Intro / Neural Systems Flashcards
Total venture capital money for pure AI startups 2014
$300million
Venture capital investment in robots in 2015
$587million (double the amount of 2014)
Percent of people in UK who have heard the term ‘machine learning’? Where from?
9%
75% mainstream media, 21% entertainment
Simple low tech example of Bio-inspired tech from burrs?
Velcro
How many neurons in your brain?
86 billion
How many synapses in your brain
10 trillion
How many possible calculations per second?
10 quadrillion
Name two Brain projects
BRAIN Initiative $300+ million
Human Brain Project €1 billion over 10 yrs
What increase needed to bridge gap between best computers and human brain?
100,000 times
Name some parts of a real neuron
Dendrites, Axon, Soma
Name parts of Synapse
Axon, Synaptic cleft, neurotransmitters, receptor molecules, dendrite
Two main types of interneurons
Inhibitory, excitatory
Three examples of output functions
Identity, step, sigmoid
Hebb’s rule is used when
updating weights in learning
Hebb’s rule suffers from
self-amplification (unbounded growth of weights)
A neural network with a single layer is called a
perceptron
Can a single perceptron separate XOR inputs?
No, not linearly separable, need multilayer
The three types of units in a multilayer neural networks?
Input, hidden and output
Which output function sound should a multi-layer network no use?
Multi-layer networks should not use linear output functions because a linear transformation remains a linear transformation. Therefore, such a network would be equivalent to a networks with a single layer.
In an MLP how is the error of the hidden units found?
Backpropagation of error
Eight steps in backpropagation of error
- Initialize weights (random, around 0)
- Present pattern
- Compute hidden
- Compute output
- Compute delta output
- Compute delta hidden
- Compute weight changes
- Update weights (back to 2)
Name a solution to overfitting
Use a validation set
