Multiphoton quantum simulation of the generalized Hopfield memory model

Gennaro Zanfardino, Stefano Paesani, Luca Leuzzi, Raffaele Santagati, Fabrizio Illuminati, Giancarlo Ruocco, Marco Leonetti · 2025

Read Multiphoton quantum simulation of the generalized Hopfield memory model by Gennaro Zanfardino, Stefano Paesani, Luca Leuzzi, Raffaele Santagati, Fabrizio Illuminati, Giancarlo Ruocco, Marco Leonetti (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-00111.txt.

Introduction

Gennaro Zanfardino and a team including Stefano Paesani and Marco Leonetti link light-based quantum devices to a classic model of memory, in a paper from twenty twenty-five. The Hopfield model treats memory as a network of simple units that settle into stored patterns. It is a basic model for both memory and machine learning.

Their setup sends several photons through an optical circuit with set phase shifts. The output counts match the behavior of a Hopfield network with four-way links between units. As more patterns are stored, the system shifts from recalling memories to a blacked-out, glassy state. The link may let light devices simulate messy, complex systems with ease.

Catalog record

CREATORGennaro Zanfardino, Stefano Paesani, Luca Leuzzi, Raffaele Santagati, Fabrizio Illuminati, Giancarlo Ruocco, Marco Leonetti
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-00111

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Multiphoton quantum simulation of the generalized Hopfield memory model Preprint: APS/123-QED Gennaro Zanfardino, Stefano Paesani, Luca Leuzzi, Raffaele Santagati, Fabrizio Illuminati, Giancarlo Ruocco, Marco Leonetti Abstract In the present work, we introduce, develop, and investigate a connection between multiphoton quantum interference, a core element of emerging photonic quantum technologies, and Hopfield-like Hamiltonians of classical neural networks, the paradigmatic models for associative memory and machine learning in systems of artificial intelligence. Specifically, we show that combining a system composed of N ph {N_{\rm ph}} indistinguishable photons in superposition over M M field modes, a controlled array of M M binary phase-shifters, and a linear-optical interferometer, yields output photon statistics described by means of a p p -body Hopfield Hamiltonian of M M Ising-like neurons ± 1 \pm 1 , with p = 2 ​ N ph p=2{N_{\rm ph}} . We investigate in detail the generalized 4 4 -body Hopfield model obtained through this procedure and show that it realizes a transition from a memory retrieval to a memory black-out regime, i.e. a spin-glass phase, as the amount of stored memory increases. The mapping enables novel routes to the realization and investigation of disordered and complex classical systems via efficient photonic quantum simulators, as well as the description of aspects of structured photonic systems in terms of classical spin Hamiltonians. I Introduction In recent years, the study of bosonic quantum interference with photons coherently evolving through linear-optical networks is of central importance in the current development of photonic quantum technologies. This phenomenon implies a wide range of applications, going from architectures for fault-tolerant photonic quantum computing Knill et al. 2001 ; Gimeno-Segovia et al. 2015 ; Bartolucci et al. 2023 to potential near-term quantum computational advantages with specialized algorithms Aaronson and Arkhipov 2013 and high-precision sensing Pirandola et al. 2018 . Recent experimental advances have enabled the generation, interference, and detection of systems with over 100 photons, achieving quantum computations at scales competitive with conventional supercomputers Zhong et al. 2020 ; Madsen et al. 2022 ; Aspuru-Guzik and Walther 2012 . In the present work, we establish and investigate the relation between multiphoton quantum interference in linear optics and generalized p -body Hopfield models of statistical physics Peretto and Niez 1986 ; Baldi and Venkatesh 1987 ; Gardner 1987 ; Abbott and Arian 1988 ; Horn and Usher 1988 ; Krotov and Hopfield 2016 ; Ramsauer et al. 2021 . The Hopfield model (HM) Hopfield 1982 is a fundamental tool in the study and characterization of memory storage and retrieval in biological neural networks Amit 1989 ; Müller et al. 1995 ; Hertz et al. 2018 , as well as in the investigation of machine learning problems Hinton and Salakhutdinov 2006 ; Hinton 2007 . The HM is a full feedback neural network providing pattern recognition, associative memory functions, error correction, and other capabilities Amit et al. 1985a ; Amit et al. 1987 . Generalized Hopfield Models (gHMs) with multi-synaptic interaction in the structure of deep neural networks are employed in studying information storage and prototype learning Krotov and Hopfield 2016 ; Ramsauer et al. 2021 . Understanding the different phase regimes and collective behaviors associated with HMs and gHMs is thus a problem at the cutting edge in the study of artificial intelligence. Here we show how photonic architectures with a structure (cf. Fig. 1 a-d) composed of i) N_{\text{ph}} photons in an initial superposition state over M optical field modes, ii) a layer of M phase shifters (one on each mode), and iii) a linear-optical interferometer implementing a general scattering matrix over the M modes, display photon statistics in the output modes that can be described through the Hamiltonian of a p=2N_{\text{ph}} -body Hopfield model of M neurons. The p=2 HM has been first optically simulated with incoherent light Farhat et al. 1985 , and successively with coherent light and a digital micromirror device architectures Leonetti et al. 2021 , showing speed advantages for the all-optical simulations of large-spin systems. Furthermore, Hopfield-inspired memory storage and computation has been recently developed Leonetti et al. 2024 . In these cases, as in the original HM Hopfield 1982 , the number P of memories that can be retrieved is a fraction of the number of neurons in the network, \alpha={P}/M<1 and cannot grow more than linearly with M . Instead, considering energy functions with interactions of order p>2 , it is possible to realize neural networks capable of retrieving a much larger number of stored memory patterns Krotov and Hopfield 2016 ; Gardner 1987 ; Horn and Usher 1988 . One drawback is that simulating the dynamics of fully connected networks with multi-neuronal synaptic couplings (i.e., in which p neurons interact simultaneously) requires a time that scales super-extensively with the number of neurons. Generically, the computation time of the energy gradient in a dynamical iteration grows like M^{p-1} , i.e., as the average connectivity per neuron. However, the analog photonic computation would drastically reduce this time. To generalize to the multi-synaptic large storage case, though, requires a further stratagem in the photonic system proposed for the energy measurement: the use of entangled photons and interferometry. The connection between linear-optical quantum systems and models from statistical physics opens ways for investigating classical p HM with quantum photonic systems, potentially enabling efficient calculations for large p and M . We illustrate this through the simulation of photonic systems implementing a 4 HM, showing evidence for the characteristic transition from retrieval to spin-glass phase emerging from the photons’ statistics. Figure 1: a) Standard description of linear optical transformation from an input configuration \vec{c} to an output configuration \vec{k} through a linear scattering matrix U . An example here is depicted for {N_{\rm ph}}=3 photons and M=5 modes. b) Schematic of the mapping of an p HM to a photonic system composed of an input state \ket{\psi} , a set of M phase-shifters which can have binary values \phi_{i}\in\{0,\pi\} and map to the M spin states \sigma_{i}=\text{exp}(i\phi_{i})=\pm 1 , and a scattering matrix S . c) Full schematic including the initial Discrete Fourier Transform U^{\text{DFT}} for generating the near-uniform input state \ket{\psi_{\text{DFT}}} . d) Simplified scheme with the set of output states \Lambda_{\text{FB}} given by fully-bunched configurations with all photons exiting in the same output mode. Panel e) reports the “on-chip” experimental scheme, correspondent to c). II Photonic mapping of the generalized Hopfield model The type of physical system we consider for the mapping is a linear optical quantum processor, composed of single photons as inputs, a linear interferometer, which can universally be implemented with only phase-shifters and beam-splitters Reck et al. 1994 ; Clements et al. 2016 , and photo-detectors. Such a platform is, in principle, compatible with scalable quantum computers Knill et al. 2001 ; Gimeno-Segovia et al. 2015 ; Bartolucci et al. 2023 as well as near-term quantum simulators Sparrow et al. 2018 , which is the type of application considered here. The advantages of photonic implementations include lighting fast transformations, inherent parallel operations, and low consumption, which are, in general, valuable properties for future generations of computing hardware Leonetti et al. 2024 ; Leonetti et al. 2021 ; Fan et al. 2023 ; Pierangeli et al. 2019 . Here, we propose to use

Read the complete text