The remarkable generalization capabilities of overparameterized neural networks remain one of the most profound mysteries in modern machine learning. Despite having sufficient capacity to simply memorize training data, these networks consistently achieve superior performance on unseen test examples. Traditional analyses focusing on weight distributions, loss landscapes, and optimization dynamics have provided partial insights but fail to capture the complex correlational structure that emerges between neurons during learning. This paper introduces a comprehensive quantum information framework for analyzing neural network representations, treating hidden layer activations as quantum states and characterizing them through density matrices, von Neumann entropy, purity, and related quantum information metrics. Through systematic experiments on synthetic classification tasks and algorithmic reasoning problems spanning over 3000 cumulative training epochs and 5000 Wang-Landau sampling iterations, we demonstrate that these quantum metrics reveal structure completely invisible to classical analysis. Our results show that stochastic gradient descent systematically reduces von Neumann entropy while increasing purity as networks learn, transitioning from highly mixed quantum states with entropy approximately 1.42 and purity approximately 0.29 to purer configurations with entropy approximately 1.19 and purity approximately 0.37 that achieve superior test accuracy of 99.5%. Using Wang-Landau sampling to explore the equilibrium entropy landscape, we discover that neural networks naturally favor high-entropy states with entropy approximately 1.42 and moderate accuracy of 62.0%, and that optimization actively purifies the quantum state to achieve task performance. On modular arithmetic tasks exhibiting the grokking phenomenon, we observe that von Neumann entropy increases dramatically from 1.55 to 3.29, representing a 112% increase, during an extended 2000-epoch memorization phase before any improvement in generalization, while purity decreases correspondingly from 0.444 to 0.042, a 90.5% decrease, approaching the maximally mixed limit of 0.01. This entropy accumulation phase appears necessary for eventual generalization, with the network building quantum correlations at a rate of approximately 0.00097 nats per epoch during the plateau. Weight entropy, in contrast, shows minimal change of only 11% increase over the same period. Correlation analysis reveals that von Neumann entropy and purity are almost perfectly anticorrelated with a correlation coefficient of 1-0.99, while weight entropy shows weak correlation with quantum metrics with absolute correlation coefficients less than 0.3. The effective rank of the density matrix stabilizes near 2.7 to 3.0 for the 6-neuron classifier, indicating that the network effectively uses approximately three independent activation modes. These findings establish quantum information metrics as powerful tools for understanding neural network learning dynamics, provide the first quantitative evidence for the entropy accumulation interpretation of grokking, and suggest that neural networks undergo a quantum-inspired phase transition during learning. We discuss implications for generalization theory, optimization algorithm design, and the statistical physics of deep learning, and outline a comprehensive research agenda for extending these methods to larger architectures and more complex tasks.
The remarkable generalization capabilities of overparameterized neural networks remain one of the most profound mysteries in modern machine learning. Despite having sufficient capacity to simply memorize training data, these networks consistently achieve superior performance on unseen test examples. Traditional analyses focusing on weight distributions, loss landscapes, and optimization dynamics have provided partial insights but fail to capture the complex correlational structure that emerges between neurons during learning. This paper introduces a comprehensive quantum information framework for analyzing neural network representations, treating hidden layer activations as quantum states and characterizing them through density matrices, von Neumann entropy, purity, and related quantum information metrics. Through systematic experiments on synthetic classification tasks and algorithmic reasoning problems spanning over 3000 cumulative training epochs and 5000 Wang-Landau sampling iterations, we demonstrate that these quantum metrics reveal structure completely invisible to classical analysis. Our results show that stochastic gradient descent systematically reduces von Neumann entropy while increasing purity as networks learn, transitioning from highly mixed quantum states with entropy approximately 1.42 and purity approximately 0.29 to purer configurations with entropy approximately 1.19 and purity approximately 0.37 that achieve superior test accuracy of 99.5%. Using Wang-Landau sampling to explore the equilibrium entropy landscape, we discover that neural networks naturally favor high-entropy states with entropy approximately 1.42 and moderate accuracy of 62.0%, and that optimization actively purifies the