Robert Hu

Hey folks, I’ve just wrapped up some work at Graphcore on numerics, FP4 pre-training and inference, and wanted to share it here:

arxiv.org/abs/2609.02846
arxiv.org/abs/2610.00053
arxiv.org/abs/2609.04105
arxiv.org/abs/2610.00049

Four papers, a few things we learned 🧵

Happy to share that "A Kernel Test for Causal Association via Noise Contrastive Backdoor Adjustment" has just been published in JMLR!

paper: jmlr.org/papers/v25/21-…

with @sejDino and @ItsAStatLife

tl;dr New hypothesis test for causal effect under the backdoor criterion 1/6

Happy to be presenting Faster Fast and Free Memory Method (F^3M) at #NeurIPS2022 Hall J #1034.

We take on kernel matrix vector multiplication (KMVM) at scale for tall and skinny data (n<=10^9, D<=7) on GPU in linear time and memory.

neurips.cc/virtual/2022/p…
#GPs
#KeOps

Happy to be presenting Pref-SHAP with @Chau9991 at #NeurIPS2022 Hall J #905.

We develop Shapley values for generalised preference models (GPM), which are preference models without the transitivity assumption (no rock paper scissor relationships).

nips.cc/virtual/2022/p…

Me and @vdwild will be presenting GWI-net tmrw at poster session 1 #509 at #NeurIPS2022! We bridge VI in infinite-dimensional spaces with BDL, where GWI-net allows us to equip deep nets with uncertainty quantification analogous to that of GPs

neurips.cc/virtual/2022/p…