AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Polarity Is All You Need to Learn and Transfer Faster

Qingyang Wang, Michael Alan Powell, Eric W Bridgeford, Ali Geisa, Joshua T Vogelstein

ICML 2023 · Architectures and efficiency › RNNs and normalization · approach: Scalable approximation and equivariant compute · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
fix weight polarities a priori to an adequate (sufficient) templateFrozen-Net on XOR-5D · XOR-5D, 60-72 training samplestraining time to reach same accuracy: takes 58% and 48% of Fluid-Net timeImproved
fix randomly chosen weight polarities (Bernoulli 0.5)Frozen-Net RAND-Polarity on XOR-5D · XOR-5Dtraining time and data to reach accuracyNo clear effect
transfer ImageNet polarity pattern and fix it (Frozen-Net IN-Polarity)AlexNet on Fashion-MNIST and CIFAR-10 · Fashion-MNIST and CIFAR-10, 100 training imagesvalidation error at convergence: 7% (Fashion-MNIST), 9.4% (CIFAR-10) less errorImproved
transfer polarity pattern without freezing (Fluid IN-Polarity) vs Fluid-Net (ablation)AlexNet · Fashion-MNIST and CIFAR-10validation error at convergence: 6% (Fashion-MNIST), 8.4% (CIFAR-10) gainImproved
add polarity freezing on top of IN-Polarity transfer (ablation)AlexNet · CIFAR-10validation error at convergence: 1% gain at 100 samples, up to 3% at 50000Improved
fix random polarities (RAND-Polarity) (ablation)AlexNet · Fashion-MNIST and CIFAR-10validation error at convergence vs Fluid-NetNo clear effect
transfer ImageNet polarity pattern (fixed) instead of ImageNet weights (finetune)AlexNet · Fashion-MNIST and CIFAR-10, limited datavalidation error at convergence: significant for Fashion-MNIST 250, 1000 and CIFAR-10 ≤1000 except 250Improved
initialize Frozen-Net with ImageNet weight magnitudes (IN-Weight) instead of random magnitudes (ablation)Frozen-Net AlexNet · Fashion-MNIST and CIFAR-10, limited datavalidation error / epochsNo clear effect
initialize Frozen-Net with ImageNet weight magnitudes when data and time are abundant (ablation)Frozen-Net AlexNet · Fashion-MNIST and CIFAR-10, abundant datavalidation accuracy: Fashion-MNIST gain ~1%, CIFAR-10 gain ~4%Improved
reduce network size for randomly polarity-fixed Frozen-Net (ablation)Frozen-Net RAND-Polarity on XOR-5D · XOR-5D, small networks (e.g., 10 hidden units)probability of learning and sample/time efficiencyHurt
increase network size for randomly polarity-fixed Frozen-Net (ablation)Frozen-Net RAND-Polarity on XOR-5D · XOR-5D, >15 hidden unitsprobability of learning XOR-5D: >99% chance at 15 hidden unitsImproved
Freeze polarities with Freeze-SGD using Adam-based training that resets sign-violating weightsFrozen-Net generally · generaloptimization disadvantageHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy