What is the difference between LSTM and fully connected LSTM?

Question

I'm currently trying to understand the difference between a vanilla LSTM and a fully connected LSTM. In a paper I'm reading, the FC-LSTM gets introduced as

FC-LSTM may be seen as a multivariate version of LSTM where the input, cell output and states are all 1D vectors

But is not really expanded further upon. Google also didn't help me much in that regard as I can't seem to find anything under that keyword.

What is the difference between the two? Also, I'm a bit confused by the quote - aren't inputs, outputs, etc. of a vanilla LSTM already 1D vectors?

Answer 1

Based on the citations in the ConvLSTM paper I have come to the conclusion that they mean the Peephole LSTM when they say fully connected LSTM. In the paper that they have taken the encoder-decoder-predictor model from, where they refer to a fully connected LSTM, a Peephole LSTM is used. Also they take their fully connected LSTM definition from this paper, which again uses the Peephole LSTM.

With that the difference would be the added "peephole connections", that lets the gate layers look at the cell states and access the constant error carousel.