Complexity Flashcards
Template neuron
Hypothetical neuron that responds to a particular template of an image.
Computational Approach
How vision works in THEORY and in PRACTICE.
Discretising template neurons…
Enables neurons to have non-overlapping narrow tuning curves. Reduces amount of neurons needed for a continuous range of parameter values.
Univariate neurons…
Respond to only one parameter.
Value unit
A univariate neuron that responds to only a small range of values.
Variable units
Univariate neurons that respond to all values of a parmeter.
Parameter space
The n-D space of all possible parameter values and combinations.
Multivariate neurons…
Neurons that respond to multiple parameters.
Exponential increase…
M=N^k
The number of template neurons (M) needed increases exponentially when the number of parameters (k) increases.
k IS the power to which N is increased, unlike polynomial, where k is raised to a fixed power (Sq).
If M=10 to the power of 10…
Then we would need 10 billion multivariate value units (neurons) to recognise a single object!
We can recognise hundreds of thousands of objects… Hence why vision is an exponential problem.
Grandmother neuron would be invariant because…
It would fire, irrespective of parameters, if the template neuron contained an image of its committed object (grandmother).
Local representation…
When the representation of an object is localised to a single cell.
Distributed representation…
When the representation of an object is distributed to a population of neurons with overlapping tuning curves (coarse coding).
Linear increase…
M=Nk N is fixed, but is multiplied by k. E.g if N=10, and k=4, then M=10x4=40 If k was doubled, then, in a linear fashion, so would M.
Polynomial increase…
M=NkSq
In this function, the increase is in proportion to the square of the number of k parameters.
k is raised to a fixed power.
e.g M=10(4Sq)=160
