QIMHNN Experimental

Quantum-Inspired Metaheuristic Neural Network

This module implements a neural network architecture inspired by quantum mechanics concepts (Superposition, Interference, and Entanglement) to potentially escape local minima and improve generalization.

Constructor

Located in Gradien.Experimental.Models.QIMHNN.

Definition

(config: QIMHNNConfig) -> Module

Configuration

type QIMHNNConfig = {
    inputDim: number,
    outputDim: number,
    hiddenDims: {number}?,     -- e.g. {64, 64}
    
    -- Quantum Features
    useSuperposition: boolean?, -- Default: true
    useInterference: boolean?,  -- Default: true
    useEntanglement: boolean?,  -- Default: true
    quantumAmplitude: number?,  -- Amplitude factor for |z| (default 0.1)
    
    -- Standard Features
    activation: string?,       -- Default activation (e.g. "ReLU")
    finalActivation: string?,  -- Output activation
    dropout: number?,          -- Dropout probability
    layerNorm: boolean?        -- Apply LayerNorm
}

Example

local model = Gradien.Experimental.Models.QIMHNN({
    inputDim = 10,
    outputDim = 2,
    hiddenDims = {32}, -- depth: 1
    useSuperposition = true,
    useInterference = true
})

Concepts

• Superposition: Weights utilize both real and imaginary components. The output incorporates a magnitude term |z|.
• Interference: Scales the output based on the phase difference between real and imaginary parts, simulating wave interference.
• Entanglement: Adds a dense connection E that mixes state information across the layer, simulating particle entanglement.