Shaolei Ren

The growing carbon footprint of artificial intelligence (AI) models, especially large ones such as GPT-3, has been undergoing public scrutiny. Unfortunately, however, the equally important and enormous water (withdrawal and consumption) footprint of AI models has remained under the radar. For example, training GPT-3 in Microsoft’s state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater, but such information has been kept a secret. More critically, the global AI demand might be accountable for 4.2 – 6.6 billion cubic meters of water withdrawal in 2027, which is more than the total annual water withdrawal of 4 – 6 Denmark or half of the United Kingdom. This is very concerning, as freshwater scarcity has become one of the most pressing challenges shared by all of us in the wake of the rapidly growing population, depleting water resources, and aging water infrastructures. To respond to the global water challenges, AI models can, and also must, take social responsibility and lead by example by addressing their own water footprint. In this paper, we provide a principled methodology to estimate the water footprint of AI models, and also discuss the unique spatial-temporal diversities of AI models’ runtime water efficiency. Finally, we highlight the necessity of holistically addressing water footprint along with carbon footprint to enable truly sustainable AI.

This paper studies learning-augmented decentralized online convex optimization in a networked multi-agent system, a challenging setting that has remained under-explored. We first consider a linear learning-augmented decentralized online algorithm (LADO-Lin) that combines a machine learning (ML) policy with a baseline expert policy in a linear manner. We show that, while LADO-Lin can exploit the potential of ML predictions to improve the average cost performance, it cannot have guaranteed worst-case performance. To address this limitation, we propose a novel online algorithm (LADO) that adaptively combines the ML policy and expert policy to safeguard the ML predictions to achieve strong competitiveness guarantees. We also prove the average cost bound for LADO, revealing the tradeoff between average performance and worst-case robustness and demonstrating the advantage of training the ML policy by explicitly considering the robustness requirement. Finally, we run an experiment on decentralized battery management for sustainable computing. Our results highlight the potential of ML augmentation to improve the average performance as well as the guaranteed worst-case performance of LADO.

Text watermarks for large language models (LLMs) have been commonly used to identify the origins of machine-generated content, which is promising for assessing liability when combating deepfake or harmful content. While existing watermarking techniques typically prioritize robustness against removal attacks, unfortunately, they are vulnerable to spoofing attacks: malicious actors can subtly alter the meanings of LLM-generated responses or even forge harmful content, potentially misattributing blame to the LLM developer. To overcome this, we introduce a bi-level signature scheme, Bileve, which embeds fine-grained signature bits for integrity checks (mitigating spoofing attacks) as well as a coarse-grained signal to trace text sources when the signature is invalid (enhancing detectability) via a novel rank-based sampling strategy. Compared to conventional watermark detectors that only output binary results, Bileve can differentiate 5 scenarios during detection, reliably tracing text provenance and regulating LLMs. The experiments conducted on OPT-1.3B and LLaMA-7B demonstrate the effectiveness of Bileve in defeating spoofing attacks with enhanced detectability.

Public models have played a crucial role in various AI applications, showcasing their proficiency in accurate predictions. However, their exclusive emphasis on prediction accuracy may not align with the diverse end objectives of downstream agents when utilized. Recognizing the public model’s predictions as a service, we advocate for integrating the objectives of downstream agents into the optimization process. Concretely, to address performance disparities and foster fairness among heterogeneous agents in training, we propose a novel q-Equitable objective. This objective, coupled with a policy gradient algorithm, is crafted to train the public model to produce a more equitable/uniform performance distribution across downstream agents, each with their unique concerns. Both theoretical analysis and empirical case studies have proven the effectiveness of our method in advancing performance equity across diverse downstream agents utilizing the public model for their decision-making.

