Hard Parameter Sharing
A parameter-sharing paradigm in multi-task learning in which all tasks share a single underlying encoder while retaining their own task heads at the output end; it uses the fewest parameters but is prone to mutual interference when tasks differ greatly.
Timeline (last 90 days)
Hard Parameter Sharing makes all tasks share the same underlying encoder and keeps only their respective task heads at the output end, saving the most parameters and minimizing overfitting risk, but tasks easily interfere with each other when they differ greatly.
Compared with Hard Parameter Sharing, MMoE usually performs more stably when multi-task correlations are weak, but its structure is more complex, its training cost is higher, and the gating weights require enough data to be learned well.
All Facts (2)
Hard Parameter Sharing makes all tasks share the same underlying encoder and keeps only their respective task heads at the output end, saving the most parameters and minimizing overfitting risk, but tasks easily interfere with each other when they differ greatly.
50%UnverifiedCompared with Hard Parameter Sharing, MMoE usually performs more stably when multi-task correlations are weak, but its structure is more complex, its training cost is higher, and the gating weights require enough data to be learned well.
50%