Dayan & Abbott Week 3 (Chapter 2 - Section 2.3)
- Many different ways of capturing the informational content of a neuron based on its spike train. Each one throws away some possibly important information.
- Spike-triggered average stimulus captures the selectivity of a neuron to dynamic stimuli.
- Retina, LGN, V1 can study responses of neurons. Generally V2 too complex and non-linear. If a lot of V1 behaviour is dependent on feedback from V2 then surely it would be just as complicated?
Estimating Firing Rates
- Reverse correlation methods used to construct a model that includes the effects of the stimulus over extended period of time.
- Assume firing rate at $t$ is a weighted sum of the values of the stimulus at earlier times.
- Are we constrained to only stimulus that we can quantify? Is the brain also constrained in this way?
The Early Visual System
- Figure 2.4: Rod and cone receptors not spiking.
- Smoother representation of light intensity, sometimes inverted.
- Why is this smoother representation only suitable in the retina? (Distances are small)
- Retinal Ganglion spiking behaviour.
- From Figure still, G2 fires when light is on, G1 fires when light is off.
- Optic Chiasm where optic nerves cross. But some do not cross. The net result is for the right cerebral hemisphere to sense and process left hemispheric vision, and vice versa.
- Neurons have receptive fields in retina + LGN and primary visual cortex.
- Illumination outside receptive field can not generate a response directly, although they can affect responses to stimuli (Is this the PP extended line example?)
- Retinal Ganglion + LGN respond best to circular spots of light surrounded by darkness or dark spots surrounded by light.
- Primary visual cortex many respond best to bars or boundaries. (Gabor patches)
- Static images are not very effective at evoking visual responses. (Are we just going to read PP into everything now?) This is also at retina/LGN level?
The Retinotopic Map
- Retinotopic map is the transformation from visual topology to corresponding locations on the cortical surface.
- Lots of interesting Figures at the bottom of this page.
- The visual world is mapped onto the cortical surface in a topographic manner. This feels like common-sense. Should it be surprising (striking)?
- Eccentricity = distance from fixation point. Azimuth = angle from horizontal.
- Pixel-by-pixel light intensities are not useful parametrisation of a visual image because neurons adapt to the overall level of illumination.
- Instead use difference between two points divided by the background level.
The Nyquist Frequency
- A visual example of temporal aliasing.
- In the retina aliasing is caused by density of photoreceptors and is spatial not temporal.
Misc Questions (not from the book)
- What do glial cells do?
- What happens when a neuron abandons a connection? Does that part of the axon form a new one somewhere?
- During refractory time of a neuron a 3GHz processor can perform more than ten million operations. Not close to what 80 billion neurons can do in parallel, but still interesting.