This paper studies online resource allocation with replenishable budgets, where budgets can be replenished on top of the initial budget and an agent sequentially chooses online allocation decisions without violating the available budget constraint at each round. We propose a novel online algorithm, called OACP (Opportunistic Allocation with Conservative Pricing), that conservatively adjusts dual variables while opportunistically utilizing available resources. OACP achieves a bounded asymptotic competitive ratio in adversarial settings as the number of decision rounds 𝑇 gets large. Importantly, the asymptotic competitive ratio of OACP is optimal in the absence of additional assumptions on budget replenishment. To further improve the competitive ratio, we make a mild assumption that there is budget replenishment every 𝑇*≥ 1 decision rounds and propose OACP+ to dynamically adjust the total budget assignment for online allocation. Next, we move beyond the worst-case and propose LA-OACP (Learning-Augmented OACP/OACP+), a novel learning-augmented algorithm for online allocation with replenishable budgets. We prove that LA-OACP can improve the average utility compared to OACP/OACP+ when the ML predictor is properly trained, while still offering worst-case utility guarantees when the ML predictions are arbitrarily wrong. Finally, we run simulation studies of sustainable AI inference powered by renewables, validating our analysis and demonstrating the empirical benefits of LA-OACP.

Online Budgeted Matching (OBM) is a classic problem with important applications in online advertising, online service matching, revenue management, and beyond. Traditional online algorithms typically assume a small bid setting, where the maximum bid-to-budget ratio (kappa) is infinitesimally small. While recent algorithms have tried to address scenarios with non-small or general bids, they often rely on the Fractional Last Matching (FLM) assumption, which allows for accepting partial bids when the remaining budget is insufficient. This assumption, however, does not hold for many applications with indivisible bids. In this paper, we remove the FLM assumption and tackle the open problem of OBM with general bids. We first establish an upper bound of 1-kappa on the competitive ratio for any deterministic online algorithm. We then propose a novel meta algorithm, called MetaAd, which reduces to different algorithms with first known provable competitive ratios parameterized by the maximum bid-to-budget ratio kappa in [0,1]. As a by-product, we extend MetaAd to the FLM setting and get provable competitive algorithms. Finally, we apply our competitive analysis to the design learning-augmented algorithms.

Fueled by the soaring popularity of large language and foundation models, the accelerated growth of artificial intelligence (AI) models’ enormous environmental footprint has come under increased scrutiny. While many approaches have been proposed to make AI more energy-efficient and environmentally friendly, environmental inequity – the fact that AI’s environmental footprint can be disproportionately higher in certain regions than in others – has emerged, raising social-ecological justice concerns. This paper takes a first step toward addressing AI’s environmental inequity by balancing its regional negative environmental impact. Concretely, we focus on the carbon and water footprints of AI model inference and propose equity-aware geographical load balancing (GLB) to explicitly address AI’s environmental impacts on the most disadvantaged regions. We run trace-based simulations by considering a set of 10 geographically-distributed data centers that serve inference requests for a large language AI model. The results demonstrate that existing GLB approaches may amplify environmental inequity while our proposed equity-aware GLB can significantly reduce the regional disparity in terms of carbon and water footprints.

Freshwater scarcity is a global problem that requires collective efforts across all industry sectors. Nevertheless, a lack of access to operational water footprint data bars many applications from exploring optimization opportunities hidden within the temporal and spatial variations. To break this barrier into research in water sustainability, we build a dataset for operation direct water usage in the cooling systems and indirect water embedded in electricity generation. Our dataset consists of the hourly water efficiency of major U.S. cities and states from 2018 to 2023. We also offer cooling system models that capture the impact of weather on water efficiency. We present a preliminary analysis of our dataset and discuss three potential applications that can benefit from it.

The enormous growth of AI computing has led to a surging demand for electricity. To stem the resulting energy cost and environmental impact, this paper explores opportunities enabled by the increasing hardware heterogeneity and introduces the concept of Geographical Server Relocation (GSR). Specifically, GSR physically balances the available AI servers across geographically distributed data centers subject to AI computing demand and power capacity constraints in each location. The key idea of GSR is to relocate older and less energy-efficient servers to regions with more renewables, better water efficiencies and/or lower electricity prices. Our case study demonstrates that, even with modest flexibility of relocation, GSR can substantially reduce the total operational environmental footprints and operation costs of AI computing. We conclude this paper by discussing the major challenges of GSR, including service migration, software management, and algorithms.

