Is acceptable or not?


#1

hi every one.I am phd student of UCLA university, and interested in both computer science (artificial intelligence) and neuroscience. i read just 1 paper about HTM and I think is better than deep learning and I decide to work on HTM. do you think HTM could be a good subject for my thesis and is there any challenge to solve it? or is acceptable to use HTM as an application problem?
would you mind send me some paper to help me? (i mean some paper that get me some idea for new work on HTM)
sincerely yours


#2

Hi Jack,

HTMs and neuroscience have many challenges left to solve. What makes a acceptable thesis topic is between you and your thesis advisor, but I can tell you that other people on this forum have written masters thesis’s about HTM’s. If you’d like to learn more about how the HTM algorithms work I’d recommend watching the “HTM School” videos on youtube. If you want to know more about the neuroscience behind HTMs then I would recommend skimming through the following two lists of papers. And if you have questions this forum is a good place to ask or to share the latest news.


#3

I’m also a PhD student. I would sincerely advise that you should NOT read one paper on a topic and decide that it is better than something else. “Better” can mean many things in many different contexts. For example, deep convolutional networks are very good at image classification. HTM is good at learning time series sequences.

Both have their strengths and weaknesses and you need to take these into account when comparing tools for particular applications.

If you really wanted, you could tackle image classification with HTM, though current results are far bellow what a basic MLP will get you.


#4

@Jonathan_Mackenzie yes, my teacher tell me the same thing when I represent to him the difference between deep learning and HTM, I tell him that HTM is better than deep learning in analysis streaming data, but in the images I don’t know who is better, he tells you can’t decide what’s the best because according to your data, your features, your dataset, your field… you can decide how is the better.
So @jack you can’t say that HTM is better than Deep learning, for example you can find some issues in deep learning and you can solve them with HTM or comparing the HTM and CNN on analysing images ( because CNN is the most popular deep on image) …


#5

Check this folder out: https://github.com/numenta/nupic.vision/tree/master/src/nupic/vision/mnist

They get 95% accuracy, other techniques listed here: http://yann.lecun.com/exdb/mnist/ get above 99% for this particular task.

Also, keep in mind that your network doesn’t necessarily have to be deep to get good results. DL nets usually need a large amount of training and data to be effective, HTM does not. Perhaps you could try a convolution encoder or layer for HTM or something. ie. you could have the conv layer + maxpool layer learn some salient features for an image.