) cortex of Japanese macaques. The short-latency eye movements are elicited by binocular disparity, and their purpose is to keep objects in the plane of regard from being seen as double images. The vergence response brings the eyes to focus on the object producing binocular disparity. Because each eye has a slightly different view, when an object is before or beyond the fixation point, the image of that object falls on different locations on the two retinas. This difference is known as retinal (or binocular) disparity.
In Dr. Miles’ experiments, the visual stimuli were large random-dot patterns, which were presented to the two eyes using a dichoptic viewing paradigm. In dichoptic viewing, images are presented to the two eyes independently. In this case the independence was achieved by using two projectors, one for each eye. Miles could create and control retinal disparity by showing the two eyes images that are slightly displaced relative to each other. The experimental procedure was to induce retinal disparity of a given amount, then measure the resultant vergence angle and vergence speed (deg/s). There was no feedback and no reinforcement. The short-latency vergence movements are an open loop phenomenon in that they are induced by the disparity stimulus and are not modulated by other external stimuli. Miles found that vergence movements occur within about 80 ms after the onset of disparity in humans and that in monkeys the latency is slightly shorter. Miles presented graphs of vergence angle versus retinal disparity. The graphs illustrated that the vergence system responds to only a narrow range of disparities. Vergence increases with increases in disparity out to a limit, beyond which the vergence response decreases with increases in disparity. After the vergence angle returns to zero, further increases in disparity have no effect on vergence. These vergence v. disparity functions were attained for two types of stimuli: correlated and antiocorrelated. When both eyes were shown black dot stimuli, the stimuli were correlated. Anticorrelation consisted of showing one eye black dots and the other eye white dots on a black background. Only the correlated stimuli produced the experience of depth, but both types of stimuli produced disparity-induced vergence. The vergence movements were in opposite directions for the correlated and anticorrelated stimuli.
It was previously shown that lesions to MST eliminate short-latency vergence responses. Intrigued by this result, Miles recorded from MST neurons in collaboration with a group from Japan. They found that about 20% of MST neurons modulate their activity in response to disparity changes. The neural activity tends to precede vergence by about 10ms, so in analyses relating unit activity to vergence, the unit data are shifted by 10ms. Miles plotted tuning curves for individual units. The curves related unit activity (spikes/s) to disparity (deg). A fuzzy logic algorithm classified the MST units into 4 categories, based on these tuning curves. The categories have been given the following names: tuned far, tuned inhibitory, near, and tuned near. Miles pointed out that the groupings are not actually categorical; they instead seem to represent different points along a continuum. The groupings accounted for much more variance in the data for correlated stimuli than for anticorrelated stimuli.
To explore the behavioral correlates of the neural activity, Miles fit the tuning curves for individual units to the vergence tuning curves. (The vergence tuning curves related vergence [deg/s] to disparity [deg]). Individual neural tuning curves coded the vergence response well for either correlated or anticorrelated stimuli, but never for both. This suggests that individual cells are not coding vergence. They instead appear to be discharging in relation to some aspect of the sensory input and or the motor response. To explore whether the population of cells encoded the vergence response, Miles fit the tuning curve for the summed neural activity to the tuning curve for vergence. Jon Horvitz asked whether the summed activity curve was obtained by simple summation, without any other operations. Dr. Miles replied that the only additional operation was that for one category of units the sign was changed. He added that this was justifiable because neural networks seem to handle sign changes easily. The population neural tuning curves fit the vergence tuning curves very well, accounting for at least 93% of the variance in the vergence curves. This was true for both the correlated and anticorrelated stimuli. This indicates that vergence is coded by the population activity. Moreover, the summed activity reproduced the idiosyncratic differences between the vergence responses of the two subjects. That is, the population tuning curve for one monkey did not fit the vergence curve for the other monkey. This means that the within-subject fits were not an artifact of the operations performed on the data.
Dr. Miles then asked, were some categories of neurons more important than others in accounting for the vergence response? Miles used two strategies to answer this question. The first strategy was to remove cells from the population, then examine whether the resulting population fit was the better or worse for it. Removing any one category of cell made the population fits worse. Removing a randomly selected subset of cells (drawn from all 4 catergories), however, had little effect on the fits. This suggests that the categories are behaviorally relevant. The second strategy was to try and obtain the subset of cells whose summed activity produces the best possible fit to the vergence data, and then to ask whether the 4 cell categories are represented equally in that subset. Dr. Miles pointed out that it is impossible to examine all possible subsets (there are 2^49 of them!). As an alternative, Miles implemented a genetic algorithm designed to give an estimate of the best subset without examining all possible subsets. The algorithm treats each cell as a gene. Chromosomes are formed by selecting a random subset of genes, and designating those genes “on” (+). The remaining genes are designated “off” (-). 5000 chromosomes were created, and then the chromosomes were mutated and mated according to the following rules:
--5% of chromosomes are randomly selected and passed on to the next generation unchanged.