The sharply increasing sizes of artificial intelligence (AI) models come with significant energy consumption and environmental footprints, which can disproportionately impact certain (often marginalized) regions and hence create environmental inequity concerns. Moreover, concerns with social inequity have also emerged, as AI computing resources may not be equitably distributed across the globe and users from certain disadvantaged regions with severe resource constraints can consistently experience inferior model performance. Importantly, the inequity concerns that encompass both social and environmental dimensions still remain unexplored and have increasingly hindered responsible AI. In this paper, we leverage the spatial flexibility of AI inference workloads and propose equitable geographical load balancing (GLB) to fairly balance AI’s regional social and environmental costs. Concretely, to penalize the disproportionately high social and environmental costs for equity, we introduce Lq norms as novel regularization terms into the optimization objective for GLB decisions. Our empirical results based on real-world AI inference traces demonstrate that while the existing GLB algorithms result in disproportionately large social and environmental costs in certain regions, our proposed equitable GLB can fairly balance AI’s negative social and environmental costs across all the regions.

This paper studies the problem of Anytime-Competitive Markov Decision Process (A-CMDP). Existing works on Constrained Markov Decision Processes (CMDPs) aim to optimize the expected reward while constraining the expected cost over random dynamics, but the cost in a specific episode can still be unsatisfactorily high. In contrast, the goal of A-CMDP is to optimize the expected reward while guaranteeing a bounded cost in each round of any episode against a policy prior. We propose a new algorithm, called Anytime-Competitive Reinforcement Learning (ACRL), which provably guarantees the anytime cost constraints. The regret analysis shows the policy asymptotically matches the optimal reward achievable under anytime constraints. Experiments on the application of carbon-intelligent computing verify the reward performance and cost constraint guarantee of ACRL.

We study a general form of Smoothed Online Convex Optimization, a.k.a. SOCO, including multi-step switching costs and feedback delay. We propose a novel machine learning (ML) augmented online algorithm, Robustness-Constrained Learning (RCL), which combines untrusted ML predictions with a trusted expert online algorithm via constrained projection to robustify the ML prediction. Specifically, we prove that RCL is able to guarantee (1+lambda)-competitiveness against any given expert for any lambda>0, while also explicitly training the ML model in a robustification-aware manner to improve the average-case performance. Importantly, RCL is the first ML-augmented algorithm with a provable robustness guarantee in the case of multi-step switching cost and feedback delay. We demonstrate the improvement of RCL in both robustness and average performance using battery management for electrifying transportation as a case study.

Artificial intelligence has become feasible on tiny edge devices, thanks to the availability of high-performance microcontroller units (MCUs) and the development of highly-efficient machine learning (ML) models. Although successful, the current design practice implementing high-performance ML models on high-end MCU devices, still presents barriers that prohibit the more widespread adoption of ML on real-world applications built on low-end devices, where power and available hardware resources have been critical factors for including ML functionality in the final product. On the one hand, these low-end MCU devices have very limited computational resources in terms of hardware resources and power budget, which makes ML inference less accessible. On the other hand, these low-end and low-cost MCUs present the leading physical carrier of real-world smart applications, making small and fast-responsive (i.e., always-on) ML models an urgent need. This paper presents MicroVSA, a library of optimized implementations of a low-dimensional computing (LDC) classifier, a recently proposed variant of vector symbolic architecture (VSA), for 8-bit, 16-bit, and 32-bit MCUs. MicroVSA achieves 1.33–10.27x speedup from the vanilla LDC, and our sample model achieves 85–93% accuracy on three most common classification tasks for always-on inference on the MCUs — myocardial infarction detection, human activity recognition, and hot word detection — while requiring only a few tens of bytes of RAM and fitting in the vast majority of tiny 8-bit MCUs. For instance, our model for detecting the phrase "Hey Snapdragon" only needs 6.75KB of flash and 0.02KB of RAM and can complete one inference in 10.6 ms on a low-end 8-bit AVR MCU or 0.68 ms on an ARM Cortex-M4 MCU, outperforming the best neural network-based model by a factor of 7.2x-113x, making it the first to run on an 8-bit MCUs and the most efficient hot word detection model. Our study suggests that ubiquitous ML deployment on extremely low-cost MCUs is possible and that more study on VSA model training, model compression, and implementation techniques is needed to further explore the possibility of lowering the cost and power of ML on edge devices.

