Search

The reduced sensitivity of mean Southern Ocean zonal transport with respect to surface wind stress magnitude changes, known as eddy saturation, is studied in an idealised analytical model. The model is based on the assumption of a balance between surface wind stress forcing and bottom dissipation in the planetary geostrophic limit, coupled to the GEOMETRIC form of the Gent--McWilliams eddy parameterisation. The assumption of a linear stratification, together with an equation for the parameterised domain integrated total eddy energy, enables the formulation of a two component dynamical system, which reduces to the non-linear oscillator of Ambaum and Novak (Q. J. R. Meteorolog. Soc. 140(685), 2680--2684, 2014) in a Hamiltonian limit. The model suggests an intrinsic oscillatory time scale for the Southern Ocean, associated with a combination of mean shear erosion by eddies and eddy energy generation by the mean shear. For Southern Ocean parameters the model suggests that perturbing the system via stochastic wind forcing may lead to relatively large excursions in eddy energy.

Spin-photon interfaces based on group-IV colour centres in diamond offer a promising platform for quantum networks. A key challenge in the field is realizing precise single-defect positioning and activation, which is crucial for scalable device fabrication. Here we address this problem by demonstrating a two-step fabrication method for tin vacancy (SnV-) centres that uses site-controlled ion implantation followed by local femtosecond laser annealing with in-situ spectral monitoring. The ion implantation is performed with sub-50 nm resolution and a dosage that is controlled from hundreds of ions down to single ions per site, limited by Poissonian statistics. Using this approach, we successfully demonstrate site-selective creation and modification of single SnV- centres. The technique opens a window onto materials tuning at the single defect level, and provides new insight into defect structures and dynamics during the annealing process. While demonstrated for SnV- centres, this versatile approach can be readily generalised to other implanted colour centres in diamond and wide-bandgap materials., Comment: 10 pages, 4 figures, methods, and supplementary information

An automated framework is presented for the numerical solution of optimal control problems with PDEs as constraints, in both the stationary and instationary settings. The associated code can solve both linear and non-linear problems, and examples for incompressible flow equations are considered. The software, which is based on a Python interface to the Firedrake system, allows for a compact definition of the problem considered by providing a few lines of code in a high-level language. The software is provided with efficient iterative linear solvers for optimal control problems with PDEs as constraints. The use of advanced preconditioning techniques results in a significant speed-up of the solution process for large-scale problems. We present numerical examples of the applicability of the software on classical control problems with PDEs as constraints.

Machine learning models have the potential to identify cardiovascular diseases (CVDs) early and accurately in primary healthcare settings, which is crucial for delivering timely treatment and management. Although population-based CVD risk models have been used traditionally, these models often do not consider variations in lifestyles, socioeconomic conditions, or genetic predispositions. Therefore, we aimed to develop machine learning models for CVD detection using primary healthcare data, compare the performance of different models, and identify the best models. We used data from the UK Biobank study, which included over 500,000 middle-aged participants from different primary healthcare centers in the UK. Data collected at baseline (2006--2010) and during imaging visits after 2014 were used in this study. Baseline characteristics, including sex, age, and the Townsend Deprivation Index, were included. Participants were classified as having CVD if they reported at least one of the following conditions: heart attack, angina, stroke, or high blood pressure. Cardiac imaging data such as electrocardiogram and echocardiography data, including left ventricular size and function, cardiac output, and stroke volume, were also used. We used 9 machine learning models (LSVM, RBFSVM, GP, DT, RF, NN, AdaBoost, NB, and QDA), which are explainable and easily interpretable. We reported the accuracy, precision, recall, and F-1 scores; confusion matrices; and area under the curve (AUC) curves., Comment: 19 pages, 3 figures

Knowing the effect of an intervention is critical for human decision-making, but current approaches for causal effect estimation rely on manual data collection and structuring, regardless of the causal assumptions. This increases both the cost and time-to-completion for studies. We show how large, diverse observational text data can be mined with large language models (LLMs) to produce inexpensive causal effect estimates under appropriate causal assumptions. We introduce NATURAL, a novel family of causal effect estimators built with LLMs that operate over datasets of unstructured text. Our estimators use LLM conditional distributions (over variables of interest, given the text data) to assist in the computation of classical estimators of causal effect. We overcome a number of technical challenges to realize this idea, such as automating data curation and using LLMs to impute missing information. We prepare six (two synthetic and four real) observational datasets, paired with corresponding ground truth in the form of randomized trials, which we used to systematically evaluate each step of our pipeline. NATURAL estimators demonstrate remarkable performance, yielding causal effect estimates that fall within 3 percentage points of their ground truth counterparts, including on real-world Phase 3/4 clinical trials. Our results suggest that unstructured text data is a rich source of causal effect information, and NATURAL is a first step towards an automated pipeline to tap this resource., Comment: 28 pages, 11 figures

