AI Benchmark Research Articles

Large Language Model Displays Emergent Ability to Interpret Novel Literary Metaphors

ABSTRACT Despite the exceptional performance of large language models (LLMs) on a wide range of tasks involving natural language processing and reasoning, there has been sharp disagreement as to whether their abilities extend to more creative human abilities. A core example is the interpretation of novel metaphors. Here we assessed the ability of GPT-4, a state-of-the-art large language model, to provide natural-language interpretations of a recent AI benchmark (Fig-QA dataset), novel literary metaphors drawn from Serbian poetry and translated into English, and entire novel English poems. GPT-4 outperformed previous AI models on the Fig-QA dataset. For metaphors drawn from Serbian poetry, human judges – blind to the fact that an AI model was involved – rated metaphor interpretations generated by GPT-4 as superior to those provided by a group of college students. In interpreting reversed metaphors, GPT-4, as well as humans, exhibited signs of sensitivity to the Gricean cooperative principle. In addition, for several novel English poems GPT-4 produced interpretations that were rated as excellent or good by a human literary critic. These results indicate that LLMs such as GPT-4 have acquired an emergent ability to interpret literary metaphors, including those embedded in novel poems.

Identifying Autism Gaze Patterns in Five-Second Data Records.

One of the most challenging problems when diagnosing autism spectrum disorder (ASD) is the need for long sets of data. Collecting data during such long periods is challenging, particularly when dealing with children. This challenge motivates the investigation of possible classifiers of ASD that do not need such long data sets. In this paper, we use eye-tracking data sets covering only 5 s and introduce one metric able to distinguish between ASD and typically developed (TD) gaze patterns based on such short time-series and compare it with two benchmarks, one using the traditional eye-tracking metrics and one state-of-the-art AI classifier. Although the data can only track possible disorders in visual attention and our approach is not a substitute to medical diagnosis, we find that our newly introduced metric can achieve an accuracy of 93% in classifying eye gaze trajectories from children with ASD surpassing both benchmarks while needing fewer data. The classification accuracy of our method, using a 5 s data series, performs better than the standard metrics in eye-tracking and is at the level of the best AI benchmarks, even when these are trained with longer time series. We also discuss the advantages and limitations of our method in comparison with the state of the art: besides needing a low amount of data, this method is a simple, understandable, and straightforward criterion to apply, which often contrasts with "black box" AI methods.

GPT-4 in Education: Evