Deep Learning | TCN Temporal Convolutional Neural Network Model Description
Table of contents
basic introduction
TCN consists of dilated, causal 1D convolutional layers with the same input and output lengths. The following sections will detail what these terms actually mean. We first describe the general architecture of convolutional sequence prediction. Our aim is to distill best practices in convolutional network design into a simple architecture that can serve as a convenient yet powerful starting point. We call the proposed architecture a Temporal Convolutional Network (TCN), emphasizing that we adopt this term not as a label for a truly new architecture, but as a simple descriptive term for a simple family. (Note that this term has been used previously (Lea et al., 2017).) The distinguishing characteristics of TCNs are:
1) Convolutions in the architecture are causal, meaning there is no information “leaking” from the future to the past;
2) The architecture can take a sequence of arbitrary length and map it to an output sequence of the same length, just like RNN. (However, tcn does not implement this function through seq2seq, but simply uses the convolutional layer)
3) In addition to this, we also highlight how to use a combination of very deep networks (with residual structures) and dilated convolutions to build very long effective history sizes (i.e. the ability of the network to look very far into the past to make predictions.
Our architecture is inspired by recent sequential data convolutional architectures (vanden Oord et al.