Large Language Models (LLMs) have become increasingly capable of handling diverse tasks with the aid of well-crafted prompts and integration of external tools, but as task complexity rises, the workflow involving LLMs can be complicated and thus challenging to implement and maintain. To address this challenge, we propose APPL, A Prompt Programming Language that acts as a bridge between computer programs and LLMs, allowing seamless embedding of prompts into Python functions, and vice versa. APPL provides an intuitive and Python-native syntax, an efficient parallelized runtime with asynchronous semantics, and a tracing module supporting effective failure diagnosis and replaying without extra costs. We demonstrate that APPL programs are intuitive, concise, and efficient through three representative scenarios: Chain-of-Thought with self-consistency (CoT-SC), ReAct tool use agent, and multi-agent chat. Experiments on three parallelizable workflows further show that APPL can effectively parallelize independent LLM calls, with a significant speedup ratio that almost matches the estimation.

We study the problem of designing minimax procedures in linear regression under the quantile risk. We start by considering the realizable setting with independent Gaussian noise, where for any given noise level and distribution of inputs, we obtain the exact minimax quantile risk for a rich family of error functions and establish the minimaxity of OLS. This improves on the known lower bounds for the special case of square error, and provides us with a lower bound on the minimax quantile risk over larger sets of distributions. Under the square error and a fourth moment assumption on the distribution of inputs, we show that this lower bound is tight over a larger class of problems. Specifically, we prove a matching upper bound on the worst-case quantile risk of a variant of the recently proposed min-max regression procedure, thereby establishing its minimaxity, up to absolute constants. We illustrate the usefulness of our approach by extending this result to all $p$-th power error functions for $p \in (2, \infty)$. Along the way, we develop a generic analogue to the classical Bayesian method for lower bounding the minimax risk when working with the quantile risk, as well as a tight characterization of the quantiles of the smallest eigenvalue of the sample covariance matrix.

Frontier Large Language Models (LLMs) can be socially discriminatory or sensitive to spurious features of their inputs. Because only well-resourced corporations can train frontier LLMs, we need robust test-time strategies to control such biases. Existing solutions, which instruct the LLM to be fair or robust, rely on the model's implicit understanding of bias. Causality provides a rich formalism through which we can be explicit about our debiasing requirements. Yet, as we show, a naive application of the standard causal debiasing strategy, counterfactual data augmentation, fails under standard assumptions to debias predictions at an individual level at test time. To address this, we develop a stratified notion of debiasing called stratified invariance, which can capture a range of debiasing requirements from population level to individual level through an additional measurement that stratifies the predictions. We present a complete observational test for stratified invariance. Finally, we introduce a data augmentation strategy that guarantees stratified invariance at test time under suitable assumptions, together with a prompting strategy that encourages stratified invariance in LLMs. We show that our prompting strategy, unlike implicit instructions, consistently reduces the bias of frontier LLMs across a suite of synthetic and real-world benchmarks without requiring additional data, finetuning or pre-training.

Training on mixtures of data distributions is now common in many modern machine learning pipelines, useful for performing well on several downstream tasks. Group distributionally robust optimization (group DRO) is one popular way to learn mixture weights for training a specific model class, but group DRO methods suffer for non-linear models due to non-convex loss functions and when the models are non-parametric. We address these challenges by proposing to solve a more general DRO problem, giving a method we call MixMax. MixMax selects mixture weights by maximizing a particular concave objective with entropic mirror ascent, and, crucially, we prove that optimally fitting this mixture distribution over the set of bounded predictors returns a group DRO optimal model. Experimentally, we tested MixMax on a sequence modeling task with transformers and on a variety of non-parametric learning problems. In all instances MixMax matched or outperformed the standard data mixing and group DRO baselines, and in particular, MixMax improved the performance of XGBoost over the only baseline, data balancing, for variations of the ACSIncome and CelebA annotations datasets.

Understanding how language model performance varies with scale is critical to benchmark and algorithm development. Scaling laws are one approach to building this understanding, but the requirement of training models across many different scales has limited their use. We propose an alternative, observational approach that bypasses model training and instead builds scaling laws from ~100 publically available models. Building a single scaling law from multiple model families is challenging due to large variations in their training compute efficiencies and capabilities. However, we show that these variations are consistent with a simple, generalized scaling law where language model performance is a function of a low-dimensional capability space, and model families only vary in their efficiency in converting training compute to capabilities. Using this approach, we show the surprising predictability of complex scaling phenomena: we show that several emergent phenomena follow a smooth, sigmoidal behavior and are predictable from small models; we show that the agent performance of models such as GPT-4 can be precisely predicted from simpler non-agentic benchmarks; and we show how to predict the impact of post-training interventions like Chain-of-Thought and Self-Consistency as language model capabilities continue to improve., Comment: Accepted at NeurIPS 2024 as a spotlight

