Neuromorphic Computing

Question 1

Computers are [serial/parallel] machines, while the brain is [serial/parallel].

Answer 1

Serial, parallel

Question 2

Neurons communicate through [cells/synapses], which are [static/configurable] chemical junctions between neurons.

Answer 2

Synapses, configurable

Question 3

The idea that neurons handle information via synapses is called the theory of…

Answer 3

Connectionism

Question 4

A neuron is composed of three parts…

Answer 4

The dendrite, the cell body, and the axon

Question 5

The dendrite in a neuron…

Answer 5

Receives the signals from a synapse

Question 6

The cell body in a neuron..

Answer 6

Processes signals received via the dendrite

Question 7

The axon in a neuron…

Answer 7

Sends signals across a synapse

Question 8

The axon adjusts the intensity of its pulse based on…

Answer 8

The strength of the incoming pulse

Question 9

The strength of a synaptic pulse represents…

Answer 9

How strongly connected the memory is

Question 10

Neural networks are discrete algorithms that are able to learn how to relate…

Answer 10

Inputs to a required output

Question 11

Common uses of neural networks are for the tasks of…

Answer 11

Regression and classification

Question 12

Neural networks must undergo [retaining/degradation] of information, in order to…

Answer 12

Degradation, in order to simulate the brain’s fault tolerance

Question 13

The term neuromorphic computing is an umbrella term for…

Answer 13

Brain-inspired techniques

Question 14

The parallel operation property of neuromorphic computers refers to…

Answer 14

All neurons being allowed to operate simultaneously

Question 15

The collocated processing and memory property of neuromorphic computers refers to…

Answer 15

Processing and memory being one and the same, meaning lower energy consumption than memory

Question 16

The scalability property of neuromorphic computers refers to…

Answer 16

The fact that we can simply add more chips with more neurons to add processing power

Question 17

The event-driven processing property of neuromorphic computers refers to…

Answer 17

Processing only occurring when a spike arrives, reducing CPU uptime

Question 18

In neuromorphic computers, each neuron has an […] which can communicate with others or itself.

Answer 18

Axon

Question 19

Instead of transmitting data via continuous signals like von Neumann architectures, neuromorphic architectures transmit data using…

Answer 19

Discrete signals

Question 20

Information in a neuromorphic computer is encoded as… (pick one)

Answer 20

Spike rates, timing or neuron population distribution

Question 21

A neuron’s charge is a value that accumulates over time from…

Answer 21

Input signals, or signals received from other neurons

Question 22

When a neuron’s charge hits a threshold…

Answer 22

The neuron fires, and emits a signal across each of its connections

Question 23

Synapses adjust their strength based on…

Answer 23

Activity patterns

Question 24

The synapse connection is strengthened if the [pre/post]-synaptic neuron fires first, otherwise it is weakened.

Answer 24

Pre

Question 25

SNNs (Synaptic Neural Networks) cannot be trained via backpropagation because...

Answer 25

The data is encoded discretely, making it non-differentiable

Question 26

Since they cannot be backpropagated, SNNs train via...

Answer 26

Training a DNN (Deep Neural Network) and adapt the weights for use in an SNN (like transfer learning!)

Question 27

A neural architecture search is often formulated as a [...]-level optimisation problem, wherein we want to optimise both...

Answer 27

Bi, structure and weights

Question 28

Neuromorphic computers experience many challenges, such as... (pick three)

Answer 28

Accessibility, usability, efficiency, cost, lack of benchmarks, too low-level

Question 29

The aim of a self-organising map is to learn how to...

Answer 29

Map points from a high-dimensional space to a low-dimensional space that preserves topological properties by mapping local and global relationships

Question 30

An example of a self-organising map is...

Answer 30

The feature map of a convolutional layer

Question 31

Self-organising maps are a form of [supervised/unsupervised] learning technique.

Answer 31

Unsupervised

Question 32

Self-organising maps make the assumption that data that belongs to the same class shares...

Answer 32

Common features that will be both identifiable and organisable by the self-organising map

Question 33

The number of input nodes in a self-organising map is...

Answer 33

As many features are in the input data

Question 34

Each input to a self-organising map corresponds to...

Answer 34

A node in the output map, chosen via some competitive process measured via some distance function

Question 35

The learning process of a self-organising map involves setting weights to small [fixed-size/random] values, finding the node [closest/furthest] to the input (this is called the Best Matching Unit), then slowly [increasing/decreasing] the size of the neighbourhood of the values until convergence.

Answer 35

Random, closest, decreasing

Question 36

The learning process of a self-organising map achieves a high early learning rate that decreases over time by ensuring that our [...] and [...] is large initially.

Answer 36

Neighbourhood size, learning rate coefficient

Question 37

Initialisation techniques are used in self-organising maps to ensure that...

Answer 37

The network unfolds properly