Hao Yu

Xiao Liu, Hao Yu, Hanchen Zhang et al. · berkeley, microsoft-research

The potential of Large Language Model (LLM) as agents has been widely acknowledged recently. Thus, there is an urgent need to quantitatively \textit{evaluate LLMs as agents} on challenging tasks in interactive environments. We present AgentBench, a multi-dimensional benchmark that consists of 8 distinct environments to assess LLM-as-Agent's reasoning and decision-making abilities. Our extensive test over \num API-based and open-sourced (OSS) LLMs shows that, while top commercial LLMs present a strong ability of acting as agents in complex environments, there is a significant disparity in performance between them and many OSS competitors that are no larger than 70B. We identify the typical reasons of failures in environments and LLMs, showing that poor long-term reasoning, decision-making, and instruction following abilities are the main obstacles for developing usable LLM agents. Improving instruction following and training on high quality multi-round alignment data could improve agent performance. And different from existing assumptions, training on code present ambivalent impacts on different agent tasks. Datasets, environments, and an integrated evaluation package for AgentBench are released at https://github.com/THUDM/AgentBench.

Wenquan Dong, Edward T. A. Mitchard, Hao Yu et al.

Accurate quantification of forest aboveground biomass (AGB) is critical for understanding carbon accounting in the context of climate change. In this study, we presented a novel attention-based deep learning approach for forest AGB estimation, primarily utilizing openly accessible EO data, including: GEDI LiDAR data, C-band Sentinel-1 SAR data, ALOS-2 PALSAR-2 data, and Sentinel-2 multispectral data. The attention UNet (AU) model achieved markedly higher accuracy for biomass estimation compared to the conventional RF algorithm. Specifically, the AU model attained an R2 of 0.66, RMSE of 43.66 Mg ha-1, and bias of 0.14 Mg ha-1, while RF resulted in lower scores of R2 0.62, RMSE 45.87 Mg ha-1, and bias 1.09 Mg ha-1. However, the superiority of the deep learning approach was not uniformly observed across all tested models. ResNet101 only achieved an R2 of 0.50, an RMSE of 52.93 Mg ha-1, and a bias of 0.99 Mg ha-1, while the UNet reported an R2 of 0.65, an RMSE of 44.28 Mg ha-1, and a substantial bias of 1.84 Mg ha-1. Moreover, to explore the performance of AU in the absence of spatial information, fully connected (FC) layers were employed to eliminate spatial information from the remote sensing data. AU-FC achieved intermediate R2 of 0.64, RMSE of 44.92 Mgha-1, and bias of -0.56 Mg ha-1, outperforming RF but underperforming AU model using spatial information. We also generated 10m forest AGB maps across Guangdong for the year 2019 using AU and compared it with that produced by RF. The AGB distributions from both models showed strong agreement with similar mean values; the mean forest AGB estimated by AU was 102.18 Mg ha-1 while that of RF was 104.84 Mg ha-1. Additionally, it was observed that the AGB map generated by AU provided superior spatial information. Overall, this research substantiates the feasibility of employing deep learning for biomass estimation based on satellite data.

Man Chen, Wenquan Dong, Hao Yu et al.

