aiqbench analysis
This research introduces "ProtoSeam," a training technique that inserts learnable class prototypes at a network's semantic interface to boost classification accuracy without altering the inference architecture. By combining quadratic consensus penalties with standard loss functions, the method achieves up to five percentage points improvement on benchmarks like CIFAR-10 and TinyImageNet while maintaining theoretical rigor. The approach offers a practical pathway for enhancing model performance through architectural agnostic training modifications rather than hardware or algorithmic overhauls.
ProtoSeam: Lifting Classifier Training with Latent Gaussian Mixture Models
arXiv:2609.35174v1 Announce Type: new Abstract: 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.
Abstract: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.
Machine Learning (cs.LG)
arXiv:2609.35174 [cs.LG]
(or arXiv:2609.35174v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.35174 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 224 words
Related coverage
Beyond Mode Collapse: Generating Diverse Synthetic Expert Conversations via Generative Flow Networks
arXiv · Oct 1, 2026
Learning Contact Representation for Leg Odometry
arXiv · Oct 14, 2026
Multiclass Classification without Labels via Posterior Simplex Geometry
arXiv · Jul 29, 2026
[AINews] The Future of Latent Space
Latent Space · Sep 25, 2026