Search

Safety and responsibility evaluations of advanced AI models are a critical but developing field of research and practice. In the development of Google DeepMind's advanced AI models, we innovated on and applied a broad set of approaches to safety evaluation. In this report, we summarise and share elements of our evolving approach as well as lessons learned for a broad audience. Key lessons learned include: First, theoretical underpinnings and frameworks are invaluable to organise the breadth of risk domains, modalities, forms, metrics, and goals. Second, theory and practice of safety evaluation development each benefit from collaboration to clarify goals, methods and challenges, and facilitate the transfer of insights between different stakeholders and disciplines. Third, similar key methods, lessons, and institutions apply across the range of concerns in responsibility and safety - including established and emerging harms. For this reason it is important that a wide range of actors working on safety evaluation and safety research communities work together to develop, refine and implement novel evaluation approaches and best practices, rather than operating in silos. The report concludes with outlining the clear need to rapidly advance the science of evaluations, to integrate new evaluations into the development and governance of AI, to establish scientifically-grounded norms and standards, and to promote a robust evaluation ecosystem., Comment: 10 pages excluding bibliography

To understand the risks posed by a new AI system, we must understand what it can and cannot do. Building on prior work, we introduce a programme of new "dangerous capability" evaluations and pilot them on Gemini 1.0 models. Our evaluations cover four areas: (1) persuasion and deception; (2) cyber-security; (3) self-proliferation; and (4) self-reasoning. We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs. Our goal is to help advance a rigorous science of dangerous capability evaluation, in preparation for future models.

We show that existing unsupervised methods on large language model (LLM) activations do not discover knowledge -- instead they seem to discover whatever feature of the activations is most prominent. The idea behind unsupervised knowledge elicitation is that knowledge satisfies a consistency structure, which can be used to discover knowledge. We first prove theoretically that arbitrary features (not just knowledge) satisfy the consistency structure of a particular leading unsupervised knowledge-elicitation method, contrast-consistent search (Burns et al. - arXiv:2212.03827). We then present a series of experiments showing settings in which unsupervised methods result in classifiers that do not predict knowledge, but instead predict a different prominent feature. We conclude that existing unsupervised methods for discovering latent knowledge are insufficient, and we contribute sanity checks to apply to evaluating future knowledge elicitation methods. Conceptually, we hypothesise that the identification issues explored here, e.g. distinguishing a model's knowledge from that of a simulated character's, will persist for future unsupervised methods., Comment: 12 pages (38 including references and appendices). First three authors equal contribution, randomised order

Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is critical for addressing extreme risks. Developers must be able to identify dangerous capabilities (through "dangerous capability evaluations") and the propensity of models to apply their capabilities for harm (through "alignment evaluations"). These evaluations will become critical for keeping policymakers and other stakeholders informed, and for making responsible decisions about model training, deployment, and security., Comment: Fixed typos; added citation

Information-theoretic approaches to active learning have traditionally focused on maximising the information gathered about the model parameters, most commonly by optimising the BALD score. We highlight that this can be suboptimal from the perspective of predictive performance. For example, BALD lacks a notion of an input distribution and so is prone to prioritise data of limited relevance. To address this we propose the expected predictive information gain (EPIG), an acquisition function that measures information gain in the space of predictions rather than parameters. We find that using EPIG leads to stronger predictive performance compared with BALD across a range of datasets and models, and thus provides an appealing drop-in replacement., Comment: Published at AISTATS 2023

We introduce a method to measure uncertainty in large language models. For tasks like question answering, it is essential to know when we can trust the natural language outputs of foundation models. We show that measuring uncertainty in natural language is challenging because of "semantic equivalence" -- different sentences can mean the same thing. To overcome these challenges we introduce semantic entropy -- an entropy which incorporates linguistic invariances created by shared meanings. Our method is unsupervised, uses only a single model, and requires no modifications to off-the-shelf language models. In comprehensive ablation studies we show that the semantic entropy is more predictive of model accuracy on question answering data sets than comparable baselines., Comment: International Conference on Learning Representations 2023 (Spotlight)

We show how to "compile" human-readable programs into standard decoder-only transformer models. Our compiler, Tracr, generates models with known structure. This structure can be used to design experiments. For example, we use it to study "superposition" in transformers that execute multi-step algorithms. Additionally, the known structure of Tracr-compiled models can serve as ground-truth for evaluating interpretability methods. Commonly, because the "programs" learned by transformers are unknown it is unclear whether an interpretation succeeded. We demonstrate our approach by implementing and examining programs including computing token frequencies, sorting, and parenthesis checking. We provide an open-source implementation of Tracr at https://github.com/google-deepmind/tracr., Comment: Presented at NeurIPS 2023 (Spotlight)

Users often ask dialogue systems ambiguous questions that require clarification. We show that current language models rarely ask users to clarify ambiguous questions and instead provide incorrect answers. To address this, we introduce CLAM: a framework for getting language models to selectively ask for clarification about ambiguous user questions. In particular, we show that we can prompt language models to detect whether a given question is ambiguous, generate an appropriate clarifying question to ask the user, and give a final answer after receiving clarification. We also show that we can simulate users by providing language models with privileged information. This lets us automatically evaluate multi-turn clarification dialogues. Finally, CLAM significantly improves language models' accuracy on mixed ambiguous and unambiguous questions relative to SotA.

