Implements the Tsetlin Machine, Coalesced Tsetlin Machine, Convolutional Tsetlin Machine, Regression Tsetlin Machine, and Weighted Tsetlin Machine, with support for continuous features, drop clause, Type III Feedback, focused negative sampling, multi-task classifier, autoencoder, literal budget, and one-vs-one multi-class classifier. TMU is written in Python with wrappers for C and CUDA-based clause evaluation and updating.
During testing of a hardware implementation of a Convolutional Coalesced TM – for MNIST – I found that setting a maximum/minimum value of the signed weights could give a boost in test accuracy, compared to training with no weight restrictions. In hardware such a limitation is required as a fixed number of bits per weight has to be allocated.
Typically, the weight limit would be rather small compared to the number of clauses and T, and this seems to be beneficial for the performance. It might affect the training time though.
It would be very useful and interesting if such a parameter could be included.
During testing of a hardware implementation of a Convolutional Coalesced TM – for MNIST – I found that setting a maximum/minimum value of the signed weights could give a boost in test accuracy, compared to training with no weight restrictions. In hardware such a limitation is required as a fixed number of bits per weight has to be allocated.
Typically, the weight limit would be rather small compared to the number of clauses and T, and this seems to be beneficial for the performance. It might affect the training time though.
It would be very useful and interesting if such a parameter could be included.