The integration of multisource remote sensing data and deep learning models offers new possibilities for accurately mapping high spatial resolution forest height. We found that GEDI relative heights (RH) metrics exhibited strong correlation with the mean of the top 10 highest trees (dominant height) measured in situ at the corresponding footprint locations. Consequently, we proposed a novel deep learning framework termed the multi-modal attention remote sensing network (MARSNet) to estimate forest dominant height by extrapolating dominant height derived from GEDI, using Setinel-1 data, ALOS-2 PALSAR-2 data, Sentinel-2 optical data and ancillary data. MARSNet comprises separate encoders for each remote sensing data modality to extract multi-scale features, and a shared decoder to fuse the features and estimate height. Using individual encoders for each remote sensing imagery avoids interference across modalities and extracts distinct representations. To focus on the efficacious information from each dataset, we reduced the prevalent spatial and band redundancies in each remote sensing data by incorporating the extended spatial and band reconstruction convolution modules in the encoders. MARSNet achieved commendable performance in estimating dominant height, with an R2 of 0.62 and RMSE of 2.82 m, outperforming the widely used random forest approach which attained an R2 of 0.55 and RMSE of 3.05 m. Finally, we applied the trained MARSNet model to generate wall-to-wall maps at 10 m resolution for Jilin, China. Through independent validation using field measurements, MARSNet demonstrated an R2 of 0.58 and RMSE of 3.76 m, compared to 0.41 and 4.37 m for the random forest baseline. Our research demonstrates the effectiveness of a multimodal deep learning approach fusing GEDI with SAR and passive optical imagery for enhancing the accuracy of high resolution dominant height estimation.

Hao Yu, Rong Jin

For SGD based distributed stochastic optimization, computation complexity, measured by the convergence rate in terms of the number of stochastic gradient calls, and communication complexity, measured by the number of inter-node communication rounds, are two most important performance metrics. The classical data-parallel implementation of SGD over $N$ workers can achieve linear speedup of its convergence rate but incurs an inter-node communication round at each batch. We study the benefit of using dynamically increasing batch sizes in parallel SGD for stochastic non-convex optimization by charactering the attained convergence rate and the required number of communication rounds. We show that for stochastic non-convex optimization under the P-L condition, the classical data-parallel SGD with exponentially increasing batch sizes can achieve the fastest known $O(1/(NT))$ convergence with linear speedup using only $\log(T)$ communication rounds. For general stochastic non-convex optimization, we propose a Catalyst-like algorithm to achieve the fastest known $O(1/\sqrt{NT})$ convergence with only $O(\sqrt{NT}\log(\frac{T}{N}))$ communication rounds.

Hao Yu, Rong Jin, Sen Yang

Recent developments on large-scale distributed machine learning applications, e.g., deep neural networks, benefit enormously from the advances in distributed non-convex optimization techniques, e.g., distributed Stochastic Gradient Descent (SGD). A series of recent works study the linear speedup property of distributed SGD variants with reduced communication. The linear speedup property enable us to scale out the computing capability by adding more computing nodes into our system. The reduced communication complexity is desirable since communication overhead is often the performance bottleneck in distributed systems. Recently, momentum methods are more and more widely adopted in training machine learning models and can often converge faster and generalize better. For example, many practitioners use distributed SGD with momentum to train deep neural networks with big data. However, it remains unclear whether any distributed momentum SGD possesses the same linear speedup property as distributed SGD and has reduced communication complexity. This paper fills the gap by considering a distributed communication efficient momentum SGD method and proving its linear speedup property.

Hao Yu, Sen Yang, Shenghuo Zhu

In distributed training of deep neural networks, parallel mini-batch SGD is widely used to speed up the training process by using multiple workers. It uses multiple workers to sample local stochastic gradient in parallel, aggregates all gradients in a single server to obtain the average, and update each worker's local model using a SGD update with the averaged gradient. Ideally, parallel mini-batch SGD can achieve a linear speed-up of the training time (with respect to the number of workers) compared with SGD over a single worker. However, such linear scalability in practice is significantly limited by the growing demand for gradient communication as more workers are involved. Model averaging, which periodically averages individual models trained over parallel workers, is another common practice used for distributed training of deep neural networks since (Zinkevich et al. 2010) (McDonald, Hall, and Mann 2010). Compared with parallel mini-batch SGD, the communication overhead of model averaging is significantly reduced. Impressively, tremendous experimental works have verified that model averaging can still achieve a good speed-up of the training time as long as the averaging interval is carefully controlled. However, it remains a mystery in theory why such a simple heuristic works so well. This paper provides a thorough and rigorous theoretical study on why model averaging can work as well as parallel mini-batch SGD with significantly less communication overhead.