New in FANN 2.1.0 (Mar 13, 2013)
- Added fann_create_train_from_callback for creating training data from a callback function
- Amended fann_create_train_from_callback for dynamic link library compatibility
- Added fann_get_activation_function and fann_get_activation_steepness matching the set functions
- Added a set of functions to allow restricted access to neural network layout, connections, and weights.
- Added FANN_SIN, FANN_COS, FANN_SIN_SYMMETRIC and FANN_COS_SYMMETRIC activation functions
- Changed the create_array functions to indicate that the layers array parameter is const i.e. input only
- Added FANN_API calling convention to fann_get_bit_fail and two cascade functions for dll compatibility
- Added FANN_EXTERNAL/FANN_API calling convention to previous extension functions for dll compatibility
- Removed warning from fann_print_connections with explicit casts
- Added fann_set/get_user_data to allow user programs to refer to their own data during callbacks
- Added documented C++ wrapper and sample C++ program
- Fixed bias count for shortcut networks in fann_get_layer_array and fann_get_bias_array
- Minor fixes of FANN_EXTERNAL/FANN_API for windows dll compatibility
- More advanced scaling functionality
- Added advanced scaling functionality to fann_cpp.h
- Additional documentation and fix of FANN_EXTERNAL for create train callback in fann_cpp.h
- Fixed problems with FANN_GAUSSIAN and FANN_GAUSSIAN_SYMMETRIC activation functions
- Added semi Widrow + Nguyen to initialization of cascade weights
- Fixed problem with RPROP, which makes it go suboptimal in the beginning
- Changed definition of MSE, so that it calculates a value pr output, and not a sum for all outputs. (MSE = MSE/ann->num_output)
- Added more benchmark data
- Added possebility to make benchmark as average of several
- Changed so that RPROP and Quickprop may not alter a weight to more than 1500