Popularity
1.9
Stable
Activity
0.0
Stable
17
11
7

Description

You can use Infer.NET to solve many different kinds of machine learning problems, from standard problems like classification, recommendation or clustering through to customised solutions to domain-specific problems. Infer.NET has been used in a wide variety of domains including information retrieval, bioinformatics, epidemiology, vision, and many others.

Infer.NET is published as open source on GitHub under the MIT license and also available as NuGet packages.

Programming language: MATLAB
License: MIT License

Infer.NET alternatives and similar packages

Based on the "Machine Learning and Data Science" category.
Alternatively, view Infer.NET alternatives based on common mentions on social networks and blogs.

Do you think we are missing an alternative of Infer.NET or a related project?

Add another 'Machine Learning and Data Science' Package

README

KJIT

The goal of this project is to learn a kernel-based message operator which takes as input all incoming messages to a factor and produces a projected outgoing expectation propagation (EP) message. In ordinary EP, computing an outgoing message may involve solving a difficult integral for minimizing the KL divergence between the tilted distribution and the approximate posterior. Such operator allows one to bypass the computation of the integral by directly mapping all incoming messages into an outgoing message. Learning of such an operator is done online during EP. The operator is termed KJIT for Kernel-based Just-In-Time learning for passing EP messages.

Full details are in our UAI 2015 paper. Supplementary material is here.

Wittawat Jitkrittum, Arthur Gretton, Nicolas Heess, 
S. M. Ali Eslami, Balaji Lakshminarayanan, Dino Sejdinovic, and Zoltán Szabó
"Kernel-Based Just-In-Time Learning for Passing Expectation Propagation Messages"
UAI, 2015

This project extends

Nicolas Heess, Daniel Tarlow, and John Winn. 
“Learning to Pass Expectation Propagation Messages.” 
NIPS, 2013.
http://media.nips.cc/nipsbooks/nipspapers/paper_files/nips26/1493.pdf.

and

S. M. Ali Eslami, Daniel Tarlow, Pushmeet Kohli, and John Winn
"Just-In-Time Learning for Fast and Flexible Inference." 
NIPS, 2014.
http://papers.nips.cc/paper/5595-just-in-time-learning-for-fast-and-flexible-inference.pdf

License

KJIT software is under MIT license.

The KJIT software relies on Infer.NET (freely available for non-commercial use) which is not included in our software. Even though the license of KJIT software is permissive, Infer.NET's license is not. Please refer to its license for details.

Repository structure

The repository contains

  1. Matlab code for experimenting in a batch learning setting. Experiments on new kernels, factors, random features, message operators are all done in Matlab in the first stage. Once the methods are developed, they are reimplemented in C# to be operable in Infer.NET framework. EP inference is implemented in C# using Infer.NET, not in Matlab. All Matlab code is in the code folder.
  2. C# code for message operators in Infer.NET framework. The code for this part is in code/KernelEP.NET which contains a C# project developed with Monodevelop (free cross-platform IDE) on Ubuntu 14.04. You should be able to use Visual studio in Windows to open the project file if it is more preferable.

All the code is written in Matlab and C# and expected to be cross-platform.

Include Infer.NET

The Matlab part of this project does not depend on the Infer.NET package. However, to use our KJIT message operator in the Infer.NET framework, you have to include Infer.NET package by taking the following steps.

  1. Download Infer.NET package from its Microsoft research page. Upon extracting the zip archive, you will see subfolders including Bin, Source, and its license. Carefully read its license.
  2. Copy Infer.Compiler.dll and Infer.Runtime.dll from the Bin folder of the extracted archive into code/KernelEP.NET/lib/Infer.NET/Bin/ of this repository. Without this step, when you open the project in Monodevelop, it will not compile due to the missing dependency.
  3. Try to build the project. There should be no errors.

Useful submodules

In the development of the code for learning an EP message operator, some commonly used functions are reimplemented to better suit the need of this project. These functions might be useful for other works. These include

  • Incomplete Cholesky factorization. This is implemented in Matlab in such a way that any kernel and any type of data (not necessarily points from Euclidean space) can be used. The full kernel matrix is not pre-loaded. Only one row of the kernel matrix is computed at a time, allowing a large kernel matrix to be factorized. In this project, points are distributions and the kernel takes two distributions as input. See IncompChol.

  • Dynamic matrix in Matlab. This is a matrix whose entries are given by a function f: (I, J) -> M where I, J are index list and M is a submatrix specified by I, J. The dynamic matrix is useful when the underlying matrix is too large to fit into memory but entries can be computed on the fly when needed. In this project, this object is used to represent the data matrix when a large number of random features are used. Multiplication (to a regular matrix or a dynamic matrix) operations are implemented. See DynamicMatrix and DefaultDynamicMatrix.

Code usage

Please feel free to contact me (see wittawat.com) regarding code usage. For fun, visualization of this repository is available here.


*Note that all licence references and agreements mentioned in the Infer.NET README section above are relevant to that project's source code only.