Anytime Tail Averaging
This work addresses memory-efficient tail averaging for streaming data applications, but it appears incremental as it builds on existing averaging techniques.
The paper tackled the problem of tail averaging in streaming data, where existing methods suffer from high memory costs, lack of availability at every timestep, or inability to handle growing windows, and proposed two low-memory techniques that provide access to the average at every timestep, with one method improving accuracy at the cost of increased memory.
Tail averaging consists in averaging the last examples in a stream. Common techniques either have a memory requirement which grows with the number of samples to average, are not available at every timestep or do not accomodate growing windows. We propose two techniques with a low constant memory cost that perform tail averaging with access to the average at every time step. We also show how one can improve the accuracy of that average at the cost of increased memory consumption.