Identifying how much a model ${\widehat{p}}_{\theta}(Y|X)$ knows about the stochastic real-world process $p(Y|X)$ it was trained on is important to ensure it avoids producing incorrect or "hallucinated" answers or taking unsafe actions. But this is difficult for generative models because probabilistic predictions do not distinguish between per-response noise (aleatoric uncertainty) and lack of knowledge about the process (epistemic uncertainty), and existing epistemic uncertainty quantification techniques tend to be overconfident when the model underfits. We propose a general strategy for teaching a model to both approximate $p(Y|X)$ and also estimate the remaining gaps between ${\widehat{p}}_{\theta}(Y|X)$ and $p(Y|X)$: train it to predict pairs of independent responses drawn from the true conditional distribution, allow it to "cheat" by observing one response while predicting the other, then measure how much it cheats. Remarkably, we prove that being good at cheating (i.e. cheating whenever it improves your prediction) is equivalent to being second-order calibrated, a principled extension of ordinary calibration that allows us to construct provably-correct frequentist confidence intervals for $p(Y|X)$ and detect incorrect responses with high probability. We demonstrate empirically that our approach accurately estimates how much models don't know across ambiguous image classification, (synthetic) language modeling, and partially-observable navigation tasks, outperforming existing techniques., Comment: Accepted at ICML 2024. 9 pages, 6 figures

Recent advances in Language Model (LM) agents and tool use, exemplified by applications like ChatGPT Plugins, enable a rich set of capabilities but also amplify potential risks - such as leaking private data or causing financial losses. Identifying these risks is labor-intensive, necessitating implementing the tools, setting up the environment for each test scenario manually, and finding risky cases. As tools and agents become more complex, the high cost of testing these agents will make it increasingly difficult to find high-stakes, long-tailed risks. To address these challenges, we introduce ToolEmu: a framework that uses an LM to emulate tool execution and enables the testing of LM agents against a diverse range of tools and scenarios, without manual instantiation. Alongside the emulator, we develop an LM-based automatic safety evaluator that examines agent failures and quantifies associated risks. We test both the tool emulator and evaluator through human evaluation and find that 68.8% of failures identified with ToolEmu would be valid real-world agent failures. Using our curated initial benchmark consisting of 36 high-stakes tools and 144 test cases, we provide a quantitative risk analysis of current LM agents and identify numerous failures with potentially severe outcomes. Notably, even the safest LM agent exhibits such failures 23.9% of the time according to our evaluator, underscoring the need to develop safer LM agents for real-world deployment.

Solid-state spin qubits within silicon crystals at mK temperatures show great promise in the realisation of a fully scalable quantum computation platform. Qubit coherence times are limited in natural silicon owing to coupling to the isotope ${ }^{29} \mathrm{Si}$ which has a non-zero nuclear spin. This work presents a method for the depletion of ${ }^{29} \mathrm{Si}$ in localised volumes of natural silicon wafers by irradiation using a 45 keV ${ }^{28} \mathrm{Si}$ focused ion beam with fluences above $1 \times 10^{19} \, \mathrm{ions} \, \mathrm{cm}^{-2}$. Nanoscale secondary ion mass spectrometry analysis of the irradiated volumes shows unprecedented quality enriched silicon that reaches a minimal residual ${ }^{29} \mathrm{Si}$ value of 2.3 $\pm$ 0.7 ppm and with residual C and O comparable to the background concentration in the unimplanted wafer. Transmission electron microscopy lattice images confirm the solid phase epitaxial re-crystallization of the as-implanted amorphous enriched volume extending over 200 nm in depth upon annealing. The ease of fabrication, requiring only commercially available natural silicon wafers and ion sources, opens the possibility for co-integration of qubits in localised highly enriched volumes with control circuitry in the surrounding natural silicon for large-scale devices., Comment: 8 pages, 4 figures, 2 tables

Designing models that are both expressive and preserve known invariances of tasks is an increasingly hard problem. Existing solutions tradeoff invariance for computational or memory resources. In this work, we show how to leverage randomness and design models that are both expressive and invariant but use less resources. Inspired by randomized algorithms, our key insight is that accepting probabilistic notions of universal approximation and invariance can reduce our resource requirements. More specifically, we propose a class of binary classification models called Randomized Linear Classifiers (RLCs). We give parameter and sample size conditions in which RLCs can, with high probability, approximate any (smooth) function while preserving invariance to compact group transformations. Leveraging this result, we design three RLCs that are provably probabilistic invariant for classification tasks over sets, graphs, and spherical data. We show how these models can achieve probabilistic invariance and universality using less resources than (deterministic) neural networks and their invariant counterparts. Finally, we empirically demonstrate the benefits of this new class of models on invariant tasks where deterministic invariant neural networks are known to struggle.