quantum state to achieve task performance. On modular arithmetic tasks exhibiting the grokking phenomenon, we observe that von Neumann entropy increases dramatically from 1.55 to 3.29, representing a 112% increase, during an extended 2000-epoch memorization phase before any improvement in generalization, while purity decreases correspondingly from 0.444 to 0.042, a 90.5% decrease, approaching the maximally mixed limit of 0.01. This entropy accumulation phase appears necessary for eventual generalization, with the network building quantum correlations at a rate of approximately 0.00097 nats per epoch during the plateau. Weight entropy, in contrast, shows minimal change of only 11% increase over the same period. Correlation analysis reveals that von Neumann entropy and purity are almost perfectly anticorrelated with a correlation coefficient of 1-0.99, while weight entropy shows weak correlation with quantum metrics with absolute correlation coefficients less than 0.3. The effective rank of the density matrix stabilizes near 2.7 to 3.0 for the 6-neuron classifier, indicating that the network effectively uses approximately three independent activation modes. These findings establish quantum information metrics as powerful tools for understanding neural network learning dynamics, provide the first quantitative evidence for the entropy accumulation interpretation of grokking, and suggest that neural networks undergo a quantum-inspired phase transition during learning. We discuss implications for generalization theory, optimization algorithm design, and the statistical physics of deep learning, and outline a comprehensive research agenda for extending these methods to larger architectures and more complex tasks.
The remarkable generalization capabilities of overparameterized neural networks remain one of the most profound mysteries in modern machine learning. Despite having sufficient capacity to simply memorize training data, these networks consistently achieve superior performance on unseen test examples. Traditional analyses focusing on weight distributions, loss landscapes, and optimization dynamics have provided partial insights but fail to capture the complex correlational structure that emerges between neurons during learning. This paper introduces a comprehensive quantum information framework for analyzing neural network representations, treating hidden layer activations as quantum states and characterizing them through density matrices, von Neumann entropy, purity, and related quantum information metrics. Through systematic experiments on synthetic classification tasks and algorithmic reasoning problems spanning over 3000 cumulative training epochs and 5000 Wang-Landau sampling iterations, we demonstrate that these quantum metrics reveal structure completely invisible to classical analysis. Our results show that stochastic gradient descent systematically reduces von Neumann entropy while increasing purity as networks learn, transitioning from highly mixed quantum states with entropy approximately 1.42 and purity approximately 0.29 to purer configurations with entropy approximately 1.19 and purity approximately 0.37 that achieve superior test accuracy of 99.5%. Using Wang-Landau sampling to explore the equilibrium entropy landscape, we discover that neural networks naturally favor high-entropy states with entropy approximately 1.42 and moderate accuracy of 62.0%, and that optimization actively purifies the quantum state to achieve task performance. On modular arithmetic tasks exhibiting the grokking phenomenon, we observe that von Neumann entropy increases dramatically from 1.55 to 3.29, representing a 112% increase, during an extended 2000-epoch memorization phase before any improvement in generalization, while purity decreases correspondingly from 0.444 to 0.042, a 90.5% decrease, approaching the maximally mixed limit of 0.01. This entropy accumulation phase appears necessary for eventual generalization, with the network building quantum correlations at a rate of approximately 0.00097 nats per epoch during the plateau. Weight entropy, in contrast, shows minimal change of only 11% increase over the same period. Correlation analysis reveals that von Neumann entropy and purity are almost perfectly anticorrelated with a correlation coefficient of 1 -0.99, while weight entropy shows weak correlation with quantum metrics with absolute correlation coefficients less than 0.3. The effective rank of the density matrix stabilizes near 2.7 to 3.0 for the 6-neuron classifier, indicating that the network effectively uses approximately three independent activation modes. These findings establish quantum information metrics as powerful tools for understanding neural network learning dynamics, provide the first quantitative evidence for the entropy accumulation interpretation of grokking, and suggest that neural networks undergo a quantum-inspired phase transition during learning. We discuss implications for generalization theory, optimization algorithm design, and the statistical physics of deep learning, and outline a comprehensive research agenda for extending these methods to larger architectures and more complex tasks.
