MicroCloud Hologram Inc. Launches Deep Spiking Quantum Neural Network Technology for Noisy Image Classification

MicroCloud Hologram Inc. (NASDAQ: HOLO), (“HOLO” or the "Company"), a technology service provider, launched a Deep Spiking Quantum Neural Network (DSQ-Net) for noisy image classification, marking an important engineering exploration in the direction of deep integration between quantum computing and neuromorphic computing.

This technology is based on HOLO's long-accumulated experience in quantum algorithms and neural network engineering.

For the first time in an enterprise-level research and development framework, it systematically introduces a Variational Quantum Circuit (VQC) auxiliary training mechanism, constructing a novel hybrid quantum-classical deep learning system.

Unlike previous studies that only used quantum circuits as feature mapping modules or quantum kernel functions, the core innovation of DSQ-Net lies in: directly embedding quantum circuits into the training process of Deep Spiking Neural Networks (SNN), serving as a key computational unit to address the problems of non-differentiable spiking events and stochastic neuronal dynamics, thereby reconstructing the trainability of SNN at the system level.

From the perspective of overall architecture, DSQ-Net adopts a clear hybrid quantum-classical layered design.

The input end first receives noisy image data, and through a classical preprocessing module, encodes the pixel information into spatio-temporal spike sequences suitable for processing by spiking neurons.

This process fully leverages the expressive advantage of SNN in the temporal dimension, such that image noise is no longer simply treated as an interference term, but is modeled and absorbed as part of the temporal signal.

Subsequently, the deep spiking neural network is responsible for extracting high-level spatio-temporal features.

Unlike traditional convolutional neural networks that rely on continuous numerical values, this SNN forms event-driven feature representations that are robust to noise through the dynamic evolution of multi-layer spiking neurons.

However, in the critical stages of weight updating and feature mapping, the system does not completely rely on classical training mechanisms; instead, it introduces variational quantum circuits as auxiliary optimization modules.

In HOLO's DSQ-Net, the quantum layer is not simply an add-on, but is designed as a computational component tightly coupled with the SNN.

Specifically, the spike statistical features from the intermediate layers of the SNN are encoded into the amplitude space of the quantum state, and the high-dimensional spike distribution is compressed and mapped into the qubit state through amplitude encoding.

Compared to angle encoding or basis state encoding, amplitude encoding has significant advantages in representation efficiency and information density, allowing a limited number of qubits to carry complex spike feature structures.

After encoding is completed, the parameterized variational quantum circuit evolves the quantum state.

This circuit consists of multiple layers of tunable quantum gates, with its parameters being collaboratively updated by the classical optimizer along with the overall training objective.

The quantum measurement results are treated as a probabilistic evaluation of the current network state, and this evaluation result in turn guides the update direction of the classical SNN weights.

Through this mechanism, the originally non-differentiable spike firing process is indirectly embedded into a differentiable and tunable quantum optimization framework.

It is worth noting that this hybrid training strategy does not simply replace classical computation with quantum computation, but fully lever...