bdschrisk
commented
8 years ago - Added parallised multi-class classification / learner with support for unlimited classes and any generator
- Added extension methods for computing min, max, std dev and mean for vectors / matrices
- Changed base model and generator implementation to include proprocessing steps (recommended for most ML algorithms)
- Changed feature normalisation methods to allow for greater flexibility and easier extension. Also added a number of new feature normalizers.
- Restructured cost functions and optimisation - including new optimiser system
- Fixed bug in Logistic Regression where the prediction was failing to compute properly
- Fixed bug in Logistic Regression where using Descriptor method would cause label vector to output 1/-1 which caused learning to fail
- Added new optimisation functionality with new optimisation approach with support for custom optimisers, including Nesterov-Accelerated gradient descent and stochastic gradient descent with momentum
- Added new Functions (in Math) including Softmax, SteepLogistic, RectifiedLinear and ClippedRectifiedLinear functions
- Added new Kernel functions for Sigmoid and Linear kernels
- Added a lightweight Support Vector Machine with SMO utilising pair selection heuristics (Working Set Selection 3) as used in LIBSVM
- Added Collaborative Filtering (Recommender) model with generator including unit tests with gradient checking
- Added new extension methods to Matrix and Vectors for shape changing, selection and other operations
- Added Scoring functionality and new model selection metrics in Learner class
- Started added deep learning model building functionality - including support of autoencoder layers (not finished)
- Started implementing Recurrent Neural Net based on Cho et al 2014 Gated Recurrent Unit LSTM variant