Abstract Analytically continuing the von Neumann entropy from Rényi entropies is a challenging task in quantum field theory. While the n -th Rényi entropy can be computed using the replica method in the path integral representation of quantum field theory, the analytic continuation can only be achieved for some simple systems on a case-by-case basis. In this work, we propose a general framework to tackle this problem using classical and quantum neural networks with supervised learning. We begin by studying several examples with known von Neumann entropy, where the input data is generated by representing TrρAn Tr ρ A n with a generating function. We adopt KerasTuner to determine the optimal network architecture and hyperparameters with limited data. In addition, we frame a similar problem in terms of quantum machine learning models, where the expressivity of the quantum models for the entanglement entropy as a partial Fourier series is established. Our proposed methods can accurately predict the von Neumann and Rényi entropies numerically, highlighting the potential of deep learning techniques for solving problems in quantum information theory.
Grokking, the delayed transition from memorization to generalization, is a fundamental phenomenon in gradient-based learning, yet its dynamics within variational quantum machine learning (QML) remain largely unexamined. In this work, we report the empirical observation of both the grokking transition and epoch-wise double descent in a two-qubit quantum neural network (QNN) under a complete parameterization of the SU(4) manifold. We demonstrate that overparameterization via increased circuit depth improves the probability of successful generalization. Notably, these architectures frequently exhibit an epoch-wise double descent in test error, degrading at a critical epoch before recovering into a generalizing state. Crucially, we identify a generalization decay in late-stage training, where the test error increases significantly despite a stagnant training loss. Bridging this behavior with algorithmic stability theory, our analysis reveals that this decay correlates with an unconstrained increase of the weight-norm, drifting away from sparse, phase-aligned harmonic solutions toward overfitted solutions in the Hilbert space. We analyze the underlying temporal dynamics of this transition, demonstrating how the onset of generalization is linked to optimization hyperparameters such as learning rate and weight decay. Finally, to mitigate late-stage decay, we introduce a weak explicit weight-norm regularization into the loss function. We demonstrate that this structural anchor stabilizes the post-grokking phase and permanently preserves generalization gains, providing a robust framework for training overparameterized quantum circuits.
Quantum machine learning is at the crossroads of two of the most exciting current areas of research: quantum computing and classical machine learning. Although the field is still in its infancy, the body of literature is already large enough to warrant several review articles [1–3]. This short survey focuses on a selection of significant recent results on the subtopic of quantum neural networks, an area that hopes to build on the enormous impact that classical neural networks have had. In particular, we concentrate on quantum generalizations of popular neural-network concepts such as Boltzmann machines, generative adversarial networks and autoencoders, as well as applications of classical neural networks to quantum information and vice versa. Examples of Boltzmann machines. In a fully connected Boltzmann machine, connections exist between all visible and hidden neurons. In a restricted Boltzmann machine (RBM), there are no connections between neurons of the same type. A deep Boltzmann machine (DBM) consists of adding one or more additional layers of hidden neurons to an RBM, with no connections between neurons in the same layer. The energy-based nature of BMs gives a natural framework for considering quantum generalizations of their behavior. The BM energy function is equivalent to the Hamiltonian of a simple Ising model and one might hope that more general Hamiltonians allowed by quantum mechanics could explain certain data sets better than classically. Amin et al. [4] adopt this approach, and take the so-called transverse field Ising Hamiltonian as the energy function. This increased generality brings certain difficulties, however, as the transverse field term makes training the quantum Boltzmann machine (QBM) non-trivial. Although the general scaling and performance of QBMs is not yet well understood, for certain small-size examples the authors observe that their QBMs can learn data distributions better than classical BMs. Kieferova and Wiebe [5] extend the formalism of Amin et al. and consider more general Hamiltonians that, unlike the transverse Ising model, are not in general believed to be classically efficiently simulatable. Furthermore, they propose new methods for training that allow for parameters corresponding to quantum terms to be learned. Finally, their QBM enables a form of quantum-state tomography that can learn non-classical data, and also generate copies of the reconstructed state. Again, small-scale numerical evidence suggests that their proposals may provide an advantage over classical BMs, although their tendency to overfit data must be investigated further. The