Periodically bad (nearly empty) atoms are removed from the dictionary and replaced by randomly generated ones. It seems that this would take several iterations before these random atoms get replaced by something useful. Another way to do this might be to select a random atom from the raw data instead. As this is already present in the raw data it is close to matching something of value already. If one wanted to get really clever, they could weight the probability of picking a particular vector from the raw data more highly if it is poorly described by the existing dictionary/code (e.g. large error, not sparse, etc.).