Creating a Model

Build a network whose input count matches the feature columns and whose output layer, activation, and loss match the prediction task.

Regression

Use a linear output for predicting a continuous numerical value:

import deepnetts.net.FeedForwardNetwork;
import deepnetts.net.layers.activation.ActivationType;
import deepnetts.net.loss.LossType;

FeedForwardNetwork neuralNet = FeedForwardNetwork.builder()
        .addInputLayer(NUM_INPUTS)
        .addOutputLayer(1, ActivationType.LINEAR)
        .lossFunction(LossType.MEAN_SQUARED_ERROR)
        .randomSeed(42)
        .build();

This is the smallest linear regression model. Add hidden layers only when the problem requires a nonlinear relationship.

Binary classification

Use one Sigmoid output to produce a value between zero and one:

FeedForwardNetwork neuralNet = FeedForwardNetwork.builder()
        .addInputLayer(NUM_INPUTS)
        .addFullyConnectedLayer(32)
        .addFullyConnectedLayer(16)
        .addOutputLayer(1, ActivationType.SIGMOID)
        .lossFunction(LossType.CROSS_ENTROPY)
        .randomSeed(42)
        .build();

Interpret the output as a class probability and apply an explicit threshold when choosing the class.

Multiclass classification

Use one Softmax output per class:

FeedForwardNetwork neuralNet = FeedForwardNetwork.builder()
        .addInputLayer(NUM_INPUTS)
        .addFullyConnectedLayer(16, ActivationType.TANH)
        .addOutputLayer(NUM_CLASSES, ActivationType.SOFTMAX)
        .lossFunction(LossType.CROSS_ENTROPY)
        .randomSeed(42)
        .build();

The order of the outputs must match the order of the one-hot target columns and the class-name mapping used during prediction.

Convolutional Neural Network

Use ConvolutionalNetwork when the input has spatial structure, such as an image. Convolutional layers learn local patterns, pooling reduces spatial dimensions, and fully connected layers combine the learned feature maps before classification.

A compact binary image classifier can be built with the current Deep Netts API:

import deepnetts.net.ConvolutionalNetwork;
import deepnetts.net.layers.Filters;
import deepnetts.net.layers.activation.ActivationType;
import deepnetts.net.loss.LossType;

ConvolutionalNetwork network = ConvolutionalNetwork.builder()
        .addInputLayer(width, height, 3)
        .addConvolutionalLayer(6, Filters.ofSize(3))
        .addMaxPoolingLayer(2, 2)
        .addFullyConnectedLayer(16)
        .addOutputLayer(1, ActivationType.SIGMOID)
        .hiddenActivationFunction(ActivationType.LEAKY_RELU)
        .lossFunction(LossType.CROSS_ENTROPY)
        .randomSeed(42)
        .build();

For grayscale input, use one channel. For multiclass classification, use one Softmax output per class:

.addInputLayer(width, height, 1)
// convolution, pooling, and fully connected layers
.addOutputLayer(NUM_CLASSES, ActivationType.SOFTMAX)
.lossFunction(LossType.CROSS_ENTROPY)

The output order must remain identical to the class-label mapping used by the dataset and prediction code. Add another convolution-and-pooling block only when the image size and problem complexity justify the additional capacity.

Check the dimensions

• Input neurons must equal the number of feature columns.
• Regression usually has one linear output per numerical target.
• Binary classification uses one Sigmoid output in the current examples.
• Multiclass classification uses one Softmax output per class.
• A CNN input must match image width, height, and channel count exactly.
• A binary CNN uses one Sigmoid output; a multiclass CNN uses one Softmax output per class.

Use a fixed random seed when you want reproducible weight initialization while comparing changes.