Backpropagation

Networks

Introduction to

Backpropagation
- In 1969 a method for learning in multi-layer network, Backpropagation, was invented by Bryson and Ho. - The Backpropagation algorithm is a sensible approach for dividing the contribution of each weight.

- Works basically the same as perceptrons

Backpropagation Learning Principles: Hidden Layers and Gradients

There are two differences for the updating rule :

1) The activation of the hidden unit is used instead of activation of the input value.
2) The rule contains a term for the gradient of the activation function.

Backpropagation Network training

1. Initialize network with random weights 2. For all training cases (called examples):
a. Present training inputs to network and calculate output b. For all layers (starting with output layer, back to input layer):
i. Compare network output with correct output (error function) ii. Adapt weights in current layer
This is what you want

Backpropagation Learning Details

Method for learning weights in feed-forward (FF) nets Cant use Perceptron Learning Rule
no teacher values are possible for hidden units

Use gradient descent to minimize the error

propagate deltas to adjust for errors backward from outputs to hidden layers forward to inputs
backward

Backpropagation Algorithm Main Idea error in hidden layers

The ideas of the algorithm can be summarized as follows :

1. Computes the error term for the output units using the observed error. 2. From output layer, repeat - propagating the error term back to the previous layer and - updating the weights between the two layers until the earliest hidden layer is reached.

Backpropagation Algorithm
Initialize weights (typically random!) Keep doing epochs For each example e in training set do
forward pass to compute
O = neural-net-output(network,e) miss = (T-O) at each output unit

backward pass to calculate deltas to weights update all weights

end until tuning set error stops improving

Forward pass explained earlier Backward pass explained in next slide

Backward Pass
Compute deltas to weights
from hidden layer to output layer

Without changing any weights (yet), compute the actual contributions

within the hidden layer(s) and compute deltas

Gradient Descent
Think of the N weights as a point in an Ndimensional space Add a dimension for the observed error Try to minimize your position on the error surface

Error Surface
error

weights Error as function of weights in multidimensional space

Compute deltas

Gradient

Trying to make error decrease the fastest Compute:

GradE = [dE/dw1, dE/dw2, . . ., dE/dwn]

Change i-th weight by

deltawi = -alpha * dE/dwi
Derivatives of error for weights

We need a derivative! Activation function must be continuous, differentiable, non-decreasing, and easy to compute

Cant use LTU

To effectively assign credit / blame to units in hidden layers, we want to look at the first derivative of the activation function Sigmoid function is easy to differentiate and easy to compute forward
Linear Threshold Units Sigmoid function

Updating hidden-to-output
We have teacher supplied desired values deltawji = * aj * (Ti - Oi) * g(ini)
= * aj * (Ti - Oi) * Oi * (1 - Oi)
for sigmoid the derivative is,

g(x) = g(x) * (1 - g(x))

derivative

alpha

Here we have general formula with derivative, next we use for sigmoid

miss

Updating interior weights

Layer k units provide values to all layer k+1 units
miss is sum of misses from all units on k+1 missj = [ ai(1- ai) (Ti - ai) wji ] weights coming into this unit are adjusted based on their contribution deltakj = * Ik * aj * (1 - aj) * missj For layer k+1

Compute deltas

How do we pick ?
1. Tuning set, or 2. Cross validation, or 3. Small for slow, conservative learning

How many hidden layers?

Usually just one (i.e., a 2-layer net) How many hidden units in the layer?
Too few ==> cant learn Too many ==> poor generalization

How big a training set?

Determine your target error rate, e Success rate is 1- e Typical training set approx. n/e, where n is the number of weights in the net Example:
e = 0.1, n = 80 weights training set size 800 trained until 95% correct training set classification should produce 90% correct classification on testing set (typical)

Examples of Backpropagation Learning

In the restaurant problem NN was worse than the decision tree

Error decreases with number of epochs

Decision tree still better for restaurant example

Examples of Backpropagation Learning

Majority example, perceptron better

Restaurant example, DT better

Backpropagation Learning Math

See next slide for explanation

Visualization of Backpropagation learning

Backprop output layer

Bias Neurons in Backpropagation Learning

bias neuron in input layer

Software for Backpropagation Learning

Training pairs
This routine calculate error for backpropagation

Run network forward. Was explained earlier

Calculate difference to desired output

Calculate total error

Software for Backpropagation Learning continuation

Here we do not use alpha, the learning rate

Update output weights

Calculate hidden difference values

Update input weights

Return total error

The general Backpropagation Algorithm for updating weights in a multilayer network Repeat until convergent

Here we use alpha, the learning rate

Go through all examples

Run network to calculate its output for this example

Compute the error in output Update weights to output layer Compute error in each hidden layer

Update weights in each hidden layer

Return learned network

Examples and Applications of ANN

Neural Network in Practice

NNs are used for classification and function approximation or mapping problems which are:

- Tolerant of some imprecision.

- Have lots of training data available.

- Hard and fast rules cannot easily be applied.

