About¶
Oxiedo¶
ormas-monitor is built by Rokib Al Dhin Raadh at Oxiedo. It is released under the Apache-2.0 licence.
Contact: raadh@oxiedo.com.
The paper¶
The dead-channel certificate is Theorem 2 of:
R. A. D. Raadh. Self-Repair as a Bounded Disturbance: Input-to-State Stability of Neural Network Training Dynamics. DeepMath 2026, The Ohio State University, 29–30 October 2026. Accepted as a poster.
The paper studies self-repair during training as a disturbance to the training dynamics. One of its results is that a Conv/Linear → BatchNorm → ReLU channel with a zero filter (or zero BatchNorm scale) and a non-positive shift sits in an absorbing dead set: gradient descent cannot leave it, and neither can any repair whose size is proportional to the weights. Only a discrete jump, such as a re-initialisation, gets out. The CIFAR-10 dead-layer experiment in Results is the experiment from the poster.
ORMAS, and how this tool relates to it¶
ORMAS is Oxiedo's research training architecture: a network design with local heads, three health signals and seven diagnoses, studied in Oxiedo's research. ormas-monitor is not ORMAS and contains none of it. The name reflects where the work came from, not what the package does.
ormas-monitor is a separate, smaller tool. It needs no architecture change: it attaches to an ordinary PyTorch model from outside, with hooks. It packages one result from that research context (the certificate) together with general-purpose monitoring, a checkpoint check, a training record and reports.
The ORMAS research results, including its accuracy numbers, are not claims about this tool and are not repeated here.
Prior art¶
Dead and dormant units are well studied. ReDo (Sokar et al., The Dormant Neuron Phenomenon in Deep Reinforcement Learning, ICML 2023) detects and resets dormant neurons. Network slimming (Liu et al., Learning Efficient Convolutional Networks through Network Slimming, ICCV 2017) drives BatchNorm scales toward zero to prune channels. ormas-monitor claims the certificate, checked at run time against its assumptions, and the packaging; not the discovery.
Citing¶
See CITATION.cff.