##Purpose This library is a general purpose Artificial Neural Network implementation in MATLAB focused on research purposes. It is designed to have a large number of often experimental features that can easily be enabled/disabled without bogging the user down in details they aren't immediately concerned with.

Last Update: 11 Mar 16
Version: 0.1.1 alpha release (stable, but not thoroughly tested)
Requirements: Matlab parallel processing toolkit, tested on R2015b
Optional: Matlab computer vision toolkit (some optional functions)

###Basic Usage: Model parameters define all hyper parameters and settings the algorithm needs. They are stored in a struct and should be generated using initialize_model. Data sets are prepared in a similar data structure created using initialize_data. A trained model, complete with metrics, trained weights, etc. is generated by calling train_ann with a model or a cell array of many models, and a data set or cell array of many datasets.

    model = initialize_model(...);
    data  = initialize_data(...);
    trained_model = train_ann(model, data);
	predictions   = predict_ann(trained_model, data);

initialize_model requires 1 parameter, everything else is optional and are defined using matlab's standard ('name',value) parameterization. Parameters for initialize_model are documented in detail later in this guide.

initialize_data requires two parameters, the feature data in NxD format (N samples by D features), and label data in NxC (N samples by C output neurons). Cross validation feature and labels, and options to normalize the data are optional and documented later in this guide.

train_ann accepts one or more models and one or more data sets and provides one or more trained_models as output. If you provide multiple models or data sets to be trained (in parallel) provide them in a cell array (see example usage). Note that you may not provide multiple models AND multiple data sets, one of the two must be single.

This function returns the same model that was created with initialize_model along with a Metrics structure added to it and weights set. This is meant to allow for easy storage of the model and parameters that were used to generate the model.

predict_ann provides predictions based on a trained model. Note that if you provide a cross validation data set this will occur automatically and the Metrics produced will reflect the results, it is common not to use this function during experimentation.

###Example usage:

% Initialize model parameters to a 3 layer network with 784 feature inputs, 
% a single 30 neuron hidden layer, and 10 output classes.
model = initialize_model( [784 30 10], ...
                          'cost_function', 'cross_entropy', ...
                          'regularization', 'L2' );

% Initialize a dataset of 1000 MNIST digits and 1000 cross validation digits
% This example assumes mnist_features and mnist_labels was already created
% by the user and contains a random set of features and labels.
data = initialize_data( mnist_features(1:1000,:), ...
                        mnist_labels(1:1000,:), ...
                        'cv_features', mnist_features(1001:2000,:), ...
                        'cv_labels',   mnist_labels(1001:2000,:), ...
                        'rescale_method', 'standardize' );

% Train the model using the dataset defined above
trained_model = train_ann(model, data);

% Plot the results
plot_n_runs_accuracy_cost( trained_model );

###initialize_model parameter options

Parameters	Value	Default	Description
layer_sizes	[784 30 30 10]	Required	Defines the network size, the example shown in the Value column creates a network of 784 inputs, two hidden layers of 30 neurons, and 10 output neurons.
'plot_title'	Text	None	The text used in the plot legend (and possibly elsewhere that a model name is needed).
'update_method'	'GD' or 'EG+-'	'GD'	The update method to use, Gradient Descent and EG+- are options (Exponentiated Gradient [2])
'U'	Numeric or 'unnormalized'	40	The regularization term used by EG+-, unused if EG+- isn't chosen as the update method. For unnormalized EG+- use 'unnormalized' instead of a numeric value. [2]
'EG_sharing'	'none', 'past_average', 'decaying_past'	'none'	Sharing method implemented by EG+/- update algorithm. When decaying_past is used be sure to include a parameter 'EG_sharing_decay_q' used to define how the decay rate is calculated.
'EG_sharing_alpha'	Numeric	0.1	Alpha parameter for EG_sharing, the percentage of weight removed from the past update and shared among previous updates
'EG_sharing_decay_q'	Numeric	0.2	Used with sharing with decaying_past, determines the rate at which decayed sharing is applied. See: https://users.soe.ucsc.edu/~manfred/pubs/J52talk.pdf
'EG_sharing_inc_biases'	true/false	false	Determines if sharing will share past biases or not (e.g. only past weights are shared).
'learning_rate'	Numeric	depends on cost function	The 'Eta' learning rate, default value set dynamically depending on the cost function chosen. 3.0 for quadratic cost and 0.5 default for cross entropy.
'num_epochs'	Numeric	30	Number of epochs of the training set to run through. Note that this parameter will change in future versions.
'mini_batch_size'	Numeric	100	Number of samples to compute in each mini batch. 1 equals stochastic, choosing a value equal to the # of epochs is equal to full batch gradient descent.
'cost_function'	'sum_of_squares', 'cross_entropy', 'loglikelihood'	'cross_entropy'	The cost function to use.
'initial_weights'	Cell array of weights	Auto initialized	The weights are initialized to random values based on the 'weight_init_method'. You can specify them manually.
'initial_biases'	Cell array of biases	Auto initialized	Bias units, works the same as initial_weights
'weight_init_method'	'1/sqrt(n)', 'gaussian-0-mean-1-std'	'1/sqrt(n)'	The method used to initialize the weights and biases.
'regularization'	'none', 'L1', 'L2'	'L2'	The regularization method to use. Note this is only used by Gradient Descent updates, not EG+-.
'lambda'	Numeric	1.0e-4	Degree of regularization to apply. Note that this is a fixed percent value, some texts, notably NeuralNetworksAndDeepLearning.com suggest to use lambda/n, but this causes an unnecessary dependence on the size of the input dataset.
'monitor_training_cost'	true/false	true	Enables generation of training data accuracy and cost function data Metrics.
'monitor_accuracy'	true/false	true	Enables generation of cross validation accuracy metrics per epoch of training.
'jl_projection'	array	[]	Defines the size that each layer will be projected down to using Johnson–Lindenstrauss projections. -1 defines no projection at that layer. Example: [-1 200 100 -1] would be a 4 layer network with projections leaving layer 2 (down to 200) and leaving layer 3 (down to 100), no projections on the input layer or output layer (there is no reason to ever project on the output layer).
'store_weight_history'	true/false	false	Stores a flattened list of all weights & biases for each minibatch in model.Metrics.weight_history. Warning, this can be a huge amount of data. Each column of the matrix contains all weights/biases from the full network. A column can be converted back into matrix form using the utility function: unflatten_weights_biases(trained_model, trained_model.Metrics.weight_history(:,i)) where i is the desired mini-batch iteration.
'verbosity'	'none', 'epoch', 'debug'	'epoch'	The amount of console output to write during single training runs.
'rng_seed'	Numeric, 'shuffle'	'shuffle'	A random seed to generate deterministic results.
'noise_function'	@function_handle, 'none'	'none'	A function that will take feature input and transform it. The noise function must take an MxD matrix of features and return a transformed MxD matrix of features. It will be called on each epoch to transform the data before processing.
'debug_check_numerical_gradients'	true/false	false	Runs a numerical gradient check at each mini batch. This is really only for debugging purposes, for example, when adding a new cost function. If verbose='debug', per-weight/bias output will be written to the console. Also note that inputs should be double precision. Numerical gradient checks must be within 1e-9 to pass. An exception will be thrown if the gradient check fails.

