Skip to main content
. Author manuscript; available in PMC: 2019 Dec 23.
Published in final edited form as: Adv Neural Inf Process Syst. 2019 Dec;32:9392–9402.

Table 2: Model characterizations:

We characterize each model’s advantages/limitations and compute the number of parameters for each baseline model, given k slices, M backbone parameters, feature representation z dimension r, and slice expert representation pi dimension h.

Method Slice-aware No manual
tuning		 Weighted
slice info.		 Avoids copies of
model (M params)		 Num. Params
Vanilla ✓ ✓ O(M+r)
HPS ✓ ✓ ✓ O(M+kr)
Manual ✓ ✓ ✓ O(M+kr)
MoE ✓ ✓ ✓ O(kM+kr)
Ours ✓ ✓ ✓ ✓ O(M+krh)