استكشف أحدث الأوراق البحثية

M-TriX

M-TriX introduces a novel hybrid-causal neural architecture based on a dynamicallylearned balance between an expressive transformation path f(x) and a stabilizingresidual path x. The central idea is that robust intelligence emerges not from depthalone nor from residual shortcuts alone, but from a precise, self-regulated blendmodulated by a learned spatial p-map 5. We prove this hypothesis through Noise-Augmented Training (v0.6.0), which forces f(x) to specialize as a "denoising expert". Controlled interventions confirm the causal necessity of this hybrid design: M-TriX: A Hybrid-Causal NeuralArchitecture for Balanced Intelligence,Stability, and Interpretability 1. Blackout Collapse (p≈0): Forcing the model to "sleep" collapses accuracy to 10.67% (chance level), proving f(x) is causally necessary7. 2. Reinforced Dense Collapse (p=1.0): Forcing the model to be "dense" also collapses accuracy to 10.33%, proving the residual path x is equally essential for stability. 3. Noise Immunity: On noisy test data, M-TriX achieves 97.69% accuracy, while the dense model collapses to 10.32%. This demonstrates that the hybrid balance—which self-organizes to a robust equilibrium (p≈0.457) 10—is not just optimal, but the only state capable of stable, intelligent reasoning. M-TriX offers a new direction for interpretable, efficient, and causally grounded deep learning.