train hardware generative models

This work addresses the limited expressivity of modern analog hardware, which is constrained by fixed differential equations and lags far behind software-defined generative models. To bridge this gap, the authors propose the Analog Interaction Systems (AIS) framework, which substantially enhances the representational capacity of analog dynamical systems through time-segmented tunable parameters and latent physical states. By integrating a Wasserstein GAN training strategy, AIS enables end-to-end trainable generative modeling without requiring trajectory alignment. This study presents the first systematic quantification of the expressivity gap between analog systems and neural networks and introduces a hardware-compatible mechanism to close it. On MNIST and Fashion-MNIST, the model achieves FID scores of 27.6 and 80.8, respectively—outperforming prior analog generative models by 3–4×—while consuming only 23 microjoules per image, offering two orders of magnitude energy savings over digital counterparts.

This study addresses the high computational cost and poor safety controllability in training ultra-large AI models. Methodologically, it introduces a novel optimization paradigm grounded in mechanistic interpretability, pioneering the application of “circuit analysis” to model gradient descent trajectories. It structures the parameter space into functional subnetworks and designs a progressive curriculum learning strategy to dynamically regulate optimization paths within a controlled environment. Key contributions include: (1) establishing a formal mapping between gradient flow dynamics and circuit-like structural representations, enabling interpretable modeling of optimization behavior; and (2) leveraging structural priors to guide curriculum design, significantly accelerating convergence while suppressing the emergence of harmful behaviors. Experiments across multiple benchmark tasks demonstrate over 30% reduction in training cost alongside improved behavioral controllability, offering a principled pathway toward efficient and safe large-model training.

This paper addresses core challenges in applying GANs to artistic generation—namely, difficulty in high-resolution modeling, training instability, and weak cross-modal adaptability—by systematically constructing a theory–application co-driven generative framework. Methodologically, it pioneers the integration of adversarial principles with multiple stabilization techniques (WGAN-GP, spectral normalization, self-attention) and task-specific artistic strategies, while incorporating comparative analysis with diffusion models and Transformer architectures. Extensive experiments evaluate mainstream variants—including DCGAN, InfoGAN, LAPGAN, and LSGAN—across diverse scenarios: high-resolution image synthesis, cross-domain style transfer, video generation, and text-to-image translation. The framework delivers consistently high-fidelity outputs and provides a fully reproducible technical pipeline. Results demonstrate substantial improvements in robustness and expressive capability of GANs for creative computing, establishing a foundational methodology for both AI-driven art research and industrial deployment.

Generative machine learning for design optimization faces dual bottlenecks: heavy reliance on large-scale labeled data and poor interpretability. Addressing these challenges in flow battery manifold design, this paper proposes a synergistic GAN–Bayesian optimization framework. It constructs a hybrid input prototype set integrating homogeneous structural constraints and heterogeneous physical features, and jointly optimizes the generative model’s latent space with a Bayesian surrogate model to enhance both interpretability and task-directedness of design representations. Experiments demonstrate that all generated designs strictly satisfy engineering feasibility constraints, cover over 92% of the critical performance domain, improve the Pareto front on the pressure-drop–uniformity trade-off curve by 37%, and enable physics-informed attribution analysis. This work establishes a novel paradigm for sample-efficient, high-reliability, and interpretable design generation.

Analog IC topology design remains heavily reliant on manual expertise, resulting in low efficiency and limited creativity. Method: This paper introduces the first generative AI framework for analog circuit topology synthesis. We construct the first large-scale analog circuit topology dataset, propose a unified, cross-circuit-type serialized graph representation, and design an analog-specific generative engine integrating graph neural networks with topology-aware encoding. Contributions/Results: The framework enables fully automated and scalable topology generation. It significantly improves topology diversity (+327%) and device density (+215%). Moreover, it synthesizes numerous high-performance, previously unreported circuit topologies, increasing the count of Pareto-optimal solutions in benchmark tasks by 4.8×. This work establishes a new paradigm for analog IC design—shifting from experience-driven to data- and model-coordinated automation.

This study addresses the challenge of evaluating fault tolerance in digital circuits under failure modes such as missing or misaligned logic gates. The authors propose a novel numerical approach based on generative adversarial networks (GANs), introducing for the first time a GAN architecture that employs complex-valued representations. This framework generates bit-level current configurations corresponding to ideal circuit behavior and compares them against actual signal responses to quantify output deviations, enabling fine-grained fault tolerance analysis. Experimental results demonstrate that the method effectively distinguishes and precisely measures the impact of various fault types on circuit outputs, significantly enhancing both the efficiency and accuracy of robustness assessment in electronic design.

This work addresses the limitation of conventional generative adversarial networks (GANs), which rely on the independent and identically distributed (i.i.d.) assumption and thus struggle with deterministic, non-random time series generated by chaotic dynamical systems. The authors propose a novel framework based on infinite-dimensional generative adversarial learning that, for the first time, provides theoretical guarantees for GANs in a non-i.i.d. setting. Specifically, they demonstrate that the invariant distribution of a chaotic system can be recovered from a single deterministic trajectory, establishing an explicit convergence rate in terms of Jensen–Shannon divergence. By integrating concepts from chaotic dynamical systems, statistical learning theory, and infinite-dimensional GAN analysis, this study rigorously establishes the feasibility of learning invariant measures from a single sample path and quantifies the convergence speed of generative models toward the true underlying distribution.

This work presents the first systematic survey and critical analysis of neural architecture search (NAS) methods tailored for generative adversarial networks (GANs), addressing the inefficiency and instability of manual GAN design, which often struggles to balance performance and generalization. The study organizes existing approaches through a structured comparison based on search strategies, evaluation metrics, and empirical performance. It advocates moving beyond conventional Inception Score (IS) and Fréchet Inception Distance (FID) toward more robust evaluation frameworks and diverse datasets. The analysis further highlights the complementary strengths of evolutionary algorithms and gradient-based methods across different scenarios. By clarifying the current limitations and untapped potential of NAS-GAN methodologies, this work establishes a foundation for future research and advances the standardization and performance of automated GAN architecture design.

This work addresses the underutilization of exponential moving average (EMA) generators in traditional GAN training, where they are employed only during inference and thus fail to leverage their inherent stability during optimization. To close this gap, the authors propose Self-Distilled GAN (SD-GAN), which uniquely integrates the EMA generator as a teacher model that guides the student generator via perceptual loss throughout training, thereby enabling end-to-end exploitation of EMA’s stabilizing properties. Theoretical analysis demonstrates that SD-GAN achieves local asymptotic stability under the Dirac-GAN setting, effectively suppressing parasitic oscillations. Extensive experiments show consistent improvements in image generation quality—particularly in FID and random-FID—across diverse architectures and datasets, along with smoother optimization trajectories. Moreover, the method proves effective for fine-tuning pretrained GANs.

This work addresses the challenges faced by physics-informed neural networks (PINNs) in solving differential equations—particularly spectral bias, stiffness, and insufficient accuracy for multiscale solutions. From the perspective of the neural tangent kernel (NTK), the study establishes the first theoretical framework to systematically analyze how the discriminator in adversarial training influences PINN dynamics, clarifies the conditions under which such approaches are effective, and provides a unified interpretation of various GAN variants within the PINN context. Building on this theory, the authors propose an efficient training algorithm that substantially mitigates ill-conditioning during optimization, achieving accuracy improvements of several orders of magnitude over existing methods across multiple benchmark problems.