What are the other advantages of TCN’s over LSTMS apart from speed, do you think TCN’s would replace RNNs in the future?

So, the question is about this TCN stands for temporal convolution networks.The broad picture is this if you have sequential data, so, for example, if you have data like stock markets or the weather data temperature every day or the pollution data in in Delhi every day or if you have something like speech where you are getting some some frequency every time step in all such temporal data, how do we process them? You will typically require models like RNNs and LSTM to do. So you have a sequence of inputs and you would like to process it let us say, take the signal that I am speaking now and convert it into words. So the input is the waveform with different time steps showing what are the different frequencies and output becomes the different words I'm saying. Now something like this is usually processed using RNNs, which were the original models proposed for sequential data. And they were later enhanced with LSTM's with the ability to store data, forget data and so on. And very recently, I think in the last couple of years, this idea of temporal convolutional networks has come temporal of course means over time, convolution neural networks is also another well known well studied field for computer vision. So the idea is combined convolution which has been used mostly in the spatial domain you convolve over space with convolution on the time domain. That's why it's called temporal convolutional networks. And as far as I've seen, some recent papers have shown that it significantly improves on LSTMs and RNNs for both some biological data and also stock market data. So all this is evolving. So I don't think there is an answer as to whether it will be replaced or not. For example, we know that in the domain of natural language processing, RNNs and LSTM which were used a lot let us say three to four years back have been almost replaced now by attention based models. You might have heard about Bert, and and transformers, which use attention instead of the sequential models. So it remains to be seen. It depends on people training large models and showing that one is better than the other PCs, I think are still very early. And maybe we should wait for one or two more years to see how things shake out.