Nowadays, there is an increasing concern in machine learning about the causes underlying unfair decision making, that is, algorithmic decisions discriminating some groups over others, especially with ...groups that are defined over protected attributes, such as gender, race and nationality. Missing values are one frequent manifestation of all these latent causes: protected groups are more reluctant to give information that could be used against them, sensitive information for some groups can be erased by human operators, or data acquisition may simply be less complete and systematic for minority groups. However, most recent techniques, libraries and experimental results dealing with fairness in machine learning have simply ignored missing data. In this paper, we present the first comprehensive analysis of the relation between missing values and algorithmic fairness for machine learning: (1) we analyse the sources of missing data and bias, mapping the common causes, (2) we find that rows containing missing values are usually fairer than the rest, which should discourage the consideration of missing values as the uncomfortable ugly data that different techniques and libraries for handling algorithmic bias get rid of at the first occasion, (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), and (4) we show that the sensitivity of six different machine‐learning techniques to missing values is usually low, which reinforces the view that the rows with missing data contribute more to fairness through the other, nonmissing, attributes. We end the paper with a series of recommended procedures about what to do with missing data when aiming for fair decision making.
ROC curves for regression Hernández-Orallo, José
Pattern recognition,
12/2013, Letnik:
46, Številka:
12
Journal Article
Recenzirano
Odprti dostop
Receiver Operating Characteristic (ROC) analysis is one of the most popular tools for the visual assessment and understanding of classifier performance. In this paper we present a new representation ...of regression models in the so-called regression ROC (RROC) space. The basic idea is to represent over-estimation against under-estimation. The curves are just drawn by adjusting a shift, a constant that is added (or subtracted) to the predictions, and plays a similar role as a threshold in classification. From here, we develop the notions of optimal operating condition, convexity, dominance, and explore several evaluation metrics that can be shown graphically, such as the area over the RROC curve (AOC). In particular, we show a novel and significant result: the AOC is equivalent to the error variance. We illustrate the application of RROC curves to resource estimation, namely the estimation of software project effort.
•We present a natural and powerful counterpart of ROC analysis for regression.•The parallelism with classification evaluation suggests high expected applicability.•RROC space and RROC curves are derived from asymmetric loss functions.•The area over the RROC curve is shown to be equivalent to the error variance.•Predictive model behaviour will be better understood by these links (MSE, Bias, Var).
The evaluation of artificial intelligence systems and components is crucial for the progress of the discipline. In this paper we describe and critically assess the different ways AI systems are ...evaluated, and the role of components and techniques in these systems. We first focus on the traditional
task
-oriented evaluation approach. We identify three kinds of evaluation: human discrimination, problem benchmarks and peer confrontation. We describe some of the limitations of the many evaluation schemes and competitions in these three categories, and follow the progression of some of these tests. We then focus on a less customary (and challenging)
ability
-oriented evaluation approach, where a system is characterised by its (cognitive) abilities, rather than by the tasks it is designed to solve. We discuss several possibilities: the adaptation of cognitive tests used for humans and animals, the development of tests derived from algorithmic information theory or more integrated approaches under the perspective of universal psychometrics. We analyse some evaluation tests from AI that are better positioned for an ability-oriented evaluation and discuss how their problems and limitations can possibly be addressed with some of the tools and ideas that appear within the paper. Finally, we enumerate a series of lessons learnt and generic guidelines to be used when an AI evaluation scheme is under consideration.
We report on a series of new platforms and events dealing with AI evaluation that may change the way in which AI systems are compared and their progress is measured. The introduction of a more ...diverse and challenging set of tasks in these platforms can feed AI research in the years to come, shaping the notion of success and the directions of the field. However, the playground of tasks and challenges presented there may misdirect the field without some meaningful structure and systematic guidelines for its organization and use. Anticipating this issue, we also report on several initiatives and workshops that are putting the focus on analyzing the similarity and dependencies between tasks, their difficulty, what capabilities they really measure and — ultimately — on elaborating new concepts and tools that can arrange tasks and benchmarks into a meaningful taxonomy.
Making sense of sensory input Evans, Richard; Hernández-Orallo, José; Welbl, Johannes ...
Artificial intelligence,
April 2021, 2021-04-00, 20210401, Letnik:
293
Journal Article
Recenzirano
Odprti dostop
This paper attempts to answer a central question in unsupervised learning: what does it mean to “make sense” of a sensory sequence? In our formalization, making sense involves constructing a symbolic ...causal theory that both explains the sensory sequence and also satisfies a set of unity conditions. The unity conditions insist that the constituents of the causal theory – objects, properties, and laws – must be integrated into a coherent whole. On our account, making sense of sensory input is a type of program synthesis, but it is unsupervised program synthesis.
Our second contribution is a computer implementation, the Apperception Engine, that was designed to satisfy the above requirements. Our system is able to produce interpretable human-readable causal theories from very small amounts of data, because of the strong inductive bias provided by the unity conditions. A causal theory produced by our system is able to predict future sensor readings, as well as retrodict earlier readings, and impute (fill in the blanks of) missing sensory readings, in any combination. In fact, it is able to do all three tasks simultaneously.
