1 RECENT DEVELOPMENTS IN MULTILAYER PERCEPTRON NEURAL NETWORKS Walter H. Delashmit Lockheed Martin Missiles and Fire Control Dallas, TX Michael T. Manry The University of Texas at Arlington Arlington, TX Memphis Area Engineering and Science Conference 2005 May 11, 2005

2 Outline of Presentation Review of Multilayer Perceptron Neural Networks Network Initial Types and Training Problems Common Starting Point Initialized Networks Dependently Initialized Networks Separating Mean Processing Summary

3 Review of Multilayer Perceptron Neural Networks

4 Typical 3 Layer MLP

5 MLP Performance Equations Mean Square Error (MSE): Output: Net Function:

6 Net Control Scales and shifts all net functions so that they do not generate small gradients and do not allow large inputs to mask the potential effects of small inputs

7 Neural Network Training Algorithms Backpropagation Training Output Weight Optimization – Hidden Weight Optimization (OWO-HWO) Full Conjugate Gradient

8 Output Weight Optimization – Hidden Weight Optimization (OWO-HWO) Used in this development Linear equations used to solve for output weights in OWO Separate error functions for each hidden unit are used and multiple sets of linear equations solved to determine the weights connecting to the hidden units in HWO

9 Network Initial Types and Training Problems

10 Problem Definition Assume that a set of MLPs of different sizes are to be designed for a given training data set Let be the set of all MLPs for that training data having N h hidden units, E int (N h ) denote the corresponding training error of am initial network that belongs to Let E f (N h ) denote the corresponding training error of a well-trained network Let N hmax denote the maximum number of hidden units for which networks are to be designed Goal: Choose a set of initial networks from {S 0, S 1, S 2, … }such that E int (0)  E int (1)  E int (2)  …. E int (N hmax ) and train the network to minimize E f (N h ) such that E f (0)  E f (1)  E f (2)  …. E f (N hmax ) Axiom 3.1: If E f (N h )  E f (N h -1) then the network having N h hidden units is useless since the training resulted in a larger, more complex network with a larger or the same training error.

11 Network Design Methodologies Design Methodology One (DM-1) – A well-organized researcher may design a set of different size networks in an orderly fashion, each with one or more hidden units than the previous network oThorough design approach oMay take longer time to design oAllows achieving a trade-off between network performance and size Design Methodology Two (DM-2) – A researcher may design different size networks in no particular order oMay be quickly pursued for only a few networks oPossible that design could be significantly improved with a bit more attention to network design

12 Three Types of Networks Defined Randomly Initialized (RI) Networks – No members of this set of networks have any initial weights and thresholds in common. Practically this means that the initial random number seeds (IRNS) are widely separated. Useful when the goal is to quickly design one or more networks of the same or different sizes whose weights are statistically independent of each other. Can be designed using DM-1 or DM-2 Common Starting Points Initialized (CSPI) Networks – When a set of networks are CSPI, each one starts with the same IRNS. These networks are useful when it is desired to make performance comparisons of networks that have the same IRNS for the starting point. Can be designed using DM-1 or DM-2 Dependently Initialized (DI) Networks – A series of networks are designed with each subsequent network having one or more hidden units than the previous network. Larger size networks are initialized using the final weights and thresholds from training a smaller size network for the values of the common weights and thresholds. DI networks are useful when the goal is a thorough analysis of network performance versus size and are most relevant to being designed using DM-1.

13 Network Properties Theorem 3.1: If two initial RI networks (1) are the same size, (2) have the same training data set and (3) the training data set has more than one unique input vector, then the hidden unit basis functions are different for the two networks. Theorem 3.2: If two CSPI networks (1) are the same size and (2) use the same algorithm for processing random numbers into weights, then they are identical. Corollary 3.2: If two initial CSPI networks are the same size and use the same algorithm for processing random numbers into weights, then they have all common basis functions.

14 Problems with MLP Training Non-monotonic E f (N h ) No standard way to initialize and train additional hidden units Net control parameters are arbitrary No procedure to initialize and train DI networks Network linear and nonlinear component interference

15 Mapping Error Examples

16 Tasks Performed in this Research Analysis of RI networks Improved Initialization in CSPI networks Improved initialization of new hidden units in DI networks Analysis of separating mean training approaches

17 CSPI and CSPI-SWI Networks Improvement to RI networks Each CSPI network starts with same IRNS Extended to CSPI-SWI (Structured Weight Initialization) networks oEvery hidden unit of the larger network has the same initial weights and threshold values as the corresponding units of the smaller network oInput to output weights and thresholds are also identical Theorem 5.1: If two CSPI networks are designed with structured weight initialization, the common subset of the hidden unit basis functions are identical. Corollary 5.1: If two CSPI networks are designed using structured weight initialization, the only initial basis functions that are not the same are the hidden unit basis functions for the additional hidden units in the larger network. Detailed flow chart for CSPI-SWI initialization in dissertation

18 CSPI-SWI Examples fmtwod

19 DI Network Development and Evaluation Improvement over RI, CSPI and CSPI-SWI networks The values of the common subset of the initial weights and thresholds for the larger network are initialized with the final weights and thresholds from a previously well-trained smaller network Designed with DM-1 Single network designs  networks are implementable After training, testing is feasible on a different set of data set

20 Create an initial network with N h hidden units Train this initial network N h  N h +p N h >N hmax ? Initialize new hidden units N h-p+1  j  N h w oh (k,j)  0, 1  k  M w hi (j,i)  RN(ind+), 1  i  N+1 Net control for w hi (j,i), 1  i  N+1 Train new network Stop Yes No Basic DI Network Flowgraph

21 Properties of DI Networks E int (N h ) < E int (N h-p ) E f (N p ) curve is monotonic non-increasing (i. e., E f (N h )  E f (N h-p )) E int (N h ) = E f (N h-p )

22 Performance Results for DI Networks with Fixed Iterations fmtwod F24F17

23 RI Network and DI Network Comparison (1) DI network: standard DI network design for N h hidden units (2) RI type 1: RI networks were designed using a single network for each value of N h and every network of size N h was trained using the value of N iter that the corresponding network was trained with for the DI network. (3) RI type 2: RI networks were designed using a single network for each value of N h and every network was trained using the total number of N iter that was used for the entire sequence of DI networks. This can be expressed by This results in the RI type 2 network actually having a larger value of N iter than the DI network.

24 RI Network and DI Network Comparison Results fm twod

25 Separating Mean Processing Techniques Bottom-Up Separating Mean Top-Down Separating Mean

26 Generate linear mapping results p t  pp tt   Train MLP using new data ppp ttx  , Bottom-Up Separating Mean Basic Idea: A linear mapping is removed from the training data. The nonlinear fit to the resulting data may perform better. Generate new desired output vector

27 Bottom-up Separating Mean Results fmpower12 single2

28 Top-Down Separating Mean Basic Idea: If we know which subsets of inputs and outputs have the same means in Signal Model 2 and 3, we can estimate and remove these means. Network performance is more robust.

29 Separating Mean Results power12

30 Conclusions On the average CSPI-SWI networks have more monotonic non- increasing MSE versus N h curves than RI networks MSE versus N h curves are always monotonic non-increasing for DI networks DI network training was improved by calculating the number of training iterations and limiting the amount of training used for previously trained units DI networks always produce more consistent MSE versus N h curves than RI, CSPI and CSPI-SWI networks Separating mean processing using both a bottom-up and top-down architecture often produce improved performance results A new technique was developed to determine which inputs and outputs are similar to use for top-down separating mean processing