Đã lưu trong:
Chi tiết về thư mục
Tác giả chính: Souilah, Rachik
Định dạng: Recurso digital
Ngôn ngữ:
Được phát hành: Zenodo 2026
Những chủ đề:
Truy cập trực tuyến:https://doi.org/10.5281/zenodo.18448348
Các nhãn: Thêm thẻ
Không có thẻ, Là người đầu tiên thẻ bản ghi này!
Mục lục:
  • <p>We address the vulnerability of Deep Neural Networks (DNNs) to catastrophic forgetting under<br>severe non-stationary data drifts (trauma). Conventional safety mechanisms, such as Gradient Clip-<br>ping, constrain the instantaneous magnitude of updates but fail to prevent the cumulative erosion of<br>prior representations under sustained stress. This paper introduces Neural Guardian, a meta-control<br>layer that enforces a long-term energetic invariant on the learning process. Unlike optimization strate-<br>gies that seek to minimize loss, Neural Guardian operates as an independent safety kernel: it moni-<br>tors an energetic proxy derived from the cumulative gradient norm and enforces a Bounded Transient<br>Release (BTR) logic. When cumulative stress exceeds a critical threshold, the kernel asserts a mul-<br>tiplicative veto (α → 0) on the optimizer, imposing a mandatory dissipation period. Experimental<br>results on a TimesFM proxy demonstrate that while Gradient Clipping only delays performance<br>collapse, Neural Guardian successfully preserves nominal task performance (Performance Decay<br>Rate ≈ 0) by strictly refusing to integrate incompatible distributional shifts. This establishes Neural<br>Guardian not as an optimizer, but as a necessary boundary condition for lifetime stability in adaptive<br>systems.</p>