HLM-Vision

HLM-Vision is the validated 2D image path for Hopfield image classification. It uses a vision-shaped frontend before the shared Hopfield core, rather than feeding raw image patches directly into a language-style stack.

Current result

The packaged HLM-Vision-Mix checkpoints reached 78.63% validation accuracy for the compact model and 83.91% for the large model on CIFAR-10.

Validated stack

convolutional frontend
-> 2D spatial mixer
-> Hopfield block stack
-> pooled classification head
-> matched train/eval inference depth

The public lesson is simple: HLM does not remove the need for world-shaped frontends. It provides a shared Hopfield core after the data has been shaped into a useful sequence, grid, waveform, point cloud, or sensor representation.

Checkpoints

Checkpoint	Dataset	Validation accuracy	Use
HLM-Vision-Mix Large	CIFAR-10	83.91%	default demo
HLM-Vision-Mix Small	CIFAR-10	78.63%	compact demo

What to claim

• HLM can support 2D image classification when paired with a convolutional frontend, spatial mixer, pooled readout, and matched inference depth.
• The result validates the vision path for demos and follow-on pilots.
• This is not presented as a CIFAR-10 state-of-the-art claim.

What stays internal

Detailed failed-run timelines, normalization changes, iteration-depth ablations, and unpublished roadmap variants remain internal engineering material. Public docs should point to the validated stack and the current demo checkpoints.

Related

• Validation Summary
• Model overview
• HLM3 language model