I come from DL background and I’m new to HTM. I find HTM’s online learning & sparsity feature very appealing, as compared to usual batched training+BPTT approach for RNNs. However are there certain benchmarks results that can readily prove HTM’s performance? Like sequential MNIST classification, Penn Tree bank, or tasks introduced in the original 1997 LSTM paper? I’m not expecting state of art results, but at least it gives an idea on what HTM can do and how well it does.
“Continuous Online Sequence Learning With An Unsupervised Neural Network Model” compares HTM to LSTM, and other sequence learning algorithms, on several sequence prediction tasks.
“Evaluating Real-Time Anomaly Detection Algorithms - The Numenta Anomaly Benchmark” presents a benchmark we’ve developed specifically for anomaly detection in streaming data, along with results comparing HTM to popular anomaly detection algorithms; NAB repo here.