M06 - Cutting edge paper

Question 1

What is this article about?

Answer 1

Hebbian deep learning without feedback. Getting rid of backpropagation and use a different way to train a network that matches more to the brain does it.

Question 2

What problem does this paper address?

Answer 2

Deep Learning models are not bio-realistic in reward assignment (backpropagation not realistic)

Question 3

What is the goal of this paper?

Answer 3

• Train a deep learning model with a Hebbian learning rule
• Create bio-plausible Machine Learning

Question 4

What were the results? (in sum)

Answer 4

With up to 5 hidden layers and an added linear classifier reached relatively high accuracies

Question 5

What is synaptic plasticity?

Answer 5

Neurons being co-active and weight updates

Question 6

What is Spike-timing dependent plasticity?

Answer 6

The timing of pre- and post-synaptic spike determines the sign of the weight update

Question 7

What are the problems with back-propagation?

Answer 7

• weight transport
• non-local plasticity
• update locking
• global loss function

Question 8

What is the problem of weight transport mean in back propagation in connection to the neural network?

Answer 8

Backpropagation of error requires synapses (=weights) to invert or work backwards. This is not biologically plausible.

Question 9

What is the problem with non-local plasticity in back propagation in connection to the neural network?

Answer 9

In NN, weights are updated if the pre- and post-synaptic cells are co-active. With BP, teaching signals are computed later and at a different place.
= computational inefficiency