NETalk (1987)
Mapping character strings into phonemes so they can be pronounced by a computer Neural network trained how to pronounce each letter in a word in a sentence, given the three letters before and three letters after it in a window Output was the correct phoneme Results
95% accuracy on the training data 78% accuracy on the test set

Other Examples
Neurogammon (Tesauro & Sejnowski, 1989)
Backgammon learning program

Speech Recognition (Waibel, 1989) Character Recognition (LeCun et al., 1989) Face Recognition (Mitchell)

ALVINN
Steer a van down the road
2-layer feedforward
using backpropagation for learning

Raw input is 480 x 512 pixel image 15x per sec

Color image preprocessed into 960 input units

4 hidden units 30 output units, each is a steering direction

Neural Network Approaches

ALVINN - Autonomous Land Vehicle In a Neural Network

 ALVINN learned as the vehicle traveled

initially by observing a human driving

Learning on-thefly

learns from its own driving by watching for future corrections

never saw bad driving didnt know what was dangerous, NOT correct computes alternate views of the road (rotations, shifts, and fill-ins) to use as bad examples keeps a buffer pool of 200 pretty old examples to avoid overfitting to only the most recent images

Feed-forward vs. Interactive Nets

Feed-forward
activation propagates in one direction We usually focus on this

Interactive
activation propagates forward & backwards propagation continues until equilibrium is reached in the network We do not discuss these networks here, complex training. May be unstable.

Ways of learning with an ANN

Add nodes & connections Subtract nodes & connections Modify connection weights
current focus can simulate first two

I/O pairs:
given the inputs, what should the output be? [typical learning problem]

More Neural Network Applications

- May provide a model for massive parallel computation. - More successful approach of parallelizing traditional serial algorithms. - Can compute any computable function. - Can do everything a normal digital computer can do. - Can do even more under some impractical assumptions.

Neural Network Approaches to driving

Use special hardware
ASIC FPGA analog

- Developed in 1993.
Output units

Hidden layer

- Performs driving with Neural Networks. - An intelligent VLSI image sensor for road following.

Input units

- Learns to filter out image details not relevant to driving.

Neural Network Approaches

Input Array Hidden Units Output units

Actual Products Available

ex1. Enterprise Miner:
- Single multi-layered feed-forward neural networks. - Provides business solutions for data mining.

ex2. Nestor:
- Uses Nestor Learning System (NLS). - Several multi-layered feed-forward neural networks. - Intel has made such a chip - NE1000 in VLSI technology.

Ex1. Software tool - Enterprise Miner

- Based on SEMMA (Sample, Explore, Modify, Model, Access) methodology.
- Statistical tools include : Clustering, decision trees, linear and logistic regression and neural networks. - Data preparation tools include : Outliner detection, variable transformation, random sampling, and partition of data sets (into training, testing and validation data sets).

Ex 2. Hardware Tool - Nestor

- With low connectivity within each layer.
- Minimized connectivity within each layer results in rapid training and efficient memory utilization, ideal for VLSI. - Composed of multiple neural networks, each specializing in a subset of information about the input patterns.

- Real time operation without the need of special computers or custom hardware DSP platforms
Software exists.

Problems with using ANNs

1. Insufficiently characterized development process compared with conventional software
What are the steps to create a neural network?

2. How do we create neural networks in a repeatable and predictable manner? 3. Absence of quality assurance methods for neural network models and implementations
How do I verify my implementation?

Solving Problem 1 The Steps to create a ANN

Define the process of developing neural networks: 1. Formally capture the specifics of the problem in a document based on a template 2. Define the factors/parameters for creation
Neural network creation parameters Performance requirements

3. Create the neural network 4. Get feedback on performance

Neural Network Development Process

Problem Specification Phase

Some factors to define in problem specification:
Type of neural networks (based on experience or published results) How to collect and transform problem data Potential input/output representations Training & testing method and data selection Performance targets (accuracy and precision)

Most important output is the ranked collection of factors/parameters

Problem 2 How to create a Neural Network

Predictability (with regard to resources)
Depending on creation approach used, record time for one iteration Use timing to predict maximum and minimum times for all of the combinations specified

Repeatability
Relevant information must be captured in problem specification and combinations of parameters

Problem 3 - Quality Assurance

Specification of generic neural network software (models and learning) Prototype of specification Comparison of a given implementation with specification prototype Allows practitioners to create arbitrary neural networks verified against models

Two Methods for Comparison

Direct comparison of outputs:
20-10-5 (with particular connections and input) Prototype <0.123892, 0.567442, 0.981194, 0.321438, 0.699115>

Implementation

<0.123892, 0.567442, 0.981194, 0.321438, 0.699115>

Verification of weights generated by learning algorithm:

20-10-5 Prototype Iteration 100 Iteration 200 . Iteration n Weight state n Weight state n Weight state 1 Weight state 2 .

Implementation Weight state 1 Weight state 2 .

Further Work on improvements

Practitioners to use the development process or at least document in problem specification Feedback from neural network development community on the content of the problem specification template Collect problem specifications and analyse to look for commonalities in problem domains and improve predictability (eg. control) More verification of specification prototype

Further Work (2)

Translation methods for formal specification Extend formal specification to new types Fully prove aspects of the specification Cross discipline data analysis methods (eg. ICA, statistical analysis) Implementation of learning on distributed systems
Peer-to-peer network systems (farm each combination of parameters to a peer)

Remain unfashionable

Summary
- Neural network is a computational model that simulate some properties of the human brain.

- The connections and nature of units determine the behavior of a neural network.
- Perceptrons are feed-forward networks that can only represent linearly separable functions.

Summary

- Given enough units, any function can be represented by Multi-layer feed-forward networks. - Backpropagation learning works on multi-layer feed-forward networks. - Neural Networks are widely used in developing artificial learning systems.

References
- Russel, S. and P. Norvig (1995). Artificial Intelligence - A Modern Approach. Upper Saddle River, NJ, Prentice Hall. - Sarle, W.S., ed. (1997), Neural Network FAQ, part 1 of 7: Introduction, periodic posting to the Usenet newsgroup comp.ai.neural-nets, URL: ftp://ftp.sas.com/pub/neural/FAQ.html

Sources
Eric Wong

Eddy Li
Martin Ho

Kitty Wong