###initialize_data parameter options

Parameters	Value	Default	Description
feature_data	MxD matrix	Required	An MxD matrix of feature data where M is the number of training samples and D is the number of features.
labels	MxC matrix	Required	The labels for the feature_data in N samples by C output neurons format
'cv_features'	MxD matrix	N/A	A matrix of feature data to use as an unseen cross validation dataset (not necessarily the same size as the feature data). Same format as feature_data.
'cv_labels'	MxC matrix	N/A	If cv_features was specified this is required, same format as labels
'rescale_method'	'none', 'standardize', 'rescale'	'none'	The method used to normalize the input features. Note that if you do not normalize the input features but have a high variance in the inputs you will receive a warning.

###train_ann This function is generally used without parameters, but a few runtime parameters are available for certain special cases.

Parameters	Value	Default	Description
workers	Integer	CPU Cores + 1	Specifies the number of parallel pool workers that will be used. Defaults to one more than the number of cpu cores. Cannot use more workers than are created by the parallel pool.
save	'path/filename'	N/A	Exports training results to disk instead of returning them from the function. This will export each training as it completes and remove it from memory. train_ann will return empty results instead of trained models.

###Plotting Results Plot functions assume a cell array of 1 or more trained models. Plot functions available: plot_n_run_accuracy_cost( trained_models ) Plots training accuracy, cross validation accuracy, and traing cost for each trained model in the cell array passed to it.

Note that figure legends will use the model parameter 'plot_title' from each model.

Parameters	Value	Default	Description
trained_models	Result from train_ann	Required	A cell array of trained models, this is the standard returned value from train_ann
'plot_training_accuracy'	true/false	true	Plot the accuracy on the training data as a function of the number of epochs
'plot_cv_accuracy'	true/false	true	Plot the accuracy on the unseen cross validation data as a function of the number of epochs on the same plot
'plot_training_cost'	true/false	true	Plot the training cost generated aggregated by training epoch

Example usage:

% Plot the Training and Cross validation accuracy only.
plot_n_run_accuracy_cost( trained_models, 'plot_cost_function', false );

###Metrics If metrics generation wasn't disabled in the model the trained model will contain a new struct called Metrics. It will contain the following data:

• training_cost per-mini batch cost function results
• training_accuracy per-epoch accuracy on the training data
• cv_accuracy per-epoch accuracy on the optional cross validation data

###Planned future enhancements:

• Built in auto encoder
• Callback capabilities to manipulate the mechanics of the network each mini batch
• Also will be used for dynamically generating plots and interactively altering parameters such as regularization during network training for real-time analysis during training.
• Support for running on the GPU (implemented but untested/nonfunctional)
• Regression

###Known issues

• Currently only classification is supported.

###Licensing and Contact This package is maintained by David Parks, UCSC -- Contact: [email protected]

The MIT License (MIT)

Copyright (c) [2016] [David Freeman Parks - [email protected]]

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

###References [1] Michael Nielsen, "Neural Networks and Deep Learning book", [online] www.neuralnetworksanddeeplearning.com
[2] Manfred Warmuth and Jyrki Kivinen, "Exponentiated Gradient versus Gradient Descent for Linear Predictors" (1997), [online] https://users.soe.ucsc.edu/~manfred/pubs/J36.pdf
[3] Andrew Ng, Jiquan Ngiam, Chuan Yu Foo, Yifan Mai, Caroline Suen, "Stanford UFLDL Tutorial on Machine Learning", [online] http://ufldl.stanford.edu/wiki/index.php/UFLDL_Tutorial