@astermind/astermind-elm/dist/core/Activations.d.ts

JavaScript Extreme Learning Machine (ELM) library for browser and Node.js.

export type ActivationName = 'relu' | 'leakyrelu' | 'leaky-relu' | 'sigmoid' | 'tanh' | 'linear' | 'identity' | 'none' | 'gelu';
export declare class Activations {
    /** Rectified Linear Unit */
    static relu(x: number): number;
    /** Leaky ReLU with configurable slope for x<0 (default 0.01) */
    static leakyRelu(x: number, alpha?: number): number;
    /** Logistic sigmoid */
    static sigmoid(x: number): number;
    /** Hyperbolic tangent */
    static tanh(x: number): number;
    /** Linear / identity activation */
    static linear(x: number): number;
    /**
     * GELU (Gaussian Error Linear Unit), tanh approximation.
     * 0.5 * x * (1 + tanh(√(2/π) * (x + 0.044715 x^3)))
     */
    static gelu(x: number): number;
    /**
     * Softmax with numerical stability and optional temperature.
     * @param arr logits
     * @param temperature >0; higher = flatter distribution
     */
    static softmax(arr: number[], temperature?: number): number[];
    /** d/dx ReLU */
    static dRelu(x: number): number;
    /** d/dx LeakyReLU */
    static dLeakyRelu(x: number, alpha?: number): number;
    /** d/dx Sigmoid = s(x)*(1-s(x)) */
    static dSigmoid(x: number): number;
    /** d/dx tanh = 1 - tanh(x)^2 */
    static dTanh(x: number): number;
    /** d/dx Linear = 1 */
    static dLinear(_: number): number;
    /**
     * d/dx GELU (tanh approximation)
     * 0.5*(1 + tanh(u)) + 0.5*x*(1 - tanh(u)^2) * du/dx
     * where u = k*(x + 0.044715 x^3), du/dx = k*(1 + 0.134145 x^2), k = sqrt(2/pi)
     */
    static dGelu(x: number): number;
    /** Apply an elementwise activation across a 2D matrix, returning a new matrix. */
    static apply(matrix: number[][], fn: (x: number) => number): number[][];
    /** Apply an elementwise derivative across a 2D matrix, returning a new matrix. */
    static applyDerivative(matrix: number[][], dfn: (x: number) => number): number[][];
    /**
     * Get an activation function by name. Case-insensitive.
     * For leaky ReLU, you can pass { alpha } to override the negative slope.
     */
    static get(name: string, opts?: {
        alpha?: number;
    }): (x: number) => number;
    /** Get derivative function by name (mirrors get). */
    static getDerivative(name: string, opts?: {
        alpha?: number;
    }): (x: number) => number;
    /** Get both forward and derivative together. */
    static getPair(name: string, opts?: {
        alpha?: number;
    }): {
        f: (x: number) => number;
        df: (x: number) => number;
    };
    /**
     * Given softmax probabilities p, returns the Jacobian J = diag(p) - p p^T
     * (Useful for analysis; not typically needed for ELM.)
     */
    static softmaxJacobian(p: number[]): number[][];
}