L8 computational neusci Flashcards
Does all scientific activity revolves around the creation, testing, refinement and debate of scientific models?
Yes
What is Newtonian Mechanics?
Able to use a small number of very simple mathematical equations to describe the movement of planets with remarkable accuracy
What is parsimony?
Decide what is important and ignore the rest
What is explanatory?
Give good descriptions
What is predictive power and falsifiability
Allow precise predictions
Models of the brain have to strike the right balance between ______/______ and ______ of the explanations or predictions they make
simplicity; parsimony; accuracy
What do compartmental models do?
stimulate individual neurons as lots (sometimes thousands) of connected tubelets and each has its equivalent circuit
Purkinje cells: Number of compartments Number of ion channels Number of different voltage-dependent channels distributed across Number of synapses What does it do?
4550;8021; 10; 3500; reproduces electrical current injection and synaptic responses
Prof Markram‘s Blue Brain project tries to model parts of the neocortex by running vast _________ on massive super computers
compartmental models
What did the Blue Brain project ignore?
make the model parsimonious by leaving out as much irrelevant detail as possible
Why there is no method for trying to decide what details are relevant and which are not for Blue Brain project?
It is unclear what aspects of brain function the blue brain is suppose to model
Did Blue Brain yield great return?
No
Which of the following are spiking models/ non-spiking models (rate based)?:
Integrate and Fire
Linear threshold or linear-non linear neurons
linear neurons
the Izhikevich neurons
spiking model:
Integrate and Fire
the Izhikevich neurons
non-spiking models (rate based):
Linear threshold or linear-non linear neurons
linear neurons
It is useful to think of neurons as ______
in a first order approximation
How to compute neurons?
take all synaptic inputs (many of which may be zero most of the time) -> multiply each input by its ‘synaptic weight’ -> sum -> use a threshold criterion/ a linear/ sigmoidal function to decide how much output to send given the weighted sum of inputs
What is the major property of biological neurons?
firing action potentials in response to synaptic input
Essential features in simplified models of spiking neurons.
Action potential’s shape is same each time, so just treat the occurence of an AP as a binary 1
All EPSPs and IPSPs eventually get summed in the cell body, so no need to model dendrites
Integrate-and-fire model 3 properties.
Point neuron; resistance+capacitance; synaptic input
Which is simpler? Rate-based/ spiking neurons?
Rate-based
Rate-based neuron models’ inputs are ______ & ______ to calculate the ‘activation’
weighted; summed
Output in rate-based models could be _____/ _____/ ______/ ________
sigmoid; threshold; threshold linear function; linear function
Computing units used in artificial neural network can be regarded as models of _________
biological neurons
Activation function captures essence of a neuron’s spiking response using ____ & _____
Binary (0/1): neuron is silent/ firing - McCulloch-Pitts threshold logic unit
Analog: number represents neuron firing rate -piecewise linear/ sigmoid
McCullough-Pitts Neurons is also known as.
Binary Threshold Neurons
What is the simplest neuron model?
Binary threshold neuron
What do binary threshold neuron model receive?
Inputs form synapses that have positive/ negative weights (excitatory/ inhibitory synapses)
What is the activation of binary threshold neuron model?
Weighted sum of inputs
How does binary threshold neuron produce a binary output of 0/1?
Depends on whether the activation is greater than a given threshold
How is binary threshold neuron similar to biological neurons?
Summate synaptic currents in their dendrites and soma, and fire an action potential or not depending on whether there is depolarization
Binary threshold neurons are “_______”
Logic Gates
How to make it possible to make networks compute any function?
By setting synaptic weights on networks of binary threshold neurons appropriately
Linear and linear-nonlinear neurons weight and sum inputs the same way like _______ do to compute activation
binary threshold neurons
Linear and linear-nonlinear produce a _______ output (their firing rate) which is either just ______ to the activation (____) or is a ______, often ____, function of the activation
real valued; proportional; linear; non-linear; sigmoidal
Name 3 functions of the sigmoids
Logistic function
Arctan function
Auditory nerve fiber rate-level functions
Which model can model the input-output relationships of sensory neurons as ‘weight matrix’?
Receptive field model
What is ‘weight matrix’
It says what the synaptic weight of each given ‘spatial’ or ‘auditory frequency channel’ would be
Name 3 examples of receptive field models.
Center-surround receptive fields in retina or LGN
Simple cell receptive fields in V1
Spectro-temporal receptive fields in auditory structures
What functions do the linear or LN models serve?
Predictive power and falsifiability
What do the linear or LN models make easy to do?
