Personal tools
You are here: Home Events ANC Workshop Talk: Charles Sutton and Christopher Ball, Chair: Ali Eslami

ANC Workshop Talk: Charles Sutton and Christopher Ball, Chair: Ali Eslami

— filed under:

What
  • ANC Workshop Talk
When Jan 17, 2012
from 11:00 AM to 12:00 PM
Where IF 4.31/4.33
Add event to calendar vCal
iCal

Charles Sutton

Machine Learning: We Do, and They Don't

Despite the important strides that have been made by machine learning research, we are still far from the level of sophistication required for a self-sufficient learning system. I will describe what I believe are some major current roadblocks, including handling of streaming data, transfer learning, and our pernicious inability to divide and conquer. In addition, there are some open problems that seem important but that I suspect are currently too hard to work on fruitfully. If I feel reckless enough, I will list those, too.(This is a first draft of a future lecture in the Hamming Seminar series.)

 

Christopher Ball

Modeling the development of color maps in primary visual cortex

Optical and 2-photon calcium imaging studies of macaque monkey primary visual cortex indicate that color-selective cells are organized according to important features of subjective color perception, rather than according to physical wavelength. Furthermore, in contrast to orientation-selective cells, which form spatially contiguous maps, color-selective cells are found in small, spatially separated blobs. We are building a developmental model of the early visual system, trained on calibrated colour images, to try to show how these two aspects of color organization could arise.  The model robustly demonstrates organization of color blobs, with color-selective cells separated from orientation-selective cells for a wide range of parameter values. However, the range of hues represented and their spatial organization do not match results from the macaque monkey, when the architecture and parameters of the model are set to match known values from the macaque. In this talk, I will present work in progress showing how we are attempting to understand why a realistic organization of hues does not develop. In particular, I will show how varying certain dimensions of the model (cone receptor sensitivities, architecture of the retinal ganglion cell/LGN pathway, and hue statistics of the natural image input) away from their biologically realistic values can result in realistic hue organization. We hope that by making the effects of these individual dimensions clear, we will be able to show that various forms of adaptation, when applied to these dimensions (e.g. homeostatic adaptation of retinal ganglion cells), will allow the full range of hue selectivities and realistic hue organization to develop in the model.