Hoorcollege 11-12

Question 1

Three main categories of learning strategies

Answer 1

Unsupervised learning: the neural network receives input from the outside world, and its synaptic weights change as a consequence of this input.

Supervised learning: the neural network receives input from the outside world and also the desired output, so that the network can change its synaptic weights to reach such output.

Reinforcement learning: the neural network receives input from the outside world, and a reward/punishment teaching signal which bias the learning towards a desired output

Question 2

biologische voorbeelden learning strategies

Answer 2

• Unsupervised learning: for example, receptive fields.
• Supervised learning: links with biological mechanisms still unclear. A good
  candidate is learning in the cerebellum (teaching signals).
• Reinforcement learning: classical conditioning.

Question 3

BCM rule

Answer 3

probeert leren in visual system te verklaren
extensie van hebb rule
lost twee belangrijke aspecten van de stabiliteitsproblem van heb op

BCM voegt een depressiefactor toe waardoor synpasen slapper kunnen worden
En het voegt een drempel toe zodat een synaps niet overgeactiveerd kan worden

Question 4

overeenkomst verschil BCM met Oja

Answer 4

overeenkomst proberen beide het stability probleem van Hebb op te lossen
Verschil Oja voegt een forgetting term toe, zodat strong synapsen niet nog sterker kunnen worden

Question 5

Waarom moeten er regels worden toegevoegd voor spikes

Answer 5

van nature is spikes niet mogelijk in computerprogrammas

Question 6

Spike Timing Dependent Plasticity STDP rule

Answer 6

pre synaps voor post synaps= versterking
post synaps voor pre synaps= verslapping

Question 7

triplet rule

Answer 7

je krijgt alleen LTP bij POST-PRE-POST activatie.

Question 8

Welke spikes en depolarisaties zijn noodzakelijk voor LTP

Answer 8

presynaptische spike en postsynaptische depolarisatie(geen spike nodig)

Question 9

burst rule

Answer 9

regel die niet hebb rule volgt. geeft aan dat burst van neurons die snel na elkaar komen, LTD opwekken.
Vissen gebruiken dit in het water om magneetvelden te detecteren.

Question 10

boolean functie AND & OR

Answer 10

twee synapsen vuren op projectordendriet, deze vuurt alleen door als beide synapsen vuren(AND) of als maar 1 van de twee vuurt(OR)

Question 11

Perceptron & classifiers

Answer 11

perceptrons worden gebruikt als classiefiers, meerde factoren worden dan berekend om een scheidslijn te maken tussen hond en kat bijvoorbeeld.

Question 11

Hoe wordt het opgelost dat de lijn van de classifiers niet getrokken kan worden omdat beide soorten door elkaar lopen qua parameters

Answer 11

een extra hidden layer bijvoorbeeld waardoor er twee lijnen komen te ontstaan

Question 11

universal approximation theorem

Answer 11

It can be mathematically proven that a feedforward (FF) neural network with at
least one hidden layer is able to approximate any continuous function, and
therefore solve in principle any classification problem given enough training,
input, etc

Question 12

deep neural networks

Answer 12

When the number of hidden layers is sufficiently large, FF neural networks are usually referred to as deep neural networks.

FF neural networks can contain more than one hidden layer. This improves their computing power to even higher levels, and allows them to solve problems like face recognition, speech-to-text translation, real-world object classification, etc.

Question 13

backpropagation

Answer 13

In a nutshell, backpropagation works like this:
1. Compare the real vs desired output of the model
2. Slightly change the weights in the direction that decreases the error
3. Repeat this process as you travel backwards in the network.

Question 14

Problems and limitations of deep neural networks:

Answer 14

Problems and limitations of deep neural networks:
* Catastrophic forgetting: networks tend to forget previously learned
information when learning new things.
* Overfitting: if the networks are trained too much, they reach bad
conclusions that are not generalizable.
* Existence of many local minima often
prevents to reach optimal solution.
* They are not biologically realistic at the moment.

Question 15

reinforcement learning

Answer 15

the neural network receives input from the outside
world, and a reward/punishment teaching signal which bias the learning
towards a desired output.
* Repetition/classic conditioning. Function of dopamine in the brain.
* Basal ganglia: action selection.
* Cerebellum: error-driven learning
* Cortex: predictive coding.

Question 16

Temporal differences (TD) rule

Answer 16

Toevoegen van een reward in de toekomst,

Question 17

the actor-critic model

Answer 17

• The actor selects a behavior (e.g. go vs no-go) based on sensory input and
  expected reward, and such behavior leads to a reward;
• The critic compares the reward obtained with the reward expected, and
  calculates the reward prediction error (via TD rule);
• Both actor and critic learn (i.e. modify their expectations) with the help of
  the prediction error.