If I understand correctly, each level of the hierarchy divides its input into chunks of variable size, but outputs a fixed amount for each chunk. The chunking i

If I understand correctly, each level of the hierarchy divides its input into chunks of variable size, but outputs a fixed amount for each chunk. The chunking is learned. The model can choose to compress data by making its input chunks bigger, depending on their content.