Easy to predict what the response to any arbitrary stimulus should be
What do the Difference of Gaussians (DoG) model of retinal ganglion cell receptive fields posit?
The weight of a spot of light on the retina as a function of distance from the receptive center is given by a broad inhibitory (negative, surround) bell curve and a sharp excitatory (positive, center) bell curve
The centre-surround structure of retinal ganglion cells turns them into “______”
spatial frequency filters
Larger RGC receptive fields are tuned to “________” structure in the visual scene, while smaller RFs are tuned to ______ structure
coarsely grained; fine-grained
What kind of filter does the Gabor Filter Model of V1 simple cells?
Linear filter
Binaural STRFs in A1 can predict ______ & _______
spatial tuning; how the neurons respond, as a function of time, to the sounds coming from right/left ear
Binaural STRFs in A1 is a hugely ______ model. Because _____
parsimonious; it takes all the complicated auditory brain stem away for simplification
Rosenblatt’s Perceptron is the first “______”, basically just a _________
artificial neural net; single binary threshold neuron
Rosenblatt’s Perceptron is used as a “______”; (output goes high if inouts belong to class ___, low for class ______)
classifier; 1; 2
Rosenblatt’s Perceptron’s weight are set through ____ by thousands of examples: If the desired output for a given input set is high, then increase the weight of those inputs that were ___, but decrease those were ___
training; high; low
Rosenblatt’s Perceptron is a form of _____ learning
associative
A ______ perceptron cannnot learn functions like ‘exclusive or’ but ______ perceptrons can in principle compute any arbitrary function
single layer; multilayer
Rosenblatt’s Perceptron uses a ____ function to compute binary outputs. Multilayer perceptrons use _____ (____/____)
threshold; sigmoids; arctan; logistic
What does multilayer perceptrons mean?
Have one or more ‘hidden layers’ of artificial neurons between input and output
Weights are still set through ______ and ______, but errors during training stages must be ‘____’ to adjust the weights of the hidden layer units
training; associative learning; backpropagated
Are there any input-output mapping that a sufficiently big multilayer perceptron cannot learn?
No
What is deep neural network?
multilayer perceptrons with more than one hidden layer
After training, neurons in subsequent layers may become sensitive to ‘features’ of ever greater _______ and ______
complexity; abstraction
In ThreeBlueOneBrown: first hidden layer: ______
second hidden layer: _______
sensitive to little edges; sensitive to more complex shapes
In both _____neural nets, the features the hidden neurons are actually sensitive to, may be obscure.
real and artificial
Deep neural networks for image classification often have a hierarchy of layers sensitive to _________.
high level features
Deep neural networks for image classification is equivalent to various parts of the _______ and the_________ or _________.
ventral ‘what’ stream; dorsal ‘where’ stream; color areas of extrastriate visual areas
A _ layer RNN with only ____ hidden neurons trained on 4.4MB text file containing all the collected works of Shakespeare produced artificial Shakespeare.
3; 512
Multilayer perceptrons need to be trained on very ___ datasets, comprising tens of thousands to millions of ______, and for each example the ______, ‘_____’ answer must be
large; examples; desired; correct
Multilayer perceptrons: The synaptic weights of the network are adjusted ______ to make the network output more similar to the _______________
gradually; correct answers in the training set
Multilayer perceptrons: “Errors” (differences between current and correct output) must “________” through the ___________
backpropagate; layers of the network
What is the criticism of multilayer perceptrons?
Real brains do not learn like this, they don’t backpropagate/ teachers give us thousands of examples
What is unsupervised learning?
Learn types of things without ‘teacher’ input which tells them ‘correct’ answers
What will network under unsupervised learning discover?
natural ‘clustering’ of certain features
What way of learning is more biological?
unsupervised learning
What is reinforcement learning?
a network is not told what the output for a certain input should be, but it is simply ‘rewarded’ for ‘good’ outputs or ‘punished’ for ‘bad’ outputs
Reinforcement learning will try to predict which output (behaviour) lead to ________.
maximum reward and minimum punishment
What in the brain is thought to implement reinforcement learning?
Levels of the neurotransmitter dopamine representing ‘reward prediction error’
Temporal difference gammon is a very modestly sized network of less than ____ artificial neurons
300
Temporal difference gammon can beat more human players at _______
backgammon
Neural network, whether biological/ artificial, compute by _______________
combining weighted inputs to calculate outputs
Networks must often be trained by _______
adjusting the weights to produce useful outputs
The trained weight matrix is a form of “________”
long term memory
Recurrent neural networks also have “ ________” through activity running _______
short term memory; in loops
Subsequent layers of the network tend to become sensitive to _______________ of the inputs
increasingly higher order features
