Dec 12, 2020 · We propose a mathematical framework to evaluate geodesics in the functional space, to find high-performance paths from a dense network to its ...

(E) A depiction of multiple sparsification paths on the loss landscape from trained network (N1) to networks on the sparse (damage) hyper-plane (N2,. N3, N4, N5) ...

Sep 8, 2024 · We propose a mathematical framework to evaluate geodesics in the functional space, to find high-performance paths from a dense network to its sparser ...

Nov 7, 2020 · We propose a mathematical framework to evaluate geodesics in the functional space, to find high-performance paths from a dense network to its ...

Dec 6, 2020 · Neural Information Processing Systems (NeurIPS) is a multi-track machine learning and computational neuroscience conference that includes ...

... for network sparsification: We apply the geodesic strategy to discover high-performance paths from a trained network (VGG11 on CIFAR-10) to a sparse ...

Jun 5, 2021 · The three applications are: (i) Sparsifying networks by traversing geodesics, (ii) Alleviating catastrophic forgetting via geodesics and (iii).

Inspired by the geometric insight, we apply our geodesic framework to 3 major applications: (i) Network sparsification (ii) Mitigating catastrophic forgetting ...

Apr 25, 2024 · Solving hybrid machine learning tasks by traversing weight space geodesics. ... Sparsifying networks by traversing Geodesics. CoRR abs/2012.09605 ...

Mar 25, 2024 · Co-authors ; Sparsifying networks by traversing Geodesics. G Raghavan, M Thomson. arXiv preprint arXiv:2012.09605, 2020. 1, 2020 ; Geometric ...