arXiv subjects

Zhendong Yu

Publications and source records attributed to Zhendong Yu.

6 recordsLinked to original sources

arXiv2026

Joint Training on AMD and NVIDIA GPUs

As large language models continue to scale, training demands on compute and system capacity grow rapidly, making single-vendor homogeneous clusters insufficient. This paper presents a technical solution for heterogeneous mixed training in AMD-NVIDIA environments. We first adopt a compatibility-oriented approach based on CPU-Forwarding Communication, with differentiated communication back-end selection across parallel groups and multi-NIC parallel data transfer. To achieve higher performance, we further propose another Device-Direct Communication approach, integrating a CPU-offloading P2P mechanism to enable direct cross-vendor GPU data transfer without host-memory staging. Experiments on LLaMA-8B and Qwen2-7B demonstrate that the proposed Device-Direct Communication approach achieves up to 98% of the throughput of an NVIDIA homogeneous system, while preserving training stability and correctness.

cs.DC↗

arXiv2025

Neural Langevin Machine: a local asymmetric learning rule can be creative

Fixed points of recurrent neural networks can be leveraged to store and generate information. These fixed points can be captured by the Boltzmann-Gibbs measure, which leads to neural Langevin dynamics that can be used to find them for generative learning of a real dataset. We call this type of generative model a neural Langevin machine, which derives an asymmetric and firing-rate-speed adjusted learning rule requiring only local neural signals, thereby bearing biological relevance in terms of local predictive learning. An interesting out-of-equilibrium regime of the generative process is revealed, together with a memorization-to-generalization transition with increasing training data size. The neuro-inspired machine can also realize a continuous exploration of the phase space for different kinds of generative images and can denoise a corrupted image as well.

q-bio.NC↗

arXiv2025

Precision Neural Network Quantization via Learnable Adaptive Modules

Quantization Aware Training (QAT) is a neural network quantization technique that compresses model size and improves operational efficiency while effectively maintaining model performance. The paradigm of QAT is to introduce fake quantization operators during the training process, allowing the model to autonomously compensate for information loss caused by quantization. Making quantization parameters trainable can significantly improve the performance of QAT, but at the cost of compromising the flexibility during inference, especially when dealing with activation values with substantially different distributions. In this paper, we propose an effective learnable adaptive neural network quantization method, called Adaptive Step Size Quantization (ASQ), to resolve this conflict. Specifically, the proposed ASQ method first dynamically adjusts quantization scaling factors through a trained module capable of accommodating different activations. Then, to address the rigid resolution issue inherent in Power of Two (POT) quantization, we propose an efficient non-uniform quantization scheme. We utilize the Power Of Square root of Two (POST) as the basis for exponential quantization, effectively handling the bell-shaped distribution of neural network weights across various bit-widths while maintaining computational efficiency through a Look-Up Table method (LUT). Extensive experimental results demonstrate that the proposed ASQ method is superior to the state-of-the-art QAT approaches. Notably that the ASQ is even competitive compared to full precision baselines, with its 4-bit quantized ResNet34 model improving accuracy by 1.2\% on ImageNet.

cs.CV↗

arXiv2024

Network reconstruction may not mean dynamics prediction

With an increasing amount of observations on the dynamics of many complex systems, it is required to reveal the underlying mechanisms behind these complex dynamics, which is fundamentally important in many scientific fields such as climate, financial, ecological, and neural systems. The underlying mechanisms are commonly encoded into network structures, e.g., capturing how constituents interact with each other to produce emergent behavior. Here, we address whether a good network reconstruction suggests a good dynamics prediction. The answer is quite dependent on the nature of the supplied (observed) dynamics sequences measured on the complex system. When the dynamics are not chaotic, network reconstruction implies dynamics prediction. In contrast, even if a network can be well reconstructed from the chaotic time series (chaos means that many unstable dynamics states coexist), the prediction of the future dynamics can become impossible as at some future point the prediction error will be amplified. This is explained by using dynamical mean-field theory on a toy model of random recurrent neural networks.

q-bio.NC↗

arXiv2024

Nonequilbrium physics of generative diffusion models

Generative diffusion models apply the concept of Langevin dynamics in physics to machine leaning, attracting a lot of interests from engineering, statistics and physics, but a complete picture about inherent mechanisms is still lacking. In this paper, we provide a transparent physics analysis of diffusion models, formulating the fluctuation theorem, entropy production, equilibrium measure, and Franz-Parisi potential to understand the dynamic process and intrinsic phase transitions. Our analysis is rooted in a path integral representation of both forward and backward dynamics, and in treating the reverse diffusion generative process as a statistical inference, where the time-dependent state variables serve as quenched disorder akin to that in spin glass theory. Our study thus links stochastic thermodynamics, statistical inference and geometry based analysis together to yield a coherent picture about how the generative diffusion models work.

cond-mat.stat-mech↗

arXiv2020

Adversarial Code Learning for Image Generation

We introduce the "adversarial code learning" (ACL) module that improves overall image generation performance to several types of deep models. Instead of performing a posterior distribution modeling in the pixel spaces of generators, ACLs aim to jointly learn a latent code with another image encoder/inference net, with a prior noise as its input. We conduct the learning in an adversarial learning process, which bears a close resemblance to the original GAN but again shifts the learning from image spaces to prior and latent code spaces. ACL is a portable module that brings up much more flexibility and possibilities in generative model designs. First, it allows flexibility to convert non-generative models like Autoencoders and standard classification models to decent generative models. Second, it enhances existing GANs' performance by generating meaningful codes and images from any part of the prior. We have incorporated our ACL module with the aforementioned frameworks and have performed experiments on synthetic, MNIST, CIFAR-10, and CelebA datasets. Our models have achieved significant improvements which demonstrated the generality for image generation tasks.

cs.CV↗