Are there changes to the aggregation model that could make it useful for training machine learning models?

[palenica]

For example, multiple-pass querying, different privacy accounting schemes, different batching schemes.

[alois-bissuel]

I guess supporting gaussian noise instead of a laplacian one would improve the utility of the API when reporting over a few dimensions.

[csharrison]

I guess supporting gaussian noise instead of a laplacian one would improve the utility of the API when reporting over a few dimensions.

Similarly, what we discussed before about moving from L1 constraints to L2 (or L0 + Linf) in this issue:
#249

I think that plus Gaussian noise is the key to unlocking the advanced DP composition tricks to improve privacy / utility in the face of multiple queries.