ProtoSeam: Lifting Classifier Training with Latent Gaussian Mixture Models
ProtoSeam lifts classifier training with class prototypes, gaining up to five points without changing inference.
ProtoSeam reformulates supervised classification by splitting a network and inserting one learnable class prototype at the interface. Training pulls features toward the class prototype and trains the second stage on samples around prototypes, with no gradient crossing the split. At inference the prototypes are removed and the original network is used unchanged. On CIFAR-10, CIFAR-100, and TinyImageNet, ResNet and vision transformer backbones gained up to five percentage points over non-lifted variants.
- Inserts one learnable prototype per class at a network split.
- No gradient crosses the interface; prototypes are discarded at inference.
- Gains up to five points on CIFAR-10, CIFAR-100, and TinyImageNet.
- Tested with ResNet and vision transformer backbones under shared tuning.
Full article130 words · extracted from arxiv.org · click to collapse
We propose a lifted reformulation of supervised classification that improves the final accuracy of standard classifiers without changing the architecture at inference time. A network $N=N_2\circ N_1$ is split at a single semantic interface and one learnable prototype per class is inserted there. Training combines a quadratic consensus penalty that pulls $N_1(x)$ toward the prototype of its class with a classification loss of $N_2$ evaluated on samples drawn around the prototypes, whereat no gradient crosses the interface. At inference the prototypes are discarded and the unmodified network $N_2\circ N_1$ is used. Across CIFAR-10, CIFAR-100, and TinyImageNet with ResNet and vision transformer backbones, lifted training improves test accuracy by up to five percentage points over variants without lifting under a shared tuning protocol. Moreover, we provide theoretical justification of those results.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.35174