Bayesian inference has theoretical attractions as a principled framework for reasoning about beliefs. However, the motivations of Bayesian inference which claim it to be the only 'rational' kind of reasoning do not apply in practice. They create a binary split in which all approximate inference is equally 'irrational'. Instead, we should ask ourselves how to define a spectrum of more- and less-rational reasoning that explains why we might prefer one Bayesian approximation to another. I explore approximate inference in Bayesian neural networks and consider the unintended interactions between the probabilistic model, approximating distribution, optimization algorithm, and dataset. The complexity of these interactions highlights the difficulty of any strategy for evaluating Bayesian approximations which focuses entirely on the method, outside the context of specific datasets and decision-problems. For given applications, the expected utility of the approximate posterior can measure inference quality. To assess a model's ability to incorporate different parts of the Bayesian framework we can identify desirable characteristic behaviours of Bayesian reasoning and pick decision-problems that make heavy use of those behaviours. Here, we use continual learning (testing the ability to update sequentially) and active learning (testing the ability to represent credence). But existing continual and active learning set-ups pose challenges that have nothing to do with posterior quality which can distort their ability to evaluate Bayesian approximations. These unrelated challenges can be removed or reduced, allowing better evaluation of approximate inference methods., Comment: Accepted as a thesis satisfying the requirements of a D.Phil at the Universty of Oxford

We investigate the benefit of treating all the parameters in a Bayesian neural network stochastically and find compelling theoretical and empirical evidence that this standard construction may be unnecessary. To this end, we prove that expressive predictive distributions require only small amounts of stochasticity. In particular, partially stochastic networks with only $n$ stochastic biases are universal probabilistic predictors for $n$-dimensional predictive problems. In empirical investigations, we find no systematic benefit of full stochasticity across four different inference modalities and eight datasets; partially stochastic networks can match and sometimes even outperform fully stochastic networks, despite their reduced memory costs., Comment: Published at AISTATS2023 (Oral)

Causal models of agents have been used to analyse the safety aspects of machine learning systems. But identifying agents is non-trivial -- often the causal model is just assumed by the modeler without much justification -- and modelling failures can lead to mistakes in the safety analysis. This paper proposes the first formal causal definition of agents -- roughly that agents are systems that would adapt their policy if their actions influenced the world in a different way. From this we derive the first causal discovery algorithm for discovering agents from empirical data, and give algorithms for translating between causal models and game-theoretic influence diagrams. We demonstrate our approach by resolving some previous confusions caused by incorrect causal modelling of agents., Comment: Some typos corrected

Training on web-scale data can take months. But most computation and time is wasted on redundant and noisy points that are already learnt or not learnable. To accelerate training, we introduce Reducible Holdout Loss Selection (RHO-LOSS), a simple but principled technique which selects approximately those points for training that most reduce the model's generalization loss. As a result, RHO-LOSS mitigates the weaknesses of existing data selection methods: techniques from the optimization literature typically select 'hard' (e.g. high loss) points, but such points are often noisy (not learnable) or less task-relevant. Conversely, curriculum learning prioritizes 'easy' points, but such points need not be trained on once learned. In contrast, RHO-LOSS selects points that are learnable, worth learning, and not yet learnt. RHO-LOSS trains in far fewer steps than prior art, improves accuracy, and speeds up training on a wide range of datasets, hyperparameters, and architectures (MLPs, CNNs, and BERT). On the large web-scraped image dataset Clothing-1M, RHO-LOSS trains in 18x fewer steps and reaches 2% higher final accuracy than uniform data shuffling., Comment: ICML 2022

We present a general framework for training safe agents whose naive incentives are unsafe. As an example, manipulative or deceptive behaviour can improve rewards but should be avoided. Most approaches fail here: agents maximize expected return by any means necessary. We formally describe settings with 'delicate' parts of the state which should not be used as a means to an end. We then train agents to maximize the causal effect of actions on the expected return which is not mediated by the delicate parts of state, using Causal Influence Diagram analysis. The resulting agents have no incentive to control the delicate state. We further show how our framework unifies and generalizes existing proposals., Comment: Presented at AAAI 2022

Pruning neural networks at initialization would enable us to find sparse models that retain the accuracy of the original network while consuming fewer computational resources for training and inference. However, current methods are insufficient to enable this optimization and lead to a large degradation in model performance. In this paper, we identify a fundamental limitation in the formulation of current methods, namely that their saliency criteria look at a single step at the start of training without taking into account the trainability of the network. While pruning iteratively and gradually has been shown to improve pruning performance, explicit consideration of the training stage that will immediately follow pruning has so far been absent from the computation of the saliency criterion. To overcome the short-sightedness of existing methods, we propose Prospect Pruning (ProsPr), which uses meta-gradients through the first few steps of optimization to determine which weights to prune. ProsPr combines an estimate of the higher-order effects of pruning on the loss and the optimization trajectory to identify the trainable sub-network. Our method achieves state-of-the-art pruning performance on a variety of vision classification tasks, with less data and in a single shot compared to existing pruning-at-initialization methods.

We propose Active Surrogate Estimators (ASEs), a new method for label-efficient model evaluation. Evaluating model performance is a challenging and important problem when labels are expensive. ASEs address this active testing problem using a surrogate-based estimation approach that interpolates the errors of points with unknown labels, rather than forming a Monte Carlo estimator. ASEs actively learn the underlying surrogate, and we propose a novel acquisition strategy, XWED, that tailors this learning to the final estimation task. We find that ASEs offer greater label-efficiency than the current state-of-the-art when applied to challenging model evaluation problems for deep neural networks., Comment: Accepted for publication at NeurIPS 2022