Blog
Notes on privacy, algorithms and the occasional implementation diary.
-
LSH applied to Transformer: The Reformer
Locality Sensitive Hashing (LSH) applied to the attention layer of a transformer reduces the inference time to almost linear
Read -
I finally implemented my own algorithm
TensorCloseTop-1 existed only as a proof. Now it runs. Here is what the theory got right and what it quietly hid.
Read -
Paper Accepted at Theory and Practice of Differential Privacy 2025
InfTDA generalises the TopDown mechanism to any tabular data with a hierarchical structure. It will be presented at TPDP 2025, at Google Mountain View.
Read