--5% of chromosomes are randomly selected, mutated, and passed on to the next generation. Mutation consisted of changing the sign (+-) on 3 randomly selected genes.
--the best chromosome is passed on to the next generation unchanged
--the remaining 90% are each mated with the best out of 10 randomly selected chromosomes. Mating consisted of randomly selecting genes from each chromosome.
The algorithm was run for 50 generations, and then the entire “evolution” process was repeated several times over from scratch. The fit between the population neural tuning curves (created by summing across all active genes in the pool) and the vergence curves improved over generations. The best population curves accounted for well over 99% of the variance in the vergence curves (r squared > 0.9995). These optimal curves included cells from all four cell categories, providing more evidence that the cell categories are meaningful and that each category of cell contributes to the population code for vergence.
Dr. Miles concluded by summarizing his talk as follows:
1. 20% of cells in MST are selective for binocular disparity.
2. A fuzzy algorithm groups these cells into 4 categories.
3. The summed activity of all MST cells closely matched the tuning curves for vergence, even reproducing the indiosyncratic differences in vergence between monkeys.
4. Excluding any of the 4 cell groups produced a significant decrease in goodness-of-fit, indicating that the population code relies on contributions from all four categories of cells.
5. A genetic algorithm revealed subsets of cells that gave almost a perfect fit to the vergence curves. These subsets always included cells from all 4 categories.
6. The population activity encodes the magnitude and direction of vergence.
7. MST cells likely play a causal role in the vergence response.
Discussion:
Dr. Goldberg: Let’s get back to the idea that for this particular example, complicated wiring is not necessary. Is that because the effector neurons are basically similar. Dr. Miles: These cells in the MST are not uniquely driven by the disparity stimulus. We already know they respond to other stimuli. We know that disparity-induced vergence is not the only thing that these cells do. Our current hypothesis is that this population of cells is united in their projections onto the vergence motor system. But some of those same cells contribute to other behaviors. They may contribute to other kinds of percepts. At the single cell level, a given cell can be contributing to several kinds of behaviors. We have the advantage that in this system we can categorize the motor responses in to orthogonal components: there’s vergence and version. These are orthogonal and complete descriptions of eye movements. So we can examine the neural responses to orthogonal motor responses. Our hypothesis is that if you record from MST in relation to the paradigm we just looked at, the population activity will relate to the vergence eye movement. But if you look at activity in relation to the orthogonal movement, the population will be flat. There will be no net activity in that population. This is because that movement is coded by another population of which some of the MST cells may be members. So at the single-cell level there is multiplexing and what determines the contributions of the cell are where it projects. And this means that a cell can make a contribution to several different behaviors or percepts. So this is very different from the grandmother cell type approach. The question we are asking is what contribution can this cell make to this behavior. And in asking this we are not saying that the cell is not contributing to anything else.
Jon Horvitz: Only about 20% in MST are disparity selective. Are the 4 classes taken from that 20%? Miles: Yes. We’re only talking about that 20%.
Horvitz: So when you say that these neurons are responding to other things as well, do you know that individual neurons are responding to multiple things, or is it that different neurons in the same vicinity are responding to different things?
Miles: We know that cells in MST respond not just to disparity. If we could hold onto individual neurons for a long enough time, we would look at responses to different stimuli and I think we would find that individual neurons are responsive.
Horvitz: So is the idea that one cell is projecting to multiple areas?
Miles: Yes.
Hovitz: And so your model would not work if the cells were only projecting to one area.
Miles: Correct. MST projects all over the place. At this point we do not know all the anatomical connections.
Aniruddha Das: Is there any functional architecture to the categorization that the cells fell into?
Miles: Unfortunately my answer is very disappointing. In order to get the greatest yield of cells, we sacrificed anatomical information. So we don’t know what kinds of cells (pyramidal, interneuron etc) we’re recording from.
Yaakov Stern: So to reiterate, if you look at these cells in response to an orthogonal behavior, these cells will not respond?
Miles: Yes, that’s our idea. The responses will cancel each other out, meaning no net activity at the population level.
Stern: Could these cells make different responses that would sum up differently in a different paradigm?
Miles: It’s possible.
Jay Edelman: Could you take half the data set and run the genetic algorithm, then test how the population curves fit the other half of the data?
Miles: Unfortunately no. The data set is too small. Obtaining these curves takes a lot of averaging, and so if we divided in the data set in half there would just not be enough to get a tuning curve.
Prepared by Michael Drew, November 25, 2002