Transcript slides
What is the Other 85% of V1 Doing? Olshausen & Field Problems in Systems Neuroscience, 2004 Brian Potetz 1/26/05 http://www.cnbc.cmu.edu/cns Only 15% Understood 1. 2. 3. 4. 5. Biased sampling of neurons Biased experimental stimulus Biased theories (towards simplicity) Neural interdependence & contextual effects Ecological deviance 15% ¼ 35% x 40% Problem 1: Biased Neural Sampling • Large cell bodies, stronger action potentials – Miss 5-10% of cells • Discarded “visually unresponsive” cells – Spontaneous, bursting, or tonic cells – Miss 5-10% of cells • Bias towards high firing rates – Miss 50-60% of cells • Generous estimate: 40% of cells are unsampled. Sustaining High Firing Rates is Difficult Energy consumption of: Cortex Whole brain Whole body • Yet even for natural images, 9Hz recorded averages are typical Problem 1: Biased Neural Sampling Lennie, “The Cost of Cortical Computation” (2003) Spike Counts of Single Neurons Follow Exponential Distribution for Natural Scenes Single Neurons (anaesthetised cat V1, ave firing rate = 4Hz) Population Average: Problem 1: Biased Neural Sampling Baddeley et al, “Responses of neurons in primary and inferior temporal visual cortices to natural scenes” (1997) So Neurons with 1Hz average rates may go unnoticed Problem 1: Biased Neural Sampling Estimating Percent of Unsampled Cells • • • Assume log-normal distribution of average firing rates (why?) Assume average average firing rate is 1Hz Assume measuring threshold is 1Hz (measured population mean) Lesson of the Rat Hippocampus • Via single electrode recording, granule dentate gyrus cells thought to be mostly high-rate (theta-cells) interneurons. • Chronic implants reveal that most are very low-rate: 0.1Hz is common. • What are non-geniculate granule cells of layer 4 doing (30:1)? Problem 5: Ecological Deviance Anaesthetized cat response to natural movie, vs linear model prediction (Gray lab) Another group (David, Vinje & Gallant), with sophisticated nonlinear models learned over natural stimulus, can capture only 20% of neural response variance, even correcting for inter-trial variance (using awake monkeys). Problem 2: Biased Stimuli • • • • If neurons are so highly nonlinear in the natural environment, why focus on linear measurement techniques? “There is no principled reason for using sinewaves to study vision” Authors suggest an alternating approach of natural stimuli, then attempt to reduce stimuli to “tease apart” phenomena. Authors advocate studying V1 further, rather than assuming current models are correct. Problem 3: Biased Theories • • • Working theories for subsets of data are more easily published (and remembered) than messy, unexplained data. Possible example: V1 as edge-detector Possible example: simple & complex cells Problem 4: Interdependence and Contextual Effects • By suppressing intracortical signals using electric simulation, LGN input was estimated causing 35% of simple cell response variance. (Chung & Ferster, 98) • By recording optical imaging, local field potential, and single cell response simultaneously, 80% of V1 cell response was attributed to ongoing population activity. (Arieli et al, 96) Extra-classical receptive fields The effect of oriented bars outside of the classical RF is likely to be only one example of many contextual effects. Problem 4: Interdependence and Contextual Effects Synchrony • The fact that EEG and local field potentials are measurable suggests that synchrony takes place in the cortex • Worgotter (1996) showed that LGN and V1 receptive fields widen as synchrony increases (as measured by EEG). Problem 4: Interdependence and Contextual Effects Alternative Theories 1. Limitations of prediction for dynamic systems 2. Sparse, overcomplete representations 3. Contour integration 4. 3D Surface representation 5. Top-down feedback, Bayesian inference 6. Dynamic routing Conclusions • • • • Natural stimulus Multi-unit recordings Chronic implants Public database of recordings for others to model