AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Neural Path Features and Neural Path Kernel : Understanding the role of gates in deep learning

Chandrashekar Lakshminarayanan, Amit Vikram Singh

NeurIPS 2020 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability · 0 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
learn gates in decoupled feature network (DLNPF) instead of fixed random gates (GCONV, Adam) (ablation)deep gated network · CIFAR-10test accuracy (%): 67.09 -> 77.12Improved
learn gates in DLNPF instead of fixed random gates (FC, Adam) (ablation)deep gated network · MNISTtest accuracy (%): 96.02 -> 98.22Improved
use fixed learnt gates (FLNPF) instead of fixed random gates (GCONV, Adam) (ablation)deep gated network · CIFAR-10test accuracy (%): 67.09 -> 79.68Improved
use fixed learnt gates (FLNPF) instead of fixed random gates (VCONV, SGD) (ablation)deep gated network · CIFAR-10test accuracy (%): 58.92 -> 63.06Improved
make feature and value network initialisation independent (FRNPF-II) versus identical (FRNPF-DI) (ablation)fixed random NPF · MNIST and CIFAR-10test accuracy (%): 67.09 -> 67.08No clear effect
replace ReLU network with FLNPF (GCONV, Adam) (ablation)ReLU CNN · CIFAR-10test accuracy (%): 80.32 -> 79.68No clear effect
use gates from later training stage (FLNPF-1 to FLNPF-6) (ablation)FLNPF · CIFAR-10 GCONVtest accuracy (%): 72 -> 79.43Improved
learn gates (FLNPF-6) versus infinite width CNTKGCONV · CIFAR-10test accuracy (%): 77.43 -> 79.43Improved
use finite width with fixed random gates versus CNTKGCONV FRNPF · CIFAR-10test accuracy (%): 77.43 -> 67.09Hurt
use Adam instead of SGD for GCONV DLNPFdeep gated network · CIFAR-10test accuracy (%): 74.57 -> 77.12Improved
add global average pooling (GCONV versus VCONV) with ReLU, AdamReLU CNN · CIFAR-10test accuracy (%): 72.43 -> 80.32Improved

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