pyhgf.model.DeadGradientWarning

exception pyhgf.model.DeadGradientWarning

Some weight matrices receive essentially no learning signal.

Raised by DeepNetwork.check_gradient_health() when part of the network is training while another part receives (near-)zero gradients — typically the matrices far from the observations. A network in this state silently degrades into a random-feature readout: the live matrices fit against a frozen upper stack, accuracy can look respectable, and nothing else reports the failure.

The two known causes are Bayesian absorption (each layer’s inference consumes most of the error message before passing the residual up) and float32 annihilation of the belief-side factor (μ − μ̂)·π at high precision.