We study how to watermark LLM outputs, i.e. embedding algorithmically detectable signals into LLM-generated text to track misuse. Unlike the current mainstream methods that work with a fixed LLM, we expand the watermark design space by including the LLM tuning stage in the watermark pipeline. While prior works focus on token-level watermark that embeds signals into the output, we design a model-level watermark that embeds signals into the LLM weights, and such signals can be detected by a paired detector. We propose a co-training framework based on reinforcement learning that iteratively (1) trains a detector to detect the generated watermarked text and (2) tunes the LLM to generate text easily detectable by the detector while keeping its normal utility. We empirically show that our watermarks are more accurate, robust, and adaptable (to new attacks). It also allows watermarked model open-sourcing. In addition, if used together with alignment, the extra overhead introduced is low - only training an extra reward model (i.e. our detector). We hope our work can bring more effort into studying a broader watermark design that is not limited to working with a fixed LLM. We open-source the code: https://github.com/xiaojunxu/learning-to-watermark-llm .

Although an ever-growing number of applications employ deep learning based systems for prediction, decision-making, or state estimation, almost no certification processes have been established that would allow such systems to be deployed in safety-critical applications. In this work we consider real-world problems arising in aviation and other safety-critical areas, and investigate their requirements for a certified model. To this end, we investigate methodologies from the machine learning research community aimed towards verifying robustness and reliability of deep learning systems, and evaluate these methodologies with regard to their applicability to real-world problems. Then, we establish a new framework towards deep learning certification based on (i) inherently safe design, and (ii) run-time error detection. Using a concrete use case from aviation, we show how deep learning models can recover disentangled variables through the use of weakly-supervised representation learning. We argue that such a system design is inherently less prone to common model failures, and can be verified to encode underlying mechanisms governing the data. Then, we investigate four techniques related to the run-time safety of a model, namely (i) uncertainty quantification, (ii) out-of-distribution detection, (iii) feature collapse, and (iv) adversarial attacks. We evaluate each for their applicability and formulate a set of desiderata that a certified model should fulfill. Finally, we propose a novel model structure that exhibits all desired properties discussed in this work, and is able to make regression and uncertainty predictions, as well as detect out-of-distribution inputs, while requiring no regression labels to train. We conclude with a discussion of the current state and expected future progress of deep learning certification, and its industrial and social implications.

Model-induced distribution shifts (MIDS) occur as previous model outputs pollute new model training sets over generations of models. This is known as model collapse in the case of generative models, and performative prediction or unfairness feedback loops for supervised models. When a model induces a distribution shift, it also encodes its mistakes, biases, and unfairnesses into the ground truth of its data ecosystem. We introduce a framework that allows us to track multiple MIDS over many generations, finding that they can lead to loss in performance, fairness, and minoritized group representation, even in initially unbiased datasets. Despite these negative consequences, we identify how models might be used for positive, intentional, interventions in their data ecosystems, providing redress for historical discrimination through a framework called algorithmic reparation (AR). We simulate AR interventions by curating representative training batches for stochastic gradient descent to demonstrate how AR can improve upon the unfairnesses of models and data ecosystems subject to other MIDS. Our work takes an important step towards identifying, mitigating, and taking accountability for the unfair feedback loops enabled by the idea that ML systems are inherently neutral and objective.

Describing the dynamics of many-electron quantum systems is crucial for applications such as predicting electronic structures in quantum chemistry, the properties of condensed matter systems, and the behaviors of complex materials. However, the real-time evolution of non-equilibrium quantum electronic systems poses a significant challenge for theoretical and computational approaches, due to the system's exploration of a vast configuration space. This work introduces a variational approach for fermionic time-dependent wave functions, surpassing mean-field approximations by capturing many-body correlations. The proposed methodology involves parameterizing the time-evolving quantum state, enabling the approximation of the state's evolution. To account for electron correlations, we employ time-dependent Jastrow factors and backflow transformations. We also show that we can incorporate neural networks to parameterize these functions. The time-dependent variational Monte Carlo technique is employed to efficiently compute the optimal time-dependent parameters. The approach is demonstrated in three distinct systems: the solvable harmonic interaction model, the dynamics of a diatomic molecule in intense laser fields, and a quenched quantum dot. In all cases, we show clear signatures of many-body correlations in the dynamics not captured by mean-field methods. The results showcase the ability of our variational approach to accurately capture the time evolution of quantum states, providing insight into the quantum dynamics of interacting electronic systems, beyond the capabilities of mean-field.

Addressing the so-called ``Red-AI'' trend of rising energy consumption by large-scale neural networks, this study investigates the actual energy consumption, as measured by node-level watt-meters, of training various fully connected neural network architectures. We introduce the BUTTER-E dataset, an augmentation to the BUTTER Empirical Deep Learning dataset, containing energy consumption and performance data from 63,527 individual experimental runs spanning 30,582 distinct configurations: 13 datasets, 20 sizes (number of trainable parameters), 8 network ``shapes'', and 14 depths on both CPU and GPU hardware collected using node-level watt-meters. This dataset reveals the complex relationship between dataset size, network structure, and energy use, and highlights the impact of cache effects. We propose a straightforward and effective energy model that accounts for network size, computing, and memory hierarchy. Our analysis also uncovers a surprising, hardware-mediated non-linear relationship between energy efficiency and network design, challenging the assumption that reducing the number of parameters or FLOPs is the best way to achieve greater energy efficiency. Highlighting the need for cache-considerate algorithm development, we suggest a combined approach to energy efficient network, algorithm, and hardware design. This work contributes to the fields of sustainable computing and Green AI, offering practical guidance for creating more energy-efficient neural networks and promoting sustainable AI.

