9 - Neural networks models for faster decision making Flashcards
What is the receptive field?
the range of sensory space to which a neuron responds
eg. receptive field area of whole visual field
What is the tuning curve?
the relationship between a stimulus feature and a neuronβs response (eg firing rate)
What do P and N stand for in the tuning curve of the random dot-motion discrimination task?
Where does it peak at the tuning curve?
-preferred and null direction of the dots (in monkey receptive field)
-peaks at P (at 180 degrees)
In random dot-motion discrimination task, what is correlation/coherence?
refers to the ratio of dots moving in the same direction
As the correlation/coherence increases, what happens to the number of correct saccades that the monkeys do?
increases
What is the random dot-motion discrimination task?
-Monkey has to make saccade based on visual stimulus of moving dot (with varying levels of coherence)
-They measure how correct the saccade was
How are the lateral interparietal cortex LIP receptive fields involved in perceptual decision making?
RFs predict direction of saccades eg.
When is monkey is shown stimulus moving to the right -> LIP receptive field predict saccade movement to right
In LIP random dot-motion tests why does it take longer for monkey to make saccades when coherence is lower?
decreased coherence means that time between stimulus and saccade is increased -> as monkey is taking longer to decide whether dot is going L or R
(monkey is accumulating evidence)
In LIP random dot-motion tests what happens when the stimulus dot is moving the opposite direction of the RF?
LIP RF FIRING RATE is after saccade is made (inhibition is seen)
In LIP random dot-motion tests, what happens when the stimulus dot is moving in opposite direction and coherence is reduced?
What is the relationship between dot stimulus in opposite direction to RF (T2) and in the direction of (T1)?
-the inhibition seen because its the opposite direction is now augmented because of the low coherence
-neurons selective for T2 seem to inhibit those selective for the opposite direction T1 ???
What are the requirements of homogenous neural network model? (alekhyas)
-All neurons (N) are identical
-Required same external input current (Iext)
-Interaction strength (wij) is uniform
If
Wππ>0: exc->exc
what is?
Wππ>0
Wππ<0
Wππ>0: exc->inh
Wππ<0: inh->exc
In Winner Takes All Dynamics (WTAD), which population is inhibitory?
bottom one (top two are excitatory
In WTAD with one inhibitory and two excitatory population, what happens when you add a strong input to the left? who wins the competition?
How does this relate to decision making in LIP and the money test?
Strong input to left -> increased activity in left -> increased activity in the inhibitory population -> increased inhibition to left and right but overall excitatory drive to left is larger -> left population wins the competition
left population has more excitatory activity than right -> thus wins competition as R is inhibited -> saccade is made to the left
In WTAD, when Ileft=Iright , who wins the competion?
either population
How are attractor states related to decisions?
reaching an attractor state = making decision
What model does the WTAD use and threshold concept?
-Leaky
Integrate and Fire Model with a logistic sigmoidal activation function
In the leaky integrate and fire WTAD model, what is x and tau?
x = population activity of neural networks
tau = time parameter
How many eqn.s for WTAD dynamics?
How many dimensions?
-3 = one for each population
-3
What are the two assumptions of the WTAD model? what does this cause mathematically
- no interactions between itself in the inhibitory network thus ->
inhibitory neurons operate near the linear part of the sigmoid π₯ππββπ(π₯ππβ ) - Inhibitory neurons are much faster than excitatory neurons πππ₯πβ«πππββ0
What is the mathematical concepts used to simplify the three dimensional WTAD model
-Separation of time scales: if some dynamic variables are much FASTER than others, replace them with their steady state.
-Linearisation: Under certain conditions, approximate nonlinear terms with linear terms.
In HH model, is the activation particle faster or slower than inactivation?
What model does this concept above inspire?
What is this mathematical concept called?
-activation is faster than inactivation (thus activation is replaced by a constant)
-thus we form Vβ and Wβ in the FN model
-separation of time scales
How is linearisation used in FN and in single rate neurons?
Wβ and linear activation function respectively
To reduce dimensionality of the 3D of WTAD model, what do you remove?
In the 2D WTAD model, what are the weights to eachother and to itself?
-the inhibitory population
- eachother: minus alpha
itself Wee-alpha
What is lateral inhibition represented by in the 2D WTAD model?
minus alpha
What mathematical concepts do you need to simplify the 3D WTAD model to 2D?
separation of time scales and linearisation
Plotting null clines of WTAD models, allows you to see what?
can be used to see which side wins
If you have a strong right current, which side wins?
If you have strong symmetric input, what are the fixed points like? Who wins competition?
-right
-three fixed points: 1x unstable and two stable . Sometimes L sometimes R
