Fichas
Self-organizing maps and neural coding
Fichas del conjunto (24)
- 1What was (Perge et al, 2009)'s equation for maximal firing rate of a myelinated neurons with axonal diameter d?y=10(d-0.46) spikes/s
- 2If cross sectional area of a neuron is A, what is the new version of Purge et al's equation?y=k(2(A/pi)^0.5-d_min) spikes/s
- 3What efficiency does coding capacity (R_info/R_max) have?0.5. Half of coding capacity is mutual information
- 4What is R_max?total entropy (information) in
the neuron’s output based on its firing rate - 5What is R_info?The estimated amount of information in the neuron’s output that refers to its input.
This is based on a linear model of the neuron and is always less than (or equal to) the actual mutual information I(x, y). - 6What did Niven et al (2003) say about R_info?Data were fitted to a straight line on log-log axes, to obtain:
log energy ∝ 1.7 log R_info,
log energy = 1.7 log R_info + 4.68,
R_info ∝ energy^(1/1.7),
R_info ∝ sqrt(energy).
(diminishing returns) - 7What is the ditance between the pdfs p_s(s) and g'(s)?Kullback-Leibler distance.
It is a measure of the difference between the pdf p_s(s) of input s (fixed), and pdf g'(s), which can be varied modifying the form of g'(s).
H(y)=int_s (p_s(s) log(g'(s)/p_s(s)) ds - 8What is the formula for output of a color cone, where i belongs to R, G?x_i = n_i +
int_lambda( L(lambda) f_i(lambda) dlambda
where
f is tuning curve
and
L is spectrum of light impinging on cone - 9What is the formula for correlation?corr(x_r,x_g)=
C/n sum( (x_R - av(x_R)) (x_G - av(x_G)) ) - 10What two parts, that can function in principle, independently, can an agent be constructed of?- An actor responsible for implementing the policy
- A critic responsible for evaluating the consequences of rewards - 11The set of templates, w, is a good encoding of data, x, if what?-maximises the data compression:
n(w)/n(x)
-minimises the reconstruction error:
||w_i - w_match(x_i) ||=
sqrt (sum ( x_i - w_match(x_i) ) ) - 12What is the general Kohonen SOM algorithm?1) generate new input vector
2) compute responses
3) locate 'winning' neuron
4) define winning neighbourhood
5) Apply hebbian update6) repeat (optional: shrink learning rate neighbourhood) - 13Describe a fishnet plot.-Each point represents the weight vector of a single unit
-weight vectors are plotted as coordinates on an input space
lines connect weight vectors of adjacent units in the network - 14What functions would biological hardware have to perform in order to implement SOM algorithm?- compute responses to input
- locate winning neuron (requires global supervision)
- update neighbourhood weights - 15What real map developments exist, which LISSOM does not display?-robust, to environmental manipulation
-stable, map structure changes minimally
-adaptive, reflects input statistics - 16What is contrast gain control?Lateral inhibitory interactions between LGN neurons make input to cortex from LGN a non-linear function of stimulus strength. This improve stability.
- 17What is homeostatic plasticity?Each cortical neuron changes the distribution of ion channels to regulate the amount of synaptic input required to elicit a spike. In the model, the output function threshold adapts to main constant activation.
- 18What does GCAL mean?Gain - Control, Adaption, Laterally connected model
- 19Where is there still work to be done on the GCAL model?Modelling inhibitory and excitatory populations explicitly.
- 20What's the difference between random maps and topological maps with regard to minimal wiring lengths?Whilst they both minimise wiring lengths connecting like tuned features,
-random maps maximise local coverage
&
-topological maps maximise local continuity - 21What is mutual information?The average amount of Shannon information (i.e. entropy) that each output value provides about the input.
- 22What is coding capacity?The total entropy (information) in the neuron’s output based on its firing rate. Some of this information is noise and therefore is unrelated to the neuron’s input.
- 23What is R_info?The estimated amount of information in the neuron’s output that refers to its input. This is based on a linear model of the neuron and is always less than (or equal to) the actual mutual information, I(x, y).
- 24Who came up with the original SOM?Von der Marlsburg (1973)
Información
- Descripción
- urmom computational ns Fichas sobre Sin título, creado por samsden el 28/05/2016.
- Creado por
- samsden
- Visitas
- 1
- Valoración
Más sobre Computational Neuroscience and Neuron Models