modelling the visual system Flashcards
(16 cards)
What are the roles of the retina, LGN, and V1 in visual processing?
The retina captures light and performs early spatial filtering using centre-surround receptive fields, such as those modelled by Difference-of-Gaussian (DoG) filters. The LGN (lateral geniculate nucleus) acts as more than a simple relay—it modulates signals with feedback from V1. V1 (primary visual cortex) processes orientation, position, and direction of stimuli and is the first cortical area where a spatially organised map of visual input is formed.
How do Difference-of-Gaussian (DoG) filters model receptive fields in the retina?
DoG filters approximate the centre-surround organisation of retinal ganglion cell receptive fields. These filters emphasise changes in light intensity (edges) and act as bandpass filters in frequency space. When applied across an image, DoG filters simulate the retinal encoding of spatial structure by enhancing informative edges while reducing redundancy.
Why is edge detection important in retinal encoding?
Natural images are sparse—many neighbouring pixels contain redundant information. Edge detection through centre-surround receptive fields reduces redundancy and highlights informative features, which supports efficient coding and optimises information transfer through the optic nerve, a bandwidth-limited channel (~10 million bits/sec).
What is the significance of neural diversity in retinal ganglion cells (RGCs)?
There are over 50 types of RGCs, each specialised for different features: some respond to colour, others to luminance, motion, or object size.
This diversity allows the retina to extract multiple parallel streams of visual information before transmission to the brain (Kim et al., 2021).
What is the role of the LGN in the visual pathway?
The LGN, part of the thalamus, receives input from retinal ganglion cells and organises this input into parallel streams
(e.g., magnocellular for ‘where’, parvocellular for ‘what’).
It sends feedforward projections to V1 and receives feedback, suggesting a dynamic role in modulating visual information, not just relaying it (Duncan & Boynton, 2003).
How are simple and complex cells in V1 modelled computationally?
Simple cells are modelled using Gabor filters—2D Gaussians modulated by sine or cosine functions. Simple cells respond if the edge is in a specific orientation
These models capture orientation selectivity and spatial frequency tuning
Complex cells are thought to pool outputs from multiple simple cells with different phases, removing phase sensitivity and enabling invariant feature detection (Hubel & Wiesel, 1968).
What are cortical maps in V1 and how are they studied?
Cortical maps in V1 reflect orderly spatial organisation of neurons based on visual features (e.g., orientation, direction).
These maps are organised according to a coordinate system transformed non-linearly. This non-linear transformation means that the centre of the field has a larger representation
What insights do models provide about V1 dynamics during perception?
Models such as those by Rankin & Chavane (2017) show that lateral selectivity and excitation width in V1 determine how activity spreads across orientation maps. Such models help explain how contextual information modulates perception through lateral interactions in V1.
What are on-centre and off-centre retinal ganglion cells, and how do they respond to light?
These are two types of retinal ganglion cells with centre-surround receptive fields that detect contrast rather than absolute brightness.
On-centre cells:
Excited when light hits the centre of their receptive field.
Inhibited when light hits the surround.
Best activated by a bright spot on a dark background.
Off-centre cells:
Excited when light hits the surround.
Inhibited when light hits the centre.
Best activated by a dark spot on a bright background.
Together, they enhance edge detection and contrast, enabling efficient visual encoding from the retina to the brain.
What is the receptive field model equation?
Write out
It is telling us the response that we would get for a receptive field like R that is centered on I image
So I = the 2d image, with positions X and Y. Each value represents the black-white colour value for that part of the image.
The reason it has two integrals is because the receptive field has a range. So you are integrating across horizontal space (x) and vertical space (y)
You can also include a non-linear sigmoidal response given by f(X,y), the result is a poisson process, or the random spiking of a single neuron
The first equation produces a number that gets larger or smaller depending on how big the response is locally. However adding the poisson distribution adds some variability . This makes it a non-deterministic/ probabalistic model
How to go from one neuron receptive field to the response of many neurons?
Add in CxCy which represents each neuron firing in response to the part of the image
So Rcx,cy, refers to the receptive field being centered (for each neuron) on x,y
Now the DoG acts as a spatial filter, and the equation shows an image convolution.
Output: there is an output for each pixel, which represents the output of the spatial filter centred on that pixel
What happens if you plot the sensitivity of the receptive field in frequency space?
- You see that there is a combined effect of a band pass filter at the peak frequency
- This means that it will allow a band of frequencies to pass through, but not frequencies that are too high or two low#
- SO what the filter is doing:
You take the image, you then pass the resulting neuron firing to the bandpass, then all you would see is the edges that are matched to that frequency!
Why do we need an edge filter?
For efficient coding of information
Modedlling simple cells as gabor functions and filters
Write out equation
To generate computed responses as models of receptive fields we use gabor functions as the filter kernel
The Gabor is a 2D gaussian multiplied by a periodic cosine function
In this equation, we use two Gabor functions and take the difference between them as the receptive field
The periodic cosine is multiplied by the preffered orientation and spatial frequency, and the preffered phase
The cosine bit means that there is a peak followed by a trough
How can you create the model of simple cells from complex cells
Write out
If you put simple cells together and then remove the phase dependency by subtracting pi/2 then you get an output that resembles complex cells
Squaring the term for each simple cell makes the complex cell nonlinear
