I ran an experiment showing that a frozen neural network exposes a compact intention signal in its early layers. AN1 reads that signal directly and reconstructs the teacher’s decisions without doing a full forward pass.
CIFAR-10 demo results:
- Teacher accuracy: 87.89%
- AN1 accuracy: 72.57%
- Speedup: 10.31×
- FLOP reduction: 1370×
- No access to the teacher’s weights
Author here.
Happy to answer questions about the setup, assumptions, and limitations.
The repo includes reproducible code for the CIFAR-10 experiment, including timing and FLOP breakdowns.
I ran an experiment showing that a frozen neural network exposes a compact intention signal in its early layers. AN1 reads that signal directly and reconstructs the teacher’s decisions without doing a full forward pass.
CIFAR-10 demo results: - Teacher accuracy: 87.89% - AN1 accuracy: 72.57% - Speedup: 10.31× - FLOP reduction: 1370× - No access to the teacher’s weights
GitHub repo: https://github.com/Anima-Core/an1-meaning-engine
Hugging Face: https://huggingface.co/Anima-Core/an1-meaning-engine/
Discussion thread: https://huggingface.co/Anima-Core/an1-meaning-engine/discuss...
Author here. Happy to answer questions about the setup, assumptions, and limitations. The repo includes reproducible code for the CIFAR-10 experiment, including timing and FLOP breakdowns.