hi every one,
as we know All machine learning algorithms are made up of two parts test and train, but I saw in HTM training videos that this algorithm did not work like other learning machines algorithms(i mean it does not test and train part). My question is then how does it work? In addition, in the code that I see, the MNIST dataset was divided into two categories of test and train. I was confused. please give me more details on how the algorithm works without test and training.
thanks a lot
hi every one,
In Numenta’s theory of the brain, and in HTM algorithm implementations, sequence memory is very important. By nature, it is unsupervised, which means you don’t train it on one set of data and test it on another. The data is a continuous stream of reality, which is used to learn a model.
MNIST is not a problem for sequence memory, so we have to supervise it. In our MNIST experiments with DL networks, we apply ideas of spatial sparsity inspired by Spatial Pooling to these networks, and we do not apply temporal sequence memory (at this time). So we keep the structures of the networks and their training data, which were established in the realm of ML, so there is are training and test data sets.
Thanks for your complete reply.can we conclude that, always when we use spatial pooling we have test and train data and just when we use temporal memory we don’t have test and train data?
is there any case that we use spatial pooling and we don’t need test and train?
Whether you use test vs training data is really up to you. Just realize that with HTM, there is no “training phase” where you learn everything, and a “test phase” where you apply the learning to data you’ve never seen. With HTM, you are always learning, and always in “test phase”.