Artificial intelligence (AI) ethics has emerged as a burgeoning yet pivotal area of scholarly research. This study conducts a comprehensive bibliometric analysis of the AI ethics literature over the past two decades. The analysis reveals a discernible tripartite progression, characterized by an incubation phase, followed by a subsequent phase focused on imbuing AI with human-like attributes, culminating in a third phase emphasizing the development of human-centric AI systems. After that, they present seven key AI ethics issues, encompassing the Collingridge dilemma, the AI status debate, challenges associated with AI transparency and explainability, privacy protection complications, considerations of justice and fairness, concerns about algocracy and human enfeeblement, and the issue of superintelligence. Finally, they identify two notable research gaps in AI ethics regarding the large ethics model (LEM) and AI identification and extend an invitation for further scholarly research.

Spatiotemporal datasets, which consist of spatially-referenced time series, are ubiquitous in many scientific and business-intelligence applications, such as air pollution monitoring, disease tracking, and cloud-demand forecasting. As modern datasets continue to increase in size and complexity, there is a growing need for new statistical methods that are flexible enough to capture complex spatiotemporal dynamics and scalable enough to handle large prediction problems. This work presents the Bayesian Neural Field (BayesNF), a domain-general statistical model for inferring rich probability distributions over a spatiotemporal domain, which can be used for data-analysis tasks including forecasting, interpolation, and variography. BayesNF integrates a novel deep neural network architecture for high-capacity function estimation with hierarchical Bayesian inference for robust uncertainty quantification. By defining the prior through a sequence of smooth differentiable transforms, posterior inference is conducted on large-scale data using variationally learned surrogates trained via stochastic gradient descent. We evaluate BayesNF against prominent statistical and machine-learning baselines, showing considerable improvements on diverse prediction problems from climate and public health datasets that contain tens to hundreds of thousands of measurements. The paper is accompanied with an open-source software package (https://github.com/google/bayesnf) that is easy-to-use and compatible with modern GPU and TPU accelerators on the JAX machine learning platform.

Signed graphs are an emergent way of representing data in a variety of contexts were conflicting interactions exist. These include data from biological, ecological, and social systems. Here we propose the concept of communicability geometry for signed graphs, proving that metrics in this space, such as the communicability distance and angles, are Euclidean and spherical. We then apply these metrics to solve several problems in data analysis of signed graphs in a unified way. They include the partitioning of signed graphs, dimensionality reduction, finding hierarchies of alliances in signed networks as well as the quantification of the degree of polarization between the existing factions in systems represented by this type of graphs.

The increasing utilization of emerging technologies in the Food & Agriculture (FA) sector has heightened the need for security to minimize cyber risks. Considering this aspect, this manuscript reviews disclosed and documented cybersecurity incidents in the FA sector. For this purpose, thirty cybersecurity incidents were identified, which took place between July 2011 and April 2023. The details of these incidents are reported from multiple sources such as: the private industry and flash notifications generated by the Federal Bureau of Investigation (FBI), internal reports from the affected organizations, and available media sources. Considering the available information, a brief description of the security threat, ransom amount, and impact on the organization are discussed for each incident. This review reports an increased frequency of cybersecurity threats to the FA sector. To minimize these cyber risks, popular cybersecurity frameworks and recent agriculture-specific cybersecurity solutions are also discussed. Further, the need for AI assurance in the FA sector is explained, and the Farmer-Centered AI (FCAI) framework is proposed. The main aim of the FCAI framework is to support farmers in decision-making for agricultural production, by incorporating AI assurance. Lastly, the effects of the reported cyber incidents on other critical infrastructures, food security, and the economy are noted, along with specifying the open issues for future development.

Yield forecasting, the science of predicting agricultural productivity before the crop harvest occurs, helps a wide range of stakeholders make better decisions around agricultural planning. This study aims to investigate whether machine learning-based yield prediction models can capably predict Kharif season rice yields at the district level in India several months before the rice harvest takes place. The methodology involved training 19 machine learning models such as CatBoost, LightGBM, Orthogonal Matching Pursuit, and Extremely Randomized Trees on 20 years of climate, satellite, and rice yield data across 247 of Indian rice-producing districts. In addition to model-building, a dynamic dashboard was built understand how the reliability of rice yield predictions varies across districts. The results of the proof-of-concept machine learning pipeline demonstrated that rice yields can be predicted with a reasonable degree of accuracy, with out-of-sample R2, MAE, and MAPE performance of up to 0.82, 0.29, and 0.16 respectively. These results outperformed test set performance reported in related literature on rice yield modeling in other contexts and countries. In addition, SHAP value analysis was conducted to infer both the importance and directional impact of the climate and remote sensing variables included in the model. Important features driving rice yields included temperature, soil water volume, and leaf area index. In particular, higher temperatures in August correlate with increased rice yields, particularly when the leaf area index in August is also high. Building on the results, a proof-of-concept dashboard was developed to allow users to easily explore which districts may experience a rise or fall in yield relative to the previous year.