While knowntobesample efficient, these methods havefailed tofully leverage recent advances indeep learning, forcing the use of less efficient but more scalable model-free methods which try to learn the values directly.

The size of the circle indicates the number of parameters. Larger circles mean more parameters.

First,theirsemantic feature extractions are outdated while state-of-the-art large-scale pre-trained language models like BERT cannot be utilized due to the lack of original text.

Pt i=1eqi(x)(eai eqi(Di 1)) which is the exactly the distribution computed byMWEM. D(x)log(D(x)) (6) The optimization problem becomesDt = argminD (X)Lmwem(D, eQt, eAt). We show the exact details ofGEM in Algorithms 2 and 3. Note that given a vector of queries Qt = hq1,...,qti,wedefinefQt() = hfq1(),...,fqt()i. B.1 Lossfunction(fork-waymarginals)anddistributionalfamily For anyz R,G(z)outputs a distribution over each attribute, which we can use to calculate the answer toaquery viafq. Empirically,wefindthatour model tends to better capture the distribution of the overall private dataset in this way (Figure 3).

Measurement uncertainty inthesystem arisesfromvariousestimation orprediction procedures in real-world that rely on sensory information ( e.g.

When question "Is isthe training?"We Weshow FPR95 (left) and AUROC (right).