In deep learning theory, the covariance matrix of the representations serves as a proxy to examine the network's trainability. Motivated by the success of Transformers, we study the covariance matrix of a modified Softmax-based attention model with skip connections in the proportional limit of infinite-depth-and-width. We show that at initialization the limiting distribution can be described by a stochastic differential equation (SDE) indexed by the depth-to-width ratio. To achieve a well-defined stochastic limit, the Transformer's attention mechanism is modified by centering the Softmax output at identity, and scaling the Softmax logits by a width-dependent temperature parameter. We examine the stability of the network through the corresponding SDE, showing how the scale of both the drift and diffusion can be elegantly controlled with the aid of residual connections. The existence of a stable SDE implies that the covariance structure is well-behaved, even for very large depth and width, thus preventing the notorious issues of rank degeneracy in deep attention models. Finally, we show, through simulations, that the SDE provides a surprisingly good description of the corresponding finite-size model. We coin the name shaped Transformer for these architectural modifications.

There is an increasing interest in mesoscale eddy parameterizations that are scale-aware, normally interpreted to mean that a parameterization does not require re-tuning of parameters as the model resolution changes. Here we examine whether Gent--McWilliams (GM) based version of GEOMETRIC, a mesoscale eddy parameterization that is constrained by a parameterized eddy energy budget, is scale-aware in its energetics. It is generally known that GM-based schemes severely damp out explicit eddies, so the parameterized component would be expected to dominate across resolutions, and we might expect a negative answer to the question of energetic scale-awareness. A consideration of why GM-based schemes damp out explicit eddies leads a suggestion for what we term a splitting procedure: a definition of a `large-scale' field is sought, and the eddy-induced velocity from the GM-scheme is computed from and acts only on the large-scale field, allowing explicit and parameterized components to co-exist. Within the context of an idealized re-entrant channel model of the Southern Ocean, evidence is provided that the GM-based version of GEOMETRIC is scale-aware in the energetics as long as we employ a splitting procedure. The splitting procedure also leads to an improved representation of mean states without detrimental effects on the variability., Comment: 26 pages, 8 figures, 1 table. Previous version has data generated from a bugged NEMO implementation of GEOMETRIC, and the results from the previous version could be regarded as a case where GEOMETRIC is used ***without*** mean flow advection of parameterised eddy energy. This bug is fixed in the present version, with minor updates to the article to reflect the change

Training algorithms, broadly construed, are an essential part of every deep learning pipeline. Training algorithm improvements that speed up training across a wide variety of workloads (e.g., better update rules, tuning protocols, learning rate schedules, or data selection schemes) could save time, save computational resources, and lead to better, more accurate, models. Unfortunately, as a community, we are currently unable to reliably identify training algorithm improvements, or even determine the state-of-the-art training algorithm. In this work, using concrete experiments, we argue that real progress in speeding up training requires new benchmarks that resolve three basic challenges faced by empirical comparisons of training algorithms: (1) how to decide when training is complete and precisely measure training time, (2) how to handle the sensitivity of measurements to exact workload details, and (3) how to fairly compare algorithms that require hyperparameter tuning. In order to address these challenges, we introduce a new, competitive, time-to-result benchmark using multiple workloads running on fixed hardware, the AlgoPerf: Training Algorithms benchmark. Our benchmark includes a set of workload variants that make it possible to detect benchmark submissions that are more robust to workload changes than current widely-used methods. Finally, we evaluate baseline submissions constructed using various optimizers that represent current practice, as well as other optimizers that have recently received attention in the literature. These baseline results collectively demonstrate the feasibility of our benchmark, show that non-trivial gaps between methods exist, and set a provisional state-of-the-art for future benchmark submissions to try and surpass., Comment: 102 pages, 8 figures, 41 tables

Automated code generation allows for a separation between the development of a model, expressed via a domain specific language, and lower level implementation details. Algorithmic differentiation can be applied symbolically at the level of the domain specific language, and the code generator reused to implement code required for an adjoint calculation. However the adjoint calculations are complicated by the well-known problem of storing or recomputing the forward data required by the adjoint, and different checkpointing strategies have been developed to tackle this problem. This article considers the combination of high-level algorithmic differentiation with step-based checkpointing schedules, with the primary application being for solvers of time-dependent partial differential equations. The focus is on algorithmic differentiation using a dynamically constructed record of forward operations, where the precise structure of the original forward calculation is unknown ahead-of-time. In addition, high-level approaches provide a simplified view of the model itself. This allows data required to restart and advance the forward, and data required to advance the adjoint, to be identified. The difference between the two types of data is here leveraged to implement checkpointing strategies with improved performance.