Introduction to the TLearn Simulator

1
Introduction to the TLearn Simulator

• CS/PY 399 Lab Presentation 5
• February 8, 2001
• Mount Union College

2
TLearn Software

• Developed by Cognitive Psychologists to study
  properties of connectionist models and learning
• Kim Plunkett, Oxford
• Experimental Psychologist
• Jeffrey Elman, U.C. San Diego
• Cognitive Psychologist
• Simulates massively-parallel networks on serial
  computer platforms

3
Notational Conventions

• TLearn uses a slightly different notation than
  that which we have been using
• Input signals are treated as nodes in the
  network, and displayed on screen as squares
• Other nodes (representing neurons) are displayed
  as circles
• Input and output values can be any real numbers
  (decimals allowed)

4
Weight Adjustments Learning

• TLearn uses a more sophisticated rule than the
  simple one seen last week
• Let tkp be the target (desired) output for node k
  on pattern p
• Let okp be the actual (obtained) output for node
  k on pattern p

5
Weight Adjustments Learning

• Error for node k on pattern p (?kp ) is the
  difference between target output and observed
  output, times the derivative of the activation
  function for node k
• why? Dont ask! (actually, this value simulates
  actual observed learning)
• ?kp (tkp - okp) okp (1 - okp)

6
Weight Adjustments Learning

• This is used to calculate adjustments to weights
• Let wkj be the weight on the connection from node
  j to node k (backwards notation is what the
  authors use)
• Let ?wkj be the change required for wkj due to
  training
• ?wkj is determined by error for node k, input
  from node j, learning rate (?)

7
Weight Adjustments Learning

• ?wkj ? ?kp ojp
• ? is small (lt 1, usually 0.05 to 0.5), to keep
  weights from making wild swings that overshoot
  goals for all patterns
• This actually makes sense . . .
• a larger error (?kp) should make ?wkj larger
• if ojp is large, it contributed a great deal to
  the error, so it should contribute a large value
  to the weight adjustment

8
Weight Adjustments Learning

• The preceding is called the delta rule
• Used in Backpropagation Training
• error adjustments are propagated backwards from
  output layer to previous layers when weight
  changes are calculated
• Luckily, the simulator will perform these
  calculations for you!
• Read more in Ch. 1 of Plunkett Elman

9
TLearn Simulation Basics

• For each problem on which you will work, the
  simulator maintains a PROJECT description file
• Each project consists of three text files
• .CF file configuration information about the
  networks architecture
• .DATA file input for each of the networks
  training cases
• .TEACH file output for each training case

10
TLearn Simulation Basics

• Each file must contain information in EXACTLY the
  format TLearn expects, or else the simulation
  wont work
• Example AND project from Chapter 3 folder
• 2 inputs, one outupt, output 1 only if both
  inputs 1

11
.DATA and .TEACH Files
12
.DATA File format

• first line distributed or localist
• to start, well always use distributed
• second line n of training cases
• next n lines inputs for each training case
• a list of v values, separated by spaces, where v
  of inputs in network

13
.TEACH File format

• first line distributed or localist
• must match mode used in .DATA file
• second line n of training cases
• next n lines outputs for each training case
• a list of w values, separated by spaces, where w
  of outputs in network
• a value may be , meaning output is ignored
  during training for this pattern

14
.CF File
15
.CF File format

• Three sections
• NODES section
• nodes of non-input units in network
• inputs of inputs to network
• outputs of output units
• output node is ___ lt which node is the output
  node?
• gt 1 output node gt syntax changes to output
  nodes are

16
.CF File format

• CONNECTIONS section
• groups 0 ( explained later )
• 1 from i1-i2 (says that node 1 gets values
  from input nodes i1 and i2)
• 1 from 0 (says that node 1 gets values
  from the bias node -- explained below)
• input nodes always start with i1, i2, etc.
• non-input nodes start with 1, 2, etc.

17
.CF File format

• SPECIAL section
• selected 1 (special simulator results
  reporting)
• weight-limit 1.00 (range of random weight
  values to use in initial network creation)

18
Bias node

• TLearn units all have same threshold
• defined by logistic function
• ? values are represented by a bias node
• connected to all non-input nodes
• signal always 1
• weight of the connection is -?
• same as a perceptron with a threshold
• example on board

19
Network Arch. with Bias Node
20
.CF File Example (Draw it!)

• NODES
• nodes 5
• inputs 3
• outputs 2
• output nodes are 4-5
• CONNECTIONS
• groups 0
• 1-3 from i1-i3
• 4-5 from 1-3
• 1-5 from 0

21
Learning to use TLearn

• Chapter 3 of the Plunkett and Elman text is a
  step-by-step description of several TLearn
  Training sessions.
• Best way to learn Hands-on! Try Lab Exercise
  5

22
Introduction to the TLearn Simulator

• CS/PY 399 Lab Presentation 5
• February 8, 2001
• Mount Union College