Tutorials

This page gives a gentle introduction into Shark. The quick tutorial section gives an introduction into the most important core components. For neural network training, following the neural entwork tutorials is very helpful for a first step. If you are already familiar with the Shark architecture, the documentation of the key concepts and list of classes can be found below:

Concept	Class List
Models	List
Loss and Cost Functions	List
Optimizers	Gradient-Based Optimization, Direct-Search Optimizers, Multi-Objective Optimizers
Objective Functions	List
Trainers	Supervised Trainers, Unsupervised Trainers
Kernels	List

Quick tutorial

In case you are new to Shark, we give you a quick tour over the core components. We first show how to set up either a traditional Makefile or a CMake file for your application program. Then we move on to a simple Hello-World example of what linear binary classification can look like in Shark. The third tutorial illustrates the model-error-optimizer trias often encountered in Shark through a simple regression task.

• Your Shark Programs

• Hello World

• General Optimization Tasks

Neural Networks

A very important class of machine-learning models are Neural Networks. This section discusses the creation and training of multi-layer neural networks

• Training Feed-Forward Neural Networks

• Data Shapes and Deep Convolutional Architectures

• Autoencoders

• Variational Autoencoders

Data Handling

Since many machine learning algorithms work on real-world datasets, we extensively cover Shark’s Data class as well as common operations on them:

• Data Containers

• Label Formats

• Importing Data

• Creating and Using Subsets of Data

• Normalization of Input Data

Specific Machine-Learning Algorithms

Here come tutorials for some selected algorithms implemented in Shark. It must be said that this is only the tip of the iceberg, many more machine learning algorithms and tools are provided by the library.

Let’s start with some classical methods:

• Principal Component Analysis

• Nearest Neighbor Classification

• Linear Discriminant Analysis

• Linear Regression

• LASSO Regression

• K-Means Clustering

Tree-based algorithms:

• Random Forest

Kernel methods – support vector machine training and model selection:

• Support Vector Machines: First Steps

• Support Vector Machines: Model Selection Using Cross-Validation and Grid-Search

• Support Vector Machines: Likelihood-based Model Selection

• Linear Kernel Combinations (and a bit of MKL)

• Linear Support Vector Machines

• Kernel Target Alignment

• Kernelized Budgeted Stochastic Gradient Descent

Optimization:Direct-Search

Shark offers many direct-search algorithms. The most important one is the CMA-ES in the single and multi-objective variants

• Direct Search Algorithms

• Covariance Matrix Adaptation Evolution Strategy

• Multi-Objective Covariance Matrix Adaptation Evolution Strategy

• Creating Multi-Objective Benchmarks with Shark

Restricted Boltzman Machines

• The RBM Module

• Training Binary Restricted Boltzmann Machines

Tools

Finally, we present functionality which are not machine learning facilities themselves, but necessary or helpful tools.

For convenience, Shark provides a statistics class wrapper, as well as generic support for serialization:

• Iterative Calculation of Statistics

• Serialization

For Shark developers

Note that Shark follows a

• Coding Convention.

If you contribute to Shark, you might also find these documents helpful:

• Adding Examples, Tests and Documentation

• Shark Conventions for Derivatives

• Data Batches