Get the latest tech news
The simple beauty of XOR floating point compression
I recently implemented a small program to visualize the inner workings of a scheme that compresses floating point timeseries by XORing subsequent values. The resulting visualizations are quite neat…
I recently implemented a small program to visualize the inner workings of a scheme that compresses floating point timeseries by XORing subsequent values. This is a fairly typical floating point time series which records the gradient norm at each iteration of a machine learning run. We can save a tiny amount of extra space if we treat it as a single precision floating point number to begin with since in that case we need one less bit to store the length of the nonzero subsequence.
Or read this on Hacker News