Skip to content

Latest commit

 

History

History
207 lines (160 loc) · 6.79 KB

PoC_Ex2_nDFMNN.md

File metadata and controls

207 lines (160 loc) · 6.79 KB
Raw

Proof of Concept: n-D Fully Monotonic Neural Networks

This script provides an illustration of the constrained deep learning for n-dimensional monotonic networks. The example will work through these steps:

  1. Generate a dataset of a 2-dimensional monotonic function.
  2. Prepare the dataset for custom training loop.
  3. Create a fully monotonic neural network (FMNN) architecture.
  4. Train the FMNN using a custom training loop and apply projected gradient descent to guarantee monotonicity.
  5. Explore 1-dimensional restrictions through the function to see monotonicity is guaranteed

Generate Data

First, take the monotonic function f(x1,x2)=x1^3+x2^3 and uniformly randomly sample this over the square region: [-1,1]x[-1,1]. You can change the number of random samples if you want to experiment.

numSamples = 1024;
[x1Train,x2Train] = meshgrid(linspace(-1,1,round(sqrt(numSamples))));
xTrain = [x1Train(:),x2Train(:)];
tTrain =xTrain(:,1).^3 + xTrain(:,2).^3;

Visualize the data.

figure;
plot3(xTrain(:,1),xTrain(:,2),tTrain,"k.")
grid on
xlabel("x1")
ylabel("x2")

Observe the overall monotonic behaviour in x2 given x1, and in x1 given x2.

Prepare Data

To prepare the data for custom training loops, add the input and response to a minibatchqueue. You can do this by creating arrayDatastore objects and combining these into a single datastore using the combine function. Form the minibatchqueue with this combined datastore object.

xds = arrayDatastore(xTrain);
tds = arrayDatastore(tTrain);
cds = combine(xds,tds);

mbqTrain = minibatchqueue(cds,2,...
    "MiniBatchSize",numSamples,...
    "OutputAsDlarray",[1 1],...
    "OutputEnvironment","cpu",...
    "MiniBatchFormat",["BC","BC"]);

Create the 2-dimensional FMNN

As discussed in AI Verification: Monotonicity, fully monotonic neural networks adhere to a specific class of neural network architectures with constraints applied to weights. In this proof of concept example, you build a simple FMNN using fully connected layers and fullsort for the gradient norm preserving activation functions. You specify a ResidualScaling=2 and pNorm=1 that balances monotonic growth with smoothness of solution and specify that the MonotonicTrend is "increasing" as it is increasing with respect to all input channels. For more information on the architectural construction and hyperparameter selections, see AI Verification: Monotonicity.

inputSize = 2;
numHiddenUnits = [16 8 4 1];
pNorm = 1;
rng(0);
fmnnet = buildConstrainedNetwork("fully-monotonic",inputSize,numHiddenUnits,...
    Activation="fullsort",...
    ResidualScaling=2,...
    MonotonicTrend="increasing",...
    pNorm=pNorm)
fmnnet = 
  dlnetwork with properties:

         Layers: [11x1 nnet.cnn.layer.Layer]
    Connections: [11x2 table]
     Learnables: [8x3 table]
          State: [0x3 table]
     InputNames: {'input'}
    OutputNames: {'addition'}
    Initialized: 1

  View summary with summary.

You can view the network architecture in deepNetworkDesigner by setting the viewNetworkDND flag to true. Otherwise, plot the network graph.

viewNetworkDND = false;
if viewNetworkDND
    deepNetworkDesigner(fmnnet) %#ok<UNRCH>
else
    figure;
    plot(fmnnet)
end

Train the FMNN

First, create a custom training options struct. For the trainMonotonicNetwork function, you can specify four hyperparameters: maxEpochs, initialLearnRate, decay, lossMetric. Additionally, if you specify a pNorm in buildConstrainedNetwork, you must also specify the same pNorm for training.

maxEpochs = 800;
initialLearnRate = 0.1;
decay = 0.05;
lossMetric = "mse";
pNorm = 1;

Train the network with these options.

trained_fmnnet = trainConstrainedNetwork("fully-monotonic",fmnnet,mbqTrain,...
   MaxEpochs=maxEpochs,...
   InitialLearnRate=initialLearnRate,...
   Decay=decay,...
   LossMetric=lossMetric,...
   pNorm=pNorm);

Network Evaluation

Examine the shape of the solution by sampling the training data and network prediction.

yPred = predict(trained_fmnnet,xTrain);

figure;
plot3(xTrain(:,1),xTrain(:,2),tTrain,"k.")
hold on
plot3(xTrain(:,1),xTrain(:,2),yPred,"r.")
xlabel("x1")
ylabel("x2")
legend("Training Data","Network Prediction",Location="northwest")

Guaranteed Monotonicity for FMNN

As discussed in AI Verification: Monotonicity, fully monotonic neural networks are monotonic in each output with respect to every input. In the n-dimensional case, you can see monotonicity through any 1-dimension restriction through the space.

Take lines passing through the origin. Specify the angle line, projected onto the x1 and x2 plane, makes with the x1 axis.

thetaDegrees = 42;

Sample the network prediction along that line and plot this 1-dimensional restriction, along with the directional derivative.

xInterval = -1:0.1:1;
n = [cosd(thetaDegrees); sind(thetaDegrees)]; % Unit normal
xSample = n.*xInterval;

[fmnnPred,dfmnnPred] = dlfeval(@computeFunctionAndDerivativeForScalarOutput,trained_fmnnet,dlarray(xSample,"CB"));
ddfmnnPred = n'*extractdata(dfmnnPred);

figure;
plot(xInterval,xSample(1,:).^3+xSample(2,:).^3,"k-.")
hold on
plot(xInterval,fmnnPred,"r-.")
plot(xInterval,ddfmnnPred,"b-.")
yline(0,"k--")
xlabel("xSlice")
legend("Ground Truth","Network Prediction","Network Gradient")

As is evident in the figure, monotonicity is given along any 1-D restriction through the surface. This is enforced by the networks architecture and constraints on the weights.

Helper Functions

function [Z,dZ] = computeFunctionAndDerivativeForScalarOutput(net,X)
% Evaluate f
Z = predict(net,X);
% Evaluate df/dx. Since Z_i depends only on X_i, the derivative of d/dX_i
% sum(Z_i) = d/dX_i Z_i
dZ = dlgradient(sum(Z,2),X);
end

Copyright 2024 The MathWorks, Inc.