Brain-Inspired Quantum Neural Architectures for Pattern Recognition: Integrating QSNN and QLSTM

TL;DR

This work investigates brain-inspired quantum architectures for anomaly detection in credit card transactions by combining QSNN and QLSTM. It introduces a two-stage pipeline where QSNN filters low-level information akin to the prefrontal cortex and QLSTM handles higher-level memorization akin to the hippocampus, with amplitude encoding requiring $log_2 N$ qubits. The authors compare four quantum architectures against classical counterparts on an imbalanced fraud dataset, reporting improved data and parameter efficiency for the quantum models. Despite favorable performance, the study notes substantial runtime costs when simulating quantum circuits on classical hardware and advocates future work on online learning, scalability, and hardware deployment.

Abstract

Recent advances in the fields of deep learning and quantum computing have paved the way for innovative developments in artificial intelligence. In this manuscript, we leverage these cutting-edge technologies to introduce a novel model that emulates the intricate functioning of the human brain, designed specifically for the detection of anomalies such as fraud in credit card transactions. Leveraging the synergies of Quantum Spiking Neural Networks (QSNN) and Quantum Long Short-Term Memory (QLSTM) architectures, our approach is developed in two distinct stages, closely mirroring the information processing mechanisms found in the brain's sensory and memory systems. In the initial stage, similar to the brain's hypothalamus, we extract low-level information from the data, emulating sensory data processing patterns. In the subsequent stage, resembling the hippocampus, we process this information at a higher level, capturing and memorizing correlated patterns. We will compare this model with other quantum models such as Quantum Neural Networks among others and their corresponding classical models.

Figures (15)

• Figure 1: Leaky Integrate-and-Fire Neuron Model SNNTraining. An insulating lipid bilayer membrane separates the interior and exterior environments. Gated ion channels enable the diffusion of charge carriers like Na+ across the membrane.\ref{['fig:b1']}. RC circuit models neuron function. When the membrane potential exceeds the threshold, a spike is generated. \ref{['fig:b2']}. Input spikes are transmitted to the neuron body through dendritic branches. Sufficient excitation accumulation triggers spike emission at the output \ref{['fig:b3']}. A simulation illustrating the membrane potential $U(t)$ with a threshold of $\theta = 0.5V$, resulting in the generation of output spikes \ref{['fig:b4']}
• Figure 2: Typical morphology of a neuron. Consisting of a cell body, or soma, which contains the nucleus and other organelles, dendrites, which are fine, branched cell processes that receive synaptic input from other neurons, one axon and synaptic terminals.
• Figure 3: SNN pipeline. Input data for an SNN can be transformed into a firing rate or other encodings to generate spikes. The network is subsequently trained to predict the correct class, employing encoding strategies such as the highest firing rate or firing first, among others
• Figure 4: LSTM Cell Architecture: Featuring three essential gates (forget, input and output gates). The $\sigma$ and $tanh$ blocks symbolize the sigmoid and hyperbolic tangent activation functions, respectively. $x_t$ denotes the input at time $t$, $h_t$ represents the hidden state, and $c_t$ signifies the cell state. The symbols $\otimes$ and $\oplus$ denote element-wise multiplication and addition, respectively.
• Figure 5: General VQC Schema. The dashed gray line encompasses the steps executed in a Quantum Processing Unit (QPU) and the dashed blue line shows the steps executed in a CPU.