We tested the engine in a diverse variety of domains, including cellular automata, rhythms and simple nursery tunes, multi-modal binding problems, occlusion tasks, and sequence induction intelligence tests. In each domain, we test our engine's ability to predict future sensor values, retrodict earlier sensor values, and impute missing sensory data. The Apperception Engine performs well in all these domains, significantly out-performing neural net baselines. We note in particular that in the sequence induction intelligence tests, our system achieved human-level performance. This is notable because our system is not a bespoke system designed specifically to solve intelligence tests, but a general-purpose system that was designed to make sense of any sensory sequence.
In curriculum learning the order of concepts is determined by the teacher but not the examples for each concept, while in machine teaching it is the examples that are chosen by the teacher to ...minimise the learning effort, though the concepts are taught in isolation. Curriculum teaching is the natural combination of both, where both concept order and the set of examples can be chosen to minimise the size of the whole teaching session. Yet, this simultaneous minimisation of teaching sets and concept order is computationally challenging, facing issues such as the “interposition” phenomenon: previous knowledge may be counter-productive. We build on a machine-teaching framework based on simplicity priors that can achieve short teaching sizes for large classes of languages. Given a set of concepts, we identify an inequality relating the sizes of example sets and concept descriptions. This leverages the definition of admissible heuristics for A* search to spot the optimal curricula by avoiding interposition, being able to find the shortest teaching sessions in a more efficient way than an exhaustive search and with the guarantees we do not have with a greedy algorithm. We illustrate these theoretical findings through case studies in a drawing domain, polygonal strokes on a grid described by a simple language implementing compositionality and recursion.
In this paper, we develop the idea of a universal anytime intelligence test. The meaning of the terms “universal” and “anytime” is manifold here: the test should be able to measure the intelligence ...of any biological or artificial system that exists at this time or in the future. It should also be able to evaluate both inept and brilliant systems (any intelligence level) as well as very slow to very fast systems (any time scale). Also, the test may be interrupted at any time, producing an approximation to the intelligence score, in such a way that the more time is left for the test, the better the assessment will be. In order to do this, our test proposal is based on previous works on the measurement of machine intelligence based on Kolmogorov complexity and universal distributions, which were developed in the late 1990s (C-tests and compression-enhanced Turing tests). It is also based on the more recent idea of measuring intelligence through dynamic/interactive tests held against a universal distribution of environments. We discuss some of these tests and highlight their limitations since we want to construct a test that is both general and practical. Consequently, we introduce many new ideas that develop early “compression tests” and the more recent definition of “universal intelligence” in order to design new “universal intelligence tests”, where a feasible implementation has been a design requirement. One of these tests is the “anytime intelligence test”, which adapts to the examinee's level of intelligence in order to obtain an intelligence score within a limited time.
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 analyse 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.
In the last 20 years the Turing test has been left further behind by new developments in artificial intelligence. At the same time, however, these developments have revived some key elements of the ...Turing test: imitation and adversarialness. On the one hand, many generative models, such as generative adversarial networks (GAN), build imitators under an adversarial setting that strongly resembles the Turing test (with the judge being a learnt discriminative model). The term “Turing learning” has been used for this kind of setting. On the other hand, AI benchmarks are suffering an adversarial situation too, with a ‘challenge-solve-and-replace’ evaluation dynamics whenever human performance is ‘imitated’. The particular AI community rushes to replace the old benchmark by a more challenging benchmark, one for which human performance would still be beyond AI. These two phenomena related to the Turing test are sufficiently distinctive, important and general for a detailed analysis. This is the main goal of this paper. After recognising the abyss that appears beyond superhuman performance, we build on Turing learning to identify two different evaluation schemas: Turing testing and adversarial testing. We revisit some of the key questions surrounding the Turing test, such as ‘understanding’, commonsense reasoning and extracting meaning from the world, and explore how the new testing paradigms should work to unmask the limitations of current and future AI. Finally, we discuss how behavioural similarity metrics could be used to create taxonomies for artificial and natural intelligence. Both testing schemas should complete a transition in which humans should give way to machines—not only as references to be imitated but also as judges—when pursuing and measuring machine intelligence.
Success in all sorts of situations is the most classical interpretation of general intelligence. Under limited resources, however, the capability of an agent must necessarily be limited too, and ...generality needs to be understood as comprehensive performance up to a level of difficulty. The degree of generality then refers to the way an agent's capability is distributed as a function of task difficulty. This dissects the notion of general intelligence into two non-populational measures, generality and capability, which we apply to individuals and groups of humans, other animals and AI systems, on several cognitive and perceptual tests. Our results indicate that generality and capability can decouple at the individual level: very specialised agents can show high capability and vice versa. The metrics also decouple at the population level, and we rarely see diminishing returns in generality for those groups of high capability. We relate the individual measure of generality to traditional notions of general intelligence and cognitive efficiency in humans, collectives, non-human animals and machines. The choice of the difficulty function now plays a prominent role in this new conception of generality, which brings a quantitative tool for shedding light on long-standing questions about the evolution of general intelligence and the evaluation of progress in Artificial General Intelligence.
PDF
