This is a new feature and cleanup release. Features: Added GEMPLP for approximate inference to the structured output framework [Jiaolong Xu]. Effeciency improvements of the FITC framework for GP inference (FITC_Laplce, FITC, VarDTC) [Wu Lin]. Added optimisation of inducing variables in sparse GP inference [Wu Lin]. Added optimisation methods for GP inference (Newton, Cholesky, LBFGS, ...) [Wu Lin]. Added Automatic Relevance Determination (ARD) kernel functionality for variational GP inference [Wu Lin]. Updated Notebook for variational GP inference [Wu Lin]. New framework for stochastic optimisation (L1/2 loss, mirror descent, proximal gradients, adagrad, SVRG, RMSProp, adadelta, ...) [Wu Lin]. New Shogun meta-language for automatically generating code listings in all target languages [Esben Sörig]. Added periodic kernel [Esben Sörig]. Add gradient output functionality in Neural Nets [Sanuj Sharma]. Bugfixes: Fixes for java_modular build using OpenJDK [Björn Esser]. Catch uncaught exceptions in Neural Net code [Khaled Nasr]. Fix build of modular interfaces with SWIG 3.0.5 on MacOSX [Björn Esser]. Fix segfaults when calling delete[] twice on SGMatrix-instances [Björn Esser]. Fix for building with full-hardening-(CXX|LD)FLAGS [Björn Esser]. Patch SWIG to fix a problem with SWIG and Python >= 3.5 [Björn Esser]. Add modshogun.rb: make sure narray is loaded before modshogun.so [Björn Esser]. set working-dir properly when running R (#2654) [Björn Esser]. Cleanup, efficiency updates, and API Changes: Added GPU based dot-products to linalg [Rahul De]. Added scale methods to linalg [Rahul De]. Added element wise products to linalg [Rahul De]. Added element-wise unary operators in linalg [Rahul De]. Dropped parameter migration framework [Heiko Strathmann]. Disabled Python integration tests by default [Sergey Lisitsyn, Heiko Strathmann].