Fernando Martínez-Plumed

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Lexin Zhou, Pablo A. Moreno-Casares, Fernando Martínez-Plumed et al. · cambridge

We introduce the fundamental ideas and challenges of Predictable AI, a nascent research area that explores the ways in which we can anticipate key validity indicators (e.g., performance, safety) of present and future AI ecosystems. We argue that achieving predictability is crucial for fostering trust, liability, control, alignment and safety of AI ecosystems, and thus should be prioritised over performance. We formally characterise predictability, explore its most relevant components, illustrate what can be predicted, describe alternative candidates for predictors, as well as the trade-offs between maximising validity and predictability. To illustrate these concepts, we bring an array of illustrative examples covering diverse ecosystem configurations. Predictable AI is related to other areas of technical and non-technical AI research, but have distinctive questions, hypotheses, techniques and challenges. This paper aims to elucidate them, calls for identifying paths towards a landscape of predictably valid AI systems and outlines the potential impact of this emergent field.

Lorenzo Pacchiardi, Konstantinos Voudouris, Ben Slater et al. · cambridge

Despite possessing impressive skills, Large Language Models (LLMs) often fail unpredictably, demonstrating inconsistent success in even basic common sense reasoning tasks. This unpredictability poses a significant challenge to ensuring their safe deployment, as identifying and operating within a reliable "safe zone" is essential for mitigating risks. To address this, we present PredictaBoard, a novel collaborative benchmarking framework designed to evaluate the ability of score predictors (referred to as assessors) to anticipate LLM errors on specific task instances (i.e., prompts) from existing datasets. PredictaBoard evaluates pairs of LLMs and assessors by considering the rejection rate at different tolerance errors. As such, PredictaBoard stimulates research into developing better assessors and making LLMs more predictable, not only with a higher average performance. We conduct illustrative experiments using baseline assessors and state-of-the-art LLMs. PredictaBoard highlights the critical need to evaluate predictability alongside performance, paving the way for safer AI systems where errors are not only minimised but also anticipated and effectively mitigated. Code for our benchmark can be found at https://github.com/Kinds-of-Intelligence-CFI/PredictaBoard

Lexin Zhou, Lorenzo Pacchiardi, Fernando Martínez-Plumed et al. · cambridge

Ensuring safe and effective use of AI requires understanding and anticipating its performance on novel tasks, from advanced scientific challenges to transformed workplace activities. So far, benchmarking has guided progress in AI, but it has offered limited explanatory and predictive power for general-purpose AI systems, given the low transferability across diverse tasks. In this paper, we introduce general scales for AI evaluation that can explain what common AI benchmarks really measure, extract ability profiles of AI systems, and predict their performance for new task instances, in- and out-of-distribution. Our fully-automated methodology builds on 18 newly-crafted rubrics that place instance demands on general scales that do not saturate. Illustrated for 15 large language models and 63 tasks, high explanatory power is unleashed from inspecting the demand and ability profiles, bringing insights on the sensitivity and specificity exhibited by different benchmarks, and how knowledge, metacognition and reasoning are affected by model size, chain-of-thought and distillation. Surprisingly, high predictive power at the instance level becomes possible using these demand levels, providing superior estimates over black-box baseline predictors based on embeddings or finetuning, especially in out-of-distribution settings (new tasks and new benchmarks). The scales, rubrics, battery, techniques and results presented here represent a major step for AI evaluation, underpinning the reliable deployment of AI in the years ahead. (Collaborative platform: https://kinds-of-intelligence-cfi.github.io/ADELE.)

Peter Romero, Fernando Martínez-Plumed, Zachary R. Tyler et al.

Comparing AI models to "human level" is often misleading when benchmark scores are incommensurate or human baselines are drawn from a narrow population. To address this, we propose a framework that calibrates items against the 'world population' and report performance on a common, human-anchored scale. Concretely, we build on a set of multi-level scales for different capabilities where each level should represent a probability of success of the whole world population on a logarithmic scale with a base $B$. We calibrate each scale for each capability (reasoning, comprehension, knowledge, volume, etc.) by compiling publicly released human test data spanning education and reasoning benchmarks (PISA, TIMSS, ICAR, UKBioBank, and ReliabilityBench). The base $B$ is estimated by extrapolating between samples with two demographic profiles using LLMs, with the hypothesis that they condense rich information about human populations. We evaluate the quality of different mappings using group slicing and post-stratification. The new techniques allow for the recalibration and standardization of scales relative to the whole-world population.

Daniel Romero-Alvarado, Fernando Martínez-Plumed, Lorenzo Pacchiardi et al.

