November  2014
Data Reading:          1. Faster PFiles preparation. A cross-validation set (~5%) is always selected. This eliminates the time-consuming steps  for         PFile concatenating and splitting          2. PDNN supports specification of a list of PFiles, for example, train.pfile.*.gz and train.pfile.[1-10].gz. No need to do PFile         concatenation.          3. PFiles are always compressed. This reduces PFiles to 1/10 of the original size          4. Supports Python pickle files. Refer to examples/mnist to see how to prepare and specify pickle files. Recipes & Architecture:          1. PDNN supports multi-task learning, and this enables DNN training over multiple languages, domains, dialects, etc.          2. Optimized CNN recipes; the CNN architecture is modifed to be the IBM style          3. Maxout network models are added          4. run_rm and run_hkust are removed (not supported anymore)          5. run_tedlium is added Results:          1. Updated results on TIMIT
April 2014
More recipes, e.g., run-dnn-fbank+pitch.sh      2D (time x frequency) convolution; a faster version of CNN with the cuda-convnet  wrapper      Scripts are simplified, verified and now become more readable      For different datasets and with benchmark results