Concept
Truncated backpropagation through time
Truncated backpropagation through time (TBPTT) is a training technique for recurrent models that limits how far gradients flow backward across time steps. Standard backpropagation through time unrolls a recurrent network over its full input sequence and computes gradients…
The rest of “Truncated backpropagation through time” is a premium feature: every concept in the library gets a precise, practitioner-focused write-up like this one, cross-linked straight from the paper summaries that use it.
Log in to unlock→