I now have a LSTM/ConvNet competitive version of my HTM-inspired algorithm. It now lives as part of a new library I am working on called Neo/RL, (neocortex + reinforcement learning). The reinforcement learning is not yet complete, but the predictive hierarchy is up and running smoothly, and it is able to match or even outperform LSTMs/ConvNets, all while being fully online without any sort of rehearsal or batching. Don’t believe me? Evaluate it for yourself! https://github.com/222464/NeoRL
I am working on additional benchmarks, here is one where I reproduced this paper which used LSTMs to predict moving digits (link to original paper: http://arxiv.org/pdf/1502.04681.pdf)
I alternate between predicting the next input based on current input and predicting the next input based on its own predictions in the video. The video is real-time.
I am also working on a text prediction benchmark, and a time series dataset benchmark (the latter of which already works, but I need the LSTM comparison working properly as well).
To truly convince people, I probably need more benchmarks still though, three is not enough! So, if you are interested in helping out on that front, let me know!
Until next time!