Methods › General › Initialization › Kaiming Initialization

Kaiming Initialization

2,931 papers tagged archive 2025-07-28

archive 2025-07-28 Description, source and code snippet are the archive's method entry.

Kaiming Initialization, or He Initialization, is an initialization method for neural networks that takes into account the non-linearity of activation functions, such as ReLU activations.

A proper initialization method should avoid reducing or magnifying the magnitudes of input signals exponentially. Using a derivation they work out that the condition to stop this happening is:

1/2nₗVar[wₗ] = 1

This implies an initialization scheme of:

wₗ ∼𝒩(0, 2/nₗ)

That is, a zero-centered Gaussian with standard deviation of √(2/nₗ) (variance shown in equation above). Biases are initialized at $0$.

Source: Delving Deep into Rectifiers: Surpassing Human-Level...See Code · pytorch/pytorch

Papers archive 2025-07-28

30 shown of 2,931, newest first. Repository counts are the archive's code-links table. A Syntology line states what Syntology ran from that paper's harvested code; it is per sample and not a correctness claim.

Tasks archive 2025-07-28

20 shown of 930 tasks the archive attaches to papers tagged with this method, by distinct papers. A task without a page in the catalog is plain text.

TaskPapers
Image Classification453
image-classification349
Semantic Segmentation285
General Classification229
Object Detection211
Classification203
Segmentation193
Transfer Learning184
object-detection177
Self-Supervised Learning159
Contrastive Learning148
Representation Learning144
Data Augmentation128
Deep Learning115
Object80
Instance Segmentation76
GPU72
Image Segmentation67
Decoder65
Diagnostic65

Usage over time archive 2025-07-28

Papers per year tagged with Kaiming Initialization: 2015 to 2025, peak 463 463 0 2015: 6 papers 2015 2016: 42 papers 2016 2017: 120 papers 2017 2018: 240 papers 2018 2019: 380 papers 2019 2020: 459 papers 2020 2021: 463 papers 2021 2022: 357 papers 2022 2023: 368 papers 2023 2024: 373 papers 2024 2025: 123 papers 2025
Papers per year the archive tags with this method, by the paper's archive date (2,931 dated). Bars are counts, not a trend claim.

Components: the archive holds no method-to-method composition, so PwC's Components table cannot be rebuilt; the Papers list carries no Results column for the same reason (the archive does not join its leaderboard rows to method tags).

Categories archive 2025-07-28

Initialization

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections