Implementation:Tensorflow Tfjs Noise Layers

Knowledge Sources	Tensorflow_Tfjs TensorFlow.js API
Domains	Deep_Learning, Layers_API, Regularization
Last Updated	2026-02-10 06:00 GMT

Overview

This module implements three noise-based regularization layers for TensorFlow.js Layers: GaussianNoise, GaussianDropout, and AlphaDropout. These layers add stochastic noise to inputs during training to reduce overfitting, and pass inputs through unchanged during inference. All three layers support masking.

Code Reference

Source Location

tfjs-layers/src/layers/noise.ts (GitHub)

Key Imports

import {add, greaterEqual, mul, randomUniform, serialization, Tensor, tidy} from '@tensorflow/tfjs-core';
import * as K from '../backend/tfjs_backend';
import {Layer, LayerArgs} from '../engine/topology';

Layer Classes

GaussianNoise

Adds additive zero-centered Gaussian noise during training.

export class GaussianNoise extends Layer {
  static className = 'GaussianNoise';
  readonly stddev: number;
  constructor(args: GaussianNoiseArgs);  // { stddev: number }
  override call(inputs: Tensor | Tensor[], kwargs: Kwargs): Tensor | Tensor[];
}

During training: output = input + N(0, stddev). During inference: identity pass-through.

GaussianDropout

Applies multiplicative Gaussian noise during training.

export class GaussianDropout extends Layer {
  static className = 'GaussianDropout';
  readonly rate: number;
  constructor(args: GaussianDropoutArgs);  // { rate: number }
  override call(inputs: Tensor | Tensor[], kwargs: Kwargs): Tensor | Tensor[];
}

During training (when 0 < rate < 1): output = input * N(1, sqrt(rate / (1 - rate))). During inference: identity pass-through.

AlphaDropout

Applies alpha dropout that preserves the mean and variance of inputs, designed for use with SELU activations (self-normalizing networks).

export class AlphaDropout extends Layer {
  static className = 'AlphaDropout';
  readonly rate: number;
  readonly noiseShape: Shape;
  constructor(args: AlphaDropoutArgs);  // { rate: number, noiseShape?: Shape }
  override call(inputs: Tensor | Tensor[], kwargs: Kwargs): Tensor | Tensor[];
}

Uses the SELU constants (alpha = 1.6732..., scale = 1.0507...) to maintain self-normalizing properties after dropout.

I/O Contract

Layer	Input	Output (training)	Output (inference)
GaussianNoise	Tensor of any shape	Tensor + Gaussian noise	Same as input
GaussianDropout	Tensor of any shape	Tensor * multiplicative noise	Same as input
AlphaDropout	Tensor of any shape	Alpha-dropout transformed tensor	Same as input

All layers preserve the input shape (output shape equals input shape).

Usage Example

import * as tf from '@tensorflow/tfjs';

const model = tf.sequential();
model.add(tf.layers.dense({units: 64, activation: 'relu', inputShape: [128]}));
model.add(tf.layers.gaussianNoise({stddev: 0.1}));

// For self-normalizing networks with SELU
const seluModel = tf.sequential();
seluModel.add(tf.layers.dense({units: 64, activation: 'selu', inputShape: [128]}));
seluModel.add(tf.layers.alphaDropout({rate: 0.1}));

Related Pages

• Tensorflow_Tfjs_Regularizers - Weight regularization (complementary approach)
• Tensorflow_Tfjs_Embedding_Layer - Embedding layer that may use noise layers downstream