Skip to main content

ANN Foundations

Source: Unit 2 §4

An artificial neural network is a computing model borrowed wholesale from the brain. Before any of the maths, it is worth seeing the thing being copied.

The biological inspiration

somagapnext neuronDendritesreceive signalsSomaprocess the informationAxontransmit the outputSynapseconnect to the next neuron
One neuron is a four-stage pipeline: collect, process, transmit, hand over. The artificial unit copies exactly this shape.
Biological partRole
DendriteReceives signals from other neurons
SynapsePoint of connection to other neurons
SomaProcesses the information
AxonTransmits the output of this neuron
NumbersThe scale the model is imitating
  • The brain is a massively parallel interconnected network of neurons, with the sense organs relaying information into the lowest layer.
  • An average human brain has roughly 101110^{11} neurons - 100 billion.
  • Parallelism buys division of work: each neuron plays one role, responding to one certain stimulus.

What ANNs are good for

Neural networks provide a robust approach to approximating real-valued, discrete-valued and vector-valued functions.

FactsWhere they earn their keep
  • Handwritten character recognition.
  • Spoken-word recognition.
  • Face recognition.
  • Self-driving cars.
  • Generally: interpreting real-world sensor data, which is noisy and high-dimensional and where no clean rule exists to write down.

When to use an ANN

DECISIONIs a neural network the right tool here?
Instances are described by many attribute-value pairsANN fitsit takes a wide input vector natively
The target is discrete-valued, real-valued, or a vector of such valuesANN fitsthe output layer can be any shape
The training examples may contain errorsANN fitsnetworks are robust to noise
Long training times are acceptableANN fitstraining is the expensive half; prediction is cheap
A human must be able to read and justify the learned rulePick a decision tree insteadANN weights are not interpretable
Pick this when: the inputs are numerous and noisy and nobody needs to read the model afterwards
Exam cueThe one-line summary

An ANN is the right answer when the question is "learn a function from a lot of noisy numeric evidence" and the wrong answer when the question is "explain the rule you learned".