Vision Flashcards
What does the optical nerve do?
Carries information to the thalmus, in the very centre of the brain.
Where in the cortex is the visual cortex?
At the back of the cortex
How is information processed in the visual cortex?
The information fans out and is integrated with other signals, from memory and other sensory modalities and other aspects of our cognition.
What side of the brain do signals from the left sides off the retina in both eyes go to?
Right side of the brain
Where is the layer that detects the light in the eye?
At the back of the eye.
What detects light in the eye?
Photoreceptors, these don’t spike, but they convert light into electrical activity
What are the two types of photoreceptors?
Rods: These are important for vision in low light
Cones: Which are responsible for colour vision and important for vision in normal lighting
What do ganglion cells do?
Aggregate activity from a number of photoreceptors, along with the activity from some inhibitory cells in the intermediate layer and their axons from the optic nerve, carrying information to the thalmus.
What are receptive fields?
The receptive field of an individual sensory neuron is the particular region of the sensory space (e.g., the body surface, or the visual field) in which a stimulus will trigger the firing of that neuron.
What is V1?
Primary visual cortex
What cells make up V1?
Simple cell that have edge like receptive fields.
Complex cells
rbar=r0+sum(wij*Iij)
What are the variables?
rbar is the firing rate
r0 is the background firing
wij is the receptive field
Iij is the illumination point in the visual field.
We need to choose __ to minimise the average squared error.
wij the receptive field.
Feature selection firing rate equation?
a_s=sum(w_ij^s*Iij)
Where a_s is the firing rate,
w_ij is the receptive field
Iij is the illumination point
What does the following equation estimate:
Iij=sum(a_sW_ij^s)
Estimates the image using the firing rates.
Reconstructed image may not be equal to the original image and will just be an approximation to it.
Roughy, the _ neurons needed to reconstruct an image, the more of the image each neuron is coding for
Lower. For this to work without having a vast number of neurons covering every possible combination of pixels, the neurons must code for features.
Let I be a image and Ibar be an approximation to the image formed from features.
Ibar is an approximation because the a_s are not just chosen to give an accurate reconstruction, but to do so in a sparse way; theyaare chosen to minimize what equation?
E=sum(I_ij-Ibar_ij)^2+ßsumf(a_s)
First term being squared error between I and Ibar, the second is the intended measure of sparseness.
ß determines the trade off between accuracy and sparseness. If ß is small, the squared error dominates if ß is large the sparseness dominates.
What is the update rule for the features W_ij^S
Wij^s->Wij^S-eta(dE/dWij^s) where eta is the learning rate and E is the average error.