II. Mechanistic Model of Receptive Fields
- The Question: How are receptive fields constructed using the neural circuitry of the
- How are these oriented receptive fields obtained from center-surround receptive fields?
(this is a mechanistic model)
II. Mechanistic Model of Receptive Fields: V1
- To answer this question, we need to look at the brain.
o Input flows from the retina to the LGN, where there are centre-surround
receptive fields to V1, where we have elongated, oriented receptive fields
o The clue to this comes from the anatomy: a single V1 cell receives inputs from a
large number of LGN cells. - Model suggested by Hubel & Wiesel in the 1960s: V1 RFs are created from converging
- Center-surround LGN RFs are displaced along preferred orientation of V1 cell
- This simple model is still controversial! It does not take into account the other inputs
that this V1 cell is receiving
- Each V1 cell receives inputs from several other V1 cells as well as the feed-forward
information from the LGN cells
- All you have to do is arrange the inputs from the LGN to have receptive fields that are
aligned in the same orientation as the V1 cell
This is an example of a "mechanistic" model. By "mechanistic" we mean:
A) The model describes what the neural system is doing.
B) The model describes how the neural system does what we have observed it to do.
C) The model describes why the neural system might do something the way it does.
D) None of these
III. Interpretive Model of Receptive Fields
- The Question: Why are receptive fields in V1 shaped in this way?
o Why do they have this orientation?
o Why are they selective for light or dark?
- What are the computational advantages of such receptive fields?
- Efficient Coding Hypothesis: Suppose the goal is to represent images
as faithfully and efficiently as possible using neurons with receptive
fields RF1, RF2, etc. which can be seen to the right
o We’re asking if these are the best/most efficient way of
representing receptive fields
- Given image I, we can reconstruct I using neural responses r , r ..1: 2
o Can we just add receptive fields together to make a picture?
Yes! We can linearly combine them using the equation
- Goal: What are the RF thai minimize the total squared pixelwise
errors between I and Î and are as independent as possible (increasing
the independence increases efficiency)? Q2
When we say "linear combination" we are talking about a specific mathematical way of
combining several things together. Which of the following looks most like an example of a