Jeff Hawkins takes the position that only Numenta sees the human brain as a temporal prediction engine based on sparse, hierarchical memory.
But actually, RNNs are great for recognizing and predicting temporal sequences (as we saw in the Karpathy post), RNNs use a sparse representation, and RNNs can be extended with hierarchical memory [1]
The big difference is that the neural network crowd are getting some spectacular results, and Numenta, well, maybe they'll be show more progress in the future.
Jeff Hawkins is super smart and a good guy, and he might get more done if he acknowledged the commonalities in the approaches rather than having to invent it all separately at Numenta. I really dont mean to be critical. Jeff inspired my own interest in machine intelligence.
[1] page 442, "Over the past year, several authors have made different proposals to augment RNNs with a memory module. Proposals include the Neural Turing Machine in which the network is augmented by a ‘tape-like’ memory that the RNN can choose to read from or write to, and memory networks, in which a regular network is augmented by a kind of associative memory. Memory networks have yielded excellent performance on standard question-answering benchmarks. The memory is used to remember the story about which the network is later asked to answer questions."
These communities have different goals from the very beginning. Pattern recognition vs. real intelligence, which is what HTM is about. Hawkins describes this gap well in his book.
But there is some cross-pollination, see the recent projects by Stan Franklin's lab on Sparse distributed memory and composite representations, it's a step towards integration with deep learning: http://ccrg.cs.memphis.edu/assets/papers/theses-dissertation...
On the other hand check out the work by Volodymyr Mnih from DeepMind https://www.cs.toronto.edu/~vmnih/, reinforcement learning with "visual attention" is a step towards consciousness models of the HTM/SDM/LIDA camp.
I was also under the misapprehension that deep learning is just about classification, but that isn't true.
Yes, reinforcement learning is the path to general intelligence, and the deep learning community is showing impressive progress on that front as well. The Deepmind demo [1] and the recent robotics work at Berkeley[2] are good examples.
Thanks for the link to Stan Franklin's work. I'm glad to hear there is work to integrate the two approaches.
It seems like theres a kind of taboo against HTM in the ML community. I guess it stems from their lack of impressive benchmarks, but I think that's kind of missing the point when it comes to HTM. Maybe HTM isn't the right solution, but I think there is a lot to be learned by using neocortex inspired models, and HTM is at least a solid step in that direction. And the work Numenta has contributed on Sparse Codings shouldn't be overlooked.
Extraordinary claims demand extraordinary evidence. Numenta has plenty of the former and nil as regards the latter.
Even if they had invented some believably interesting task and done fair comparisons with other methods and shown that HTM succeeds where others fail, it would be considered worth a look by the wider machine learning community.
Because it doesn't meet academic standards for publishing in NIPS and ICML, the most prestigious machine learning conferences.
Edit: To clarify, research papers generally cite other peer-reviewed research papers in similar venues preferentially. ML papers should mostly be citing ML papers in high-quality, peer-reviewed venues. HTM doesn't have papers like this to cite.
https://en.wikipedia.org/wiki/Hierarchical_temporal_memory#D...