encoder bottleneck in seq-to-seq LSTMs

transformer architecture

The fixed-length encoder vector is the specific failure that attention-based direct addressing was introduced to remove, motivating the architectural replacement of RNN components.

1 source videos