4 ms·

I made a tiny paper for a masters thesis on KF for recurrent neural network training in 2002. 15 or 20 Neurons (a recurrent layer of a handful of neurons) was m

I made a tiny paper for a masters thesis on KF for recurrent neural network training in 2002. 15 or 20 Neurons (a recurrent layer of a handful of neurons) was massive in terms of the compute time needed to train it. To even get it off the ground we had to rewrite all the matlab into C++. Recurrent networks for modeling dynamic systems where all the rage they but I haven’t really kept track of where the field went. Did the deep learning revolution completely drop recurrent connections?

mjburgess 6y ago
They became a major phenomenon in NLP (ie., sequences of text strings) -- but have since been less prominent. RNNs are still a significant part of the toolbox, and would be found in several chapters in any applied NN book.
danans 6y ago
Last I heard, KFs are still essential for simultaneous localization and mapping (SLAM) applications.