practical value of BMs, either classical or quantum, relies in part on the ease with which they are trained. This in turn depends on sampling efficiently from the Boltzmann distribution. Classically this can be done approximately only for the special case of restricted Boltzmann machines (RBM), i.e. where connections are between visible and hidden units, and no connections exist between neurons of the same type. Wiebe et al. [6] propose two quantum algorithms for Boltzmann sampling, based on starting from a state corresponding to the product distribution that is closest to the target Boltzmann distribution. While small-scale numerical studies are encouraging, true bounds on the running time of their algorithms rely on a problem-dependent parameter κ, the scaling of which is hard to estimate. The hope remains though that empirically their algorithms work well in practice, even for multi-layer and fully connected BMs. The expressive power of BMs for compactly representing complicated probability distributions has led to another line of investigation, namely their use for representing quantum many-body wave functions. The authors apply this technique to two prototypical spin models and achieve high accuracy in both ground-state estimation and dynamical evolution, improving on existing state-of-the-art techniques. Their results indicate that RBM states can capture complex many-body correlations in an efficient manner—evidence reinforced by the work of Deng et al. [8], who show that certain topological quantum states in one, two and three dimensions can be represented by RBM states with a number of parameters only linear in the system size. In a subsequent study, however, Gao and Duan [9] prove that although RBM states can indeed represent many highly entangled states, their expressive power is limited in that they cannot represent arbitrary states generated by polynomial-sized quantum circuits, nor those that are the ground states of k-local Hamiltonians with polynomial-sized energy gaps, unless the polynomial hierarchy from computational complexity theory collapses. On the other hand, they show that these two important classes of states can be efficiently represented by a deep Boltzmann machine of three layers. Intriguingly though, another paper by Deng et al. [10] gives an analytic construction of a family of RBM states with a number of parameters linear in the system size that exhibits maximal volume-law entanglement, and that cannot be represented efficiently by either matrix product states (MPS) or tensor-network (TN) states. This shows that the expressive power of even efficiently parametrized RBM states is not limited by entanglement in the same way that other popular compact representations are. This gives the hope that RBM states may prove a valuable tool in understanding entangled quantum systems intractable by traditional methods. Levine et al. [11] establish another connection between the many-body wave function and neural networks, this time in the form of convolutional arithmetic circuits (ConvAC), a variant of the popular convolutional neural networks in widespread use today. By proving a one-to-one equivalence between a ConvAC circuit and TN quantum states, the authors are able to use tools from quantum information theory to analyze the entanglement present in the corresponding TN, and quantify the neural network's ability to model complex correlations between its inputs. This in turn can be used to obtain bounds on the expressivity of a given ConvAC architecture and be of practical use in choosing network parameters tailored to the data. While ConvAC are not widely used in practice, the authors empirically validate that the insights gained can be applied successfully to more common convolutional neural-network designs. Generative adversarial networks (GAN) are machine-learning algorithms in which two players—a generator and a discriminator—compete against each another. The generator's goal is to create and send fake data samples to the discriminator and maximize the probability that the discriminator erroneously believes that the sample comes from some true data distribution. The discriminator attempts to discern real from fake data as best as possible. Under reasonable assumptions the game converges to a point where the discriminator is unable to discriminate with probability greater than 1/2 (i.e. no better than random guessing), and the generator in fact has learned to generate data from the true distribution. Lloyd and Weedbrook [12] consider the natural extension in which one or both of the generator and discriminator are equipped with quantum information processors. When the data to be learned are a quantum state and both parties have quantum processors, the authors show that the unique end point of the adversarial game also coincides with the generator producing a quantum state corresponding to the true data statistics, and the discriminator unable to discern true from false data better than random guessing. They further argue that a different outcome may hold when the true samples are produced by the measurement of some quantum state, and the generator is purely classical. The concept of quantum supremacy is based on the assumption that there exist probability distributions accessible to a quantum computer that no classical computer can efficiently sample from. If the true data correspond to one of these distributions, then no