There are two main observations:
- Saturating counters should probably have two or three bits and the get worse with more, most likely due to the longer learning/re-learning phase.
- The (default) behaviour of hash functions: the bigger the image size to map into, the less collisions and therefore the higher the accuracy.