Some Applications of Artificial Neural Networks

1
Some Applications of Artificial Neural Networks

• From Zuradas book

2
Neural Networks as Classifiers

• This is one of the best known applications of
  neural nets
• Examples text recognizers, land feature
  recognizer, etc.
• However, to begin with well look at a really
  simple classifier.

3
A linear classifier of cube vertices according to
this expression
4
What to do when membership requirements are more
complex?

• The previous example was easy because membership
  requirement was simple
• The neural net we used mapped the entire 3D space
  into just 2 pts 1 and 1
• The thresholding element used an abrupt sgn
  function
• This function has no inverse.
• Cant train network easily or at all
• Bad if membership is not explicitly stated
  mathematically

5
Squashed sgn function

• NNs for real, complex classification problems
  use a squashed or smooth sgn function
• Such a function is not really a sgn function
  anymore, of course
• But it has an inverse
• The invertibility makes the network more
  trainable
• The training procedures are pretty independent of
  the network architecture or the problem involved

6
EEG Spike Recognition

• ANNs have been used successfully for preliminary
  EEG processing (Eberhart Dobbins 1990)
• Up to 64 electrodes
• There are lots of data from all night monitoring
• Too much work for doctors, need help
• ANN good enough for preprocessing to recognize
  certain kinds of signals such as spikes

7
Details

• Sampled 4 channels of EEG waves interest
• 200 or 250 times a second
• 240ms time window
• Yields 48 or 60 data samples from each channel

8
Details (2)

• Fed data from each channel an ANN with squashed
  sgn
• 41 units arranged in 3 layers does data
  processing
• Two outputs, one identifying the input as a
  spike, the other as a non-spike

9
Design and training teams

• Network designed by a team of signal processing
  engineers
• Training done with 4-6 neurologists who
  identified spikes or non-spikes
• Network trained extensively using both spikes and
  non-spikes

10
Results

• Impressive
• 6 series of comprehensive tests reported
• First, network was tested with spike/non-spike
  waves used earlier for training
• All spikes positively identifed
• Only 2 of 260 non-spikes misclassified as spikes
• Training with totally new data also found all
  spikes and misidentified a few non-spikes
• Since spikes are bad this simply result in
  false alarms, not undetected trouble
• These results were considered better than that
  which was required for practical application in
  hospitals

11
Function Approximation

• Some functions can be slow to compute
• But we can often approximate them with a
  feed-forward neural network (layers with
  connections in forward (input-to-output)
  direction only

12
Function Approximation

• 21 training points
• 31 parameters!
• 20 weights
• 11 biases

13
ALVINN Vehicle Driving System (Pomerleau, 1989)

• 2-level ANN used for figuring out how much to
  turn the steering wheel
• 45 output neurons with 0/1 outputs
• Middle is straight
• Leftmost one represents most extreme left turn
• Etc.

14
ALVINN inputs

• Consist of video and range information 1 extra
  input
• Video info provided by 30x32 retina, depicting
  road scene
• Resulting 960 segments each coded into input
  proportional to intensity of blue
• Blue band has highest contrast betw road and
  non-road)
• Range (distance) info provided by a second
  retina, 8x32
• These units get signal from a laser rangefinder
• 1 extra input says whether road is light or dark
  at prev time step

15
ALVINN training and performance

• Trained with computer-generated road images
• Involved 1200 different combinations of scenes,
  curvatures, lighting conditions and distortion
  levels
• Entire driver implemented on an on-board computer
  and a modified Chevy van!
• Performed comparably to the best traditional
  vision-based navigation systems evaluated under
  similar coditions
• Training was done in half-an-hour!
• Was training done on board?
• For comparison -- Algorithm-based drivers take
  months for algorithm development