classical generator can efficiently generate fake data reproducing the true statistics and, in principle, the discriminator can successfully discriminate true from fake data. However, whether the discriminator can efficiently find the measurements required to do so is unclear. Quantum-computing techniques have also found applications in autoencoders—neural networks where a larger number of input and output neurons are connected via a smaller number of intermediary latent neurons. The goal is to tune the weights of the network so that a training set of classical bit strings can be approximately recovered after being passed through the autoencoder. Romero et al. [13] consider instead a training set of quantum states, and ask whether a quantum circuit with a similar bottleneck structure can compress and approximately recover the data. The circuits that they consider consist of gates chosen from some set of parametrized unitary operations, and arranged in some fixed order. The parameters are then optimized classically so that the circuit best predicts the training-set labels. The authors demonstrate the validity of this approach by showing that ground states of the Hubbard model and certain molecules can indeed be compressed and recovered by such circuits. However, in general, the performance of such quantum autoencoders will be highly dependent on the kind of programmable circuit on which they are based, and at present there does not seem to be a structured way of determining this. Quantum neural networks are still in their infancy and, while the above works are important steps towards bridging our understanding between classical neural networks and quantum computing, many questions remain unanswered. If the expressivity of Boltzmann machine states is not restricted by entanglement in the same way that MPS and tensor-network states are, what is the best way of understanding their limitations? How can physicists use the correspondence between ConvAC and TN to their advantage in finding efficient, expressive representations of many-body wave functions? Can classical data distributions be learned more efficiently by quantum GAN than classical GAN? Whatever the outcome of these and many other related questions, this promises to be an exciting and active area of research in the years to come.
Abstract Quantum Convolutional Neural Networks(QCNN) are a promising framework in QuantumMachine Learning(QML). The study analyzes QCNNs’ expressibility, trainability, and optimizers theoretically. This approach utilizes statistical learning and quantum information theory to assess QCNNs’ effectiveness in imitating quantum states, optimizing parameters, and executing intricate machine learning(ML) tasks. For this purpose, the Haar measure is used to evaluate the QCNN expressibility, yielding a score of 0.7720, acoverage entropy of 1.4023, and an effective state-space dimension of 166.59/200. This indicates that QCNNs can simulate massive quantum states and achieve a peak expressibility score of 0.8120 at layer 3, before decreasing due to overparameterization. Additionally, we examine gradient variance and find that the barren plateau indicator is 0.0818 and the mean gradient norm is 1.1837, both of which are below the theoretical threshold of 0.1, proving that the QCNN can avoid barren plateaus and train perfectly. Furthermore, the obtained Fisher Information Matrix(FIM) condition number of 2.05 ×10³ indicates a well-conditioned optimization landscape. We also utilized the Hessian matrix for optimization analysis, identifying a saddle point with 56 zero, 11 positive,and 5 negative eigenvalues. Where QCNNs required only 72 parameters, an 11.8-fold reduction from classical CNNs’ 848 parameters, but expressibility stays the same, assuming smooth convergence. Despite a gradient noise factor near 1, QCNNs have only a 43% expressibility decrease at 5% noise levels. Finally, the outcomes prove that QCNNs outperform classical CNNs in ML applications with respect to expressibility, trainability, and optimization.
Neural quantum states (NQS) provide flexible and compact wave-function parametrizations for numerical studies of quantum many-body physics. In particular, NQS aim to circumvent the exponential scaling of the Hilbert space by compressing quantum many-body wave functions with a tractable amount of parameters. While inspired by deep learning, it remains unclear to what extent NQS share characteristics with neural networks used for standard machine learning tasks. We demonstrate that, in a simplified supervised setting, NQS exhibit the double descent phenomenon, a key feature of modern deep learning, where generalization worsens as network size increases before improving again in an overparameterized regime. Notably, we find the second descent to occur only for network sizes much larger than the Hilbert space dimension, i.e., network sizes that are out of reach for problems of practical interest. Within our setting, this observation places typical NQS in the underparameterized regime. We also observe that the optimal network size in the underparameterized regime depends on the number of unique training samples. While the double descent phenomenon does indeed translate to the NQS setting, potential practical consequences of our findings point more toward the need for symmetry-aware, physics-informed architecture design, rather than directly adopting machine learning heuristics.