The current approximation models use inducing points to reduce the number of data points. Do there exist approximation models specifically for multi-output models so that the number of output channels is more scalable (ie. doesn't go exponential with the number of output channels).
tdewolff
commented
9 months ago
The current approximation models use inducing points to reduce the number of data points. Do there exist approximation models specifically for multi-output models so that the number of output channels is more scalable (ie. doesn't go exponential with the number of output channels).
One example is for the CONV kernel: https://proceedings.neurips.cc/paper_files/paper/2008/file/149e9677a5989fd342ae44213df68868-Paper.pdf