Lightweight Neural Network++

File Release Notes and Changelog

Release Name: 0.97

Notes: New version with possibility to do make install and including ocr example by Peter Van Rossum

---

Changes: version 0.97 -- 8/3/2004 - Added possibility to derive a new class from trainer - New Makefile with dynamic libraries and possibility to do make install version 0.95 -- 3 March 2004 - Improved trainer class - Class iomanage for reading target pairs - Visual Network Trainer improved and draws graphics with gnuplot version 0.9 -- 27 February 2004 - The new class trainer makes very easy the training - New directory tests and new examples (wine) - Change of default max random values of the weights - In directory tests "visualtrainer", a graphical interface (in tcl/tk) for training networks. version 0.88 -- 22 February 2004 - New constructor taking a vector<int> of numbers of neurons in layers - Sigmoidal function now multiplies by invinterval instead of dividing by interval - Sigmoidal function approximates the bounds with 0, 1 (-1, 1) and not by the min or max value in table. - Fixed the destructor (now deletes the sumdeltas fields) - Now reading from file is smart enough to import files in old lwneuralnet (C-library) format version 0.85 -- - Fixed the copy() method for the copy constructor - generate_sigma with array min_entry e max_entry - Added possibility to switch min value and max value of tanh function in generate_sigma (options -t, -u) version 0.8 -- 8 February 2004 - Using inv_no_of_patterns in batch training - Added possibility to combine SuperSab and batch training with the end_batch_ssab() method. Seems to be very fast! - control that nus are in the range [minnu maxnu ] when these values are changed - method ssab_stats() version 0.7 -- 7 February 2004 - Bug fixed in batch training - Added possibility to change values of parameters in SuperSab training - Some bugs corrected in tests directory version 0.61 - More documentation version 0.6 - First release