AI evaluation has primarily focused on measuring capabilities, with formal approaches inspired from Item Response Theory (IRT) being increasingly applied. Yet propensities - the tendencies of models to exhibit particular behaviours - play a central role in determining both performance and safety outcomes. However, traditional IRT describes a model's success on a task as a monotonic function of model capabilities and task demands, an approach unsuited to propensities, where both excess and deficiency can be problematic. Here, we introduce the first formal framework for measuring AI propensities by using a bilogistic formulation for model success, which attributes high success probability when the model's propensity is within an "ideal band". Further, we estimate the limits of the ideal band using LLMs equipped with newly developed task-agnostic rubrics. Applying our framework to six families of LLM models whose propensities are incited in either direction, we find that we can measure how much the propensity is shifted and what effect this has on the tasks. Critically, propensities estimated using one benchmark successfully predict behaviour on held-out tasks. Moreover, we obtain stronger predictive power when combining propensities and capabilities than either separately. More broadly, our framework showcases how rigorous propensity measurements can be conducted and how it yields gains over solely using capability evaluations to predict AI behaviour.

Daniel Romero-Alvarado, Fernando Martínez-Plumed, José Hernández-Orallo

An AI assessor is an external, ideally indepen-dent system that predicts an indicator, e.g., a loss value, of another AI system. Assessors can lever-age information from the test results of many other AI systems and have the flexibility of be-ing trained on any loss function or scoring rule: from squared error to toxicity metrics. Here we address the question: is it always optimal to train the assessor for the target metric? Or could it be better to train for a different metric and then map predictions back to the target metric? Us-ing twenty regression and classification problems with tabular data, we experimentally explore this question for, respectively, regression losses and classification scores with monotonic and non-monotonic mappings and find that, contrary to intuition, optimising for more informative met-rics is not generally better. Surprisingly, some monotonic transformations are promising. For example, the logistic loss is useful for minimis-ing absolute or quadratic errors in regression, and the logarithmic score helps maximise quadratic or spherical scores in classification.

Radosvet Desislavov, Fernando Martínez-Plumed, José Hernández-Orallo

The progress of some AI paradigms such as deep learning is said to be linked to an exponential growth in the number of parameters. There are many studies corroborating these trends, but does this translate into an exponential increase in energy consumption? In order to answer this question we focus on inference costs rather than training costs, as the former account for most of the computing effort, solely because of the multiplicative factors. Also, apart from algorithmic innovations, we account for more specific and powerful hardware (leading to higher FLOPS) that is usually accompanied with important energy efficiency optimisations. We also move the focus from the first implementation of a breakthrough paper towards the consolidated version of the techniques one or two year later. Under this distinctive and comprehensive perspective, we study relevant models in the areas of computer vision and natural language processing: for a sustained increase in performance we see a much softer growth in energy consumption than previously anticipated. The only caveat is, yet again, the multiplicative factor, as future AI increases penetration and becomes more pervasive.

Fernando Martínez-Plumed, Cèsar Ferri, David Nieves et al.

The causes underlying unfair decision making are complex, being internalised in different ways by decision makers, other actors dealing with data and models, and ultimately by the individuals being affected by these decisions. One frequent manifestation of all these latent causes arises in the form of missing values: protected groups are more reluctant to give information that could be used against them, delicate information for some groups can be erased by human operators, or data acquisition may simply be less complete and systematic for minority groups. As a result, missing values and bias in data are two phenomena that are tightly coupled. However, most recent techniques, libraries and experimental results dealing with fairness in machine learning have simply ignored missing data. In this paper, we claim that fairness research should not miss the opportunity to deal properly with missing data. To support this claim, (1) we analyse the sources of missing data and bias, and we map the common causes, (2) we find that rows containing missing values are usually fairer than the rest, which should not be treated as the uncomfortable ugly data that different techniques and libraries get rid of at the first occasion, and (3) we study the trade-off between performance and fairness when the rows with missing values are used (either because the technique deals with them directly or by imputation methods). We end the paper with a series of recommended procedures about what to do with missing data when aiming for fair decision making.

Fernando Martínez-Plumed, José Hernández-Orallo

Item response theory (IRT) can be applied to the analysis of the evaluation of results from AI benchmarks. The two-parameter IRT model provides two indicators (difficulty and discrimination) on the side of the item (or AI problem) while only one indicator (ability) on the side of the respondent (or AI agent). In this paper we analyse how to make this set of indicators dual, by adding a fourth indicator, generality, on the side of the respondent. Generality is meant to be dual to discrimination, and it is based on difficulty. Namely, generality is defined as a new metric that evaluates whether an agent is consistently good at easy problems and bad at difficult ones. With the addition of generality, we see that this set of four key indicators can give us more insight on the results of AI benchmarks. In particular, we explore two popular benchmarks in AI, the Arcade Learning Environment (Atari 2600 games) and the General Video Game AI competition. We provide some guidelines to estimate and interpret these indicators for other AI benchmarks and competitions.

