3 ms·

Thanks. One thought I’ve had is to wonder if attention isn’t kind of like a way to unroll RNNs to make training more efficient. Any truth in that?