Many problems, such as online ad display, can be formulated as online bipartite matching. The crucial challenge lies in the nature of sequentially-revealed online item information, based on which we make irreversible matching decisions at each step. While numerous expert online algorithms have been proposed with bounded worst-case competitive ratios, they may not offer satisfactory performance in average cases. On the other hand, reinforcement learning (RL) has been applied to improve the average performance, but it lacks robustness and can perform arbitrarily poorly. In this paper, we propose a novel RL-based approach to edge-weighted online bipartite matching with robustness guarantees (LOMAR), achieving both good average-case and worst-case performance. The key novelty of LOMAR is a new online switching operation which, based on a judicious condition to hedge against future uncertainties, decides whether to follow the expert’s decision or the RL decision for each online item. We prove that for any ρ∈[0,1], LOMAR is ρ-competitive against any given expert online algorithm. To improve the average performance, we train the RL policy by explicitly considering the online switching operation. Finally, we run empirical experiments to demonstrate the advantages of LOMAR compared to existing baselines.

We study online convex optimization with switching costs, a practically important but also extremely challenging problem due to the lack of complete offline information. By tapping into the power of machine learning (ML) based optimizers, ML-augmented online algorithms (also referred to as expert calibration in this paper) have been emerging as state of the art, with provable worst-case performance guarantees. Nonetheless, by using the standard practice of training an ML model as a standalone optimizer and plugging it into an ML-augmented algorithm, the average cost performance can be highly unsatisfactory. In order to address the "how to learn" challenge, we propose EC-L2O (expert-calibrated learning to optimize), which trains an ML-based optimizer by explicitly taking into account the downstream expert calibrator. To accomplish this, we propose a new differentiable expert calibrator that generalizes regularized online balanced descent and offers a provably better competitive ratio than pure ML predictions when the prediction error is large. For training, our loss function is a weighted sum of two different losses — one minimizing the average ML prediction error for better robustness, and the other one minimizing the post-calibration average cost. We also provide theoretical analysis for EC-L2O, highlighting that expert calibration can be even beneficial for the average cost performance and that the high-percentile tail ratio of the cost achieved by EC-L2O to that of the offline optimal oracle (i.e., tail cost ratio) can be bounded. Finally, we test EC-L2O by running simulations for sustainable datacenter demand response. Our results demonstrate that EC-L2O can empirically achieve a lower average cost as well as a lower competitive ratio than the existing baseline algorithms.

Convolutional neural networks (CNNs) are used in numerous real-world applications such as vision-based autonomous driving and video content analysis. To run CNN inference on various target devices, hardware-aware neural architecture search (NAS) is crucial. A key requirement of efficient hardware-aware NAS is the fast evaluation of inference latencies in order to rank different architectures. While building a latency predictor for each target device has been commonly used in state of the art, this is a very time-consuming process, lacking scalability in the presence of extremely diverse devices. In this work, we address the scalability challenge by exploiting latency monotonicity — the architecture latency rankings on different devices are often correlated. When strong latency monotonicity exists, we can re-use architectures searched for one proxy device on new target devices, without losing optimality. In the absence of strong latency monotonicity, we propose an efficient proxy adaptation technique to significantly boost the latency monotonicity. Finally, we validate our approach and conduct experiments with devices of different platforms on multiple mainstream search spaces, including MobileNet-V2, MobileNet-V3, NAS-Bench-201, ProxylessNAS and FBNet. Our results highlight that, by using just one proxy device, we can find almost the same Pareto-optimal architectures as the existing per-device NAS, while avoiding the prohibitive cost of building a latency predictor for each device.