Lidia Contreras-Ochando, César Ferri, José Hernández-Orallo et al.

Given one or two examples, humans are good at understanding how to solve a problem independently of its domain, because they are able to detect what the problem is and to choose the appropriate background knowledge according to the context. For instance, presented with the string "8/17/2017" to be transformed to "17th of August of 2017", humans will process this in two steps: (1) they recognise that it is a date and (2) they map the date to the 17th of August of 2017. Inductive Programming (IP) aims at learning declarative (functional or logic) programs from examples. Two key advantages of IP are the use of background knowledge and the ability to synthesise programs from a few input/output examples (as humans do). In this paper we propose to use IP as a means for automating repetitive data manipulation tasks, frequently presented during the process of {\em data wrangling} in many data manipulation problems. Here we show that with the use of general-purpose declarative (programming) languages jointly with generic IP systems and the definition of domain-specific knowledge, many specific data wrangling problems from different application domains can be automatically solved from very few examples. We also propose an integrated benchmark for data wrangling, which we share publicly for the community.

Enrique Fernández-Macías, Emilia Gómez, José Hernández-Orallo et al.

This paper presents a multidisciplinary task approach for assessing the impact of artificial intelligence on the future of work. We provide definitions of a task from two main perspectives: socio-economic and computational. We propose to explore ways in which we can integrate or map these perspectives, and link them with the skills or capabilities required by them, for humans and AI systems. Finally, we argue that in order to understand the dynamics of tasks, we have to explore the relevance of autonomy and generality of AI systems for the automation or alteration of the workplace.

Fernando Martínez-Plumed, Shahar Avin, Miles Brundage et al.

We reframe the analysis of progress in AI by incorporating into an overall framework both the task performance of a system, and the time and resource costs incurred in the development and deployment of the system. These costs include: data, expert knowledge, human oversight, software resources, computing cycles, hardware and network facilities, and (what kind of) time. These costs are distributed over the life cycle of the system, and may place differing demands on different developers and users. The multidimensional performance and cost space we present can be collapsed to a single utility metric that measures the value of the system for different stakeholders. Even without a single utility function, AI advances can be generically assessed by whether they expand the Pareto surface. We label these types of costs as neglected dimensions of AI progress, and explore them using four case studies: Alpha* (Go, Chess, and other board games), ALE (Atari games), ImageNet (Image classification) and Virtual Personal Assistants (Siri, Alexa, Cortana, and Google Assistant). This broader model of progress in AI will lead to novel ways of estimating the potential societal use and impact of an AI system, and the establishment of milestones for future progress.

Fernando Martínez-Plumed, Cèsar Ferri, José Hernández-Orallo et al.

The application of cognitive mechanisms to support knowledge acquisition is, from our point of view, crucial for making the resulting models coherent, efficient, credible, easy to use and understandable. In particular, there are two characteristic features of intelligence that are essential for knowledge development: forgetting and consolidation. Both plays an important role in knowledge bases and learning systems to avoid possible information overflow and redundancy, and in order to preserve and strengthen important or frequently used rules and remove (or forget) useless ones. We present an incremental, long-life view of knowledge acquisition which tries to improve task after task by determining what to keep, what to consolidate and what to forget, overcoming The Stability-Plasticity dilemma. In order to do that, we rate rules by introducing several metrics through the first adaptation, to our knowledge, of the Minimum Message Length (MML) principle to a coverage graph, a hierarchical assessment structure which treats evidence and rules in a unified way. The metrics are not only used to forget some of the worst rules, but also to set a consolidation process to promote those selected rules to the knowledge base, which is also mirrored by a demotion system. We evaluate the framework with a series of tasks in a chess rule learning domain.

Fernando Martínez-Plumed, Cèsar Ferri, José Hernández-Orallo et al.

In this paper, we push forward the idea of machine learning systems whose operators can be modified and fine-tuned for each problem. This allows us to propose a learning paradigm where users can write (or adapt) their operators, according to the problem, data representation and the way the information should be navigated. To achieve this goal, data instances, background knowledge, rules, programs and operators are all written in the same functional language, Erlang. Since changing operators affect how the search space needs to be explored, heuristics are learnt as a result of a decision process based on reinforcement learning where each action is defined as a choice of operator and rule. As a result, the architecture can be seen as a 'system for writing machine learning systems' or to explore new operators where the policy reuse (as a kind of transfer learning) is allowed. States and actions are represented in a Q matrix which is actually a table, from which a supervised model is learnt. This makes it possible to have a more flexible mapping between old and new problems, since we work with an abstraction of rules and actions. We include some examples sharing reuse and the application of the system gErl to IQ problems. In order to evaluate gErl, we will test it against some structured problems: a selection of IQ test tasks and some experiments on some structured prediction problems (list patterns).