search

kashefy / elm

The ELM Library Project, a framework for building learning machines.
Other
4 stars 0 forks source link

readme

The ELM framework

Build Status Build status Coverage Status License

This started off as a a C++ implementation of a Spike-based Expectation Maximization (SEM) model, a biologically inspired model for object detection and recognition. It then grew into a more generic framework called ELM to support experimenting with as many models as possible, providing an interface for adding custom algorithms and methods. Whoever said coding in C++ couldn't be done as smoothly as scripting languages?

Background

Nessler, Pfeiffer and Maass originally propose the SEM model in 2010. They demonstrate how a spiking neuronal network governed by spike-timing-dependent-plasticity (STDP) and a stochastic winner-take-all (WTA) can learn and predict hidden causes through unsupervised learning from visual input.

This implementation is work by Youssef Kashef. A generalized framework is layed out in order to facilitate building and experimenting with models. The uses of such models could be for Computer Vision applications or solving a good old-fashioned Machine Learning problem.

An extended SEM model is used to showcase this framework. It extends the original SEM with hierarchical learning of features and selective visual attention. This work was as part of a Masters Thesis at ETH Zürich titled "Scale- and Translation-Invariant Unsupervised Learning of Hidden Causes Using Spiking Neurons with Selective Visual Attention" The extended SEM and its specific layers are detached from the generic ELM framework but continue subscribing to it.

Milestones

  1. Matlab implementation by Michael Pfeiffer and authors of 2010 SEM paper (not included).
  2. Java implementation + Evaluation in Matlab
  3. SEM model extended with scale and translation invariance through hiearchical learning
  4. C++ implementation and code refactoring in progress (partially complete)
  5. Separate SEM specific layers, hello ELM framework
  6. Learning layers and module

Anticipated milestones:

Getting your framework build up and running

The build is IDE agnostic and platform independent. After setting up all required dependencies, you configure the build for your toolchain/IDE of choice and build it. Required tools:

Dependencies:

You'll need to set up the following dependencies.You can build them from source or install them via your platforms package management software (e.g. Ubuntu's apt-get, Fedora's yum)

Configure and Build:

We use cmake for configuring the build. This allows building for a variety of toolchains and IDEs and keeping the code independent of such.

Example cmake configuration command if dependencies are install in the usual system paths (e.g. /usr/lib):

cmake $SRC_DIR/elm
make

Example cmake configuration command for when dependencies are in custom paths, or if you want to link against dependencies other than those installed under the usual system paths:

cmake -DGTEST_ROOT=<gtest build dir> -DOpenCV_DIR=<opencv build dir> -DBOOST_ROOT=<boost build dir> -DBoost_NO_SYSTEM_PATHS=ON $SRC_DIR/elm

cmake -DGTEST_ROOT=~/src/gtest-1.7.0 -DOpenCV_DIR=~/build/opencv/share/ -DBOOST_ROOT=~/build/boost/boost_1_57_0/ -DBoost_NO_SYSTEM_PATHS=ON $SRC_DIR/elm

Build issues

Linking error when building in Ubuntu: The error message looks something like ...undefined reference to symbol 'pthread_key_delete@@GLIBC_2.2.5'... ...'pthread_key_delete@@GLIBC_2.2.5' is defined in DSO /lib/x86_64-linux-gnu/libpthread.so.0 so try adding it to the linker command line...

Resolve by adding -pthread or -lpthread to the linker. You can do so by reconifuring CMake and adding the option: -DCMAKE_EXE_LINKER_FLAGS="-pthread"

Contribution guidelines

How do I generate the documentation?

cd doxygen ./docs/doxygenConf*

Whom do I talk to?