10 - Synaptic plasticity and learning Flashcards
Is 95% of brain learning supervised or unsupervised?
unsupervised
What is unsupervised learning?
What does it make neurons do?
learning from input statistics only
- allow neurons to specialise by responding to certain patterns of input activities but not others
What are some forms of unsupervised learning?
development, learning, memory
What is an example of activity-dependent synaptic plasticity (unsupervised learning)?
Hebbian learning
What happens to amplitude of APs in Long-term Potentiation LTP?
How do induce LTP experimentally?
-increase in amplitude of post synaptic potential (V) / synaptic weight is increased
-increase current of presynaptic neuron/ increases spike of pre ->
What is LTD?
What effect is seen on postsynaptic neuron as a result of LTD?
-long-term depression
-reduced amplitude of AP
What sort if plasticity is LTD and LTP an example of?
Hebbian Plasticity
What is an example of synaptic plasticity in visual system?
How does this affect the eyes?
-ocular dominance columns in V1
-you have an eye which is more dominant than the other
How many neurons in the synaptic plasticity model?
What model is used thus?
two presynaptic, one postsynaptic
rate neuron model (like we saw in feedforward networks
What do x1 and x2 represent in synaptic plasticity model?
x1 left presynaptic neuron and vice versa
What is the assumption of the synaptic plasticity model?
What mathematical concept is used to put this assumption into place?
-Synaptic dynamics Is slower than neural dynamics: ππ€β«ππ
-separation of time scales
ππβ0 β π¦=π€1π₯1 + π€2π₯2
What are the computation properties of synaptic plasticity?
-locality: synapses do not communicate with eachother, ONLY itself -> thus weight of synapse depends on self
-stability: unlimited growth/decay of synaptic strength is inhibited
-development of selectivity to imput by increasing weight from that input (perceptual decision making/competition)
What learning rate is ππ€ replaced with in the synaptic plasticity model?
ππ€ with the now familiar learning rate πΌ=1/ππ€.
What does the coactivation of pre and postsynaptic neurons (x and y) of these types of plasticity lead to?
Standard Hebbian
Ani-Hebbian
Why ?
standard ->LTP (alpha is positive)
anti-hebbian -> LTD (because alpha is negative)
-/+ axy
Is non-hebbian plasticity activated by coactivation?
What is the result of presynatic activation and then postsynaptic activation for non-hebbian (seperate)
-no activated by just either pre or post synaptic neuron activation
presynatic potentiation
postsynaptic depression
Is non-hebbian plasiticty dependent on synaptic weights?
What is the equation?
no because it is works with either pre or postsynaptic activation (no coativation)
pre =ax or post =-ay
What happens to the learning rate when you have metplasticity?
you have weight-dependent learning rates πΌ=πΌ(π€π )
Does Standard Hebbian plasticity follow the computational properties of synaptic plasticity? (locality, stability, competition)
Why?
What do you add to standard hebbian plasticity eqn. to solve this?
-not stable and not competitive (yes locality)
2πΌπ¦^2 > 0
because y squared is always positive and goes to infinity
-stability -> weight threshold and a weight decay TERM (introduces LTD mechanism)
BCM model:
π¦<π_πβ ?
π¦>π_πβ ?
LTD
LTP
respectively
What does the BCM model separate?
Phase planes of BCM model, what are the axes?
-LTP from LTD from the introduction of SLIDING threshold of theta
- horizontal = y, vertical = dw/dt
Phase planes of BCM model:
What is the shape of the curve?
What direction does the curve shift when LTP and LTD happens?
-sigmoidal ish (bit more curved at start)
-left: LTD
right: LTP
What is the condition of theta and y in the BCM model?
Condition: ππ must grow more rapidly than π¦
How does BCM model cause stability when LTP happens?
What biologically concept is this an example of?
-An increase in π€_π due to LTP increases postsynaptic activity π¦, increasing the threshold π_π, making it harder to induce LTP, stabilising the dynamics.
-homeostasis
