The authors view a classifier network (N = N_2 \circ N_1) as a composition at a single semantic interface.
Lifted training with prototypes improves classifier accuracy without changing inference architecture
By splitting a network at a semantic interface and training with decoupled losses across a Gaussian prototype layer, the method boosts test accuracy by up to five percentage points on standard benchmarks.
Research Lab
Robert Lampel · Timon Klein · Sebastian Sager
Otto von Guericke University Magdeburg · Max Planck Institute for Dynamics of Complex Technical Systems Magdeburg
Research Digest··3 min read
The authors propose ProtoSeam, a training method that splits a classifier into two subnetworks and inserts a learnable Gaussian prototype per class at the interface.
Why this paper
From Max Planck Institute for Dynamics of Complex Technical Systems Magdeburg and Otto von Guericke University Magdeburg
In one line
ProtoSeam improves classifier accuracy by inserting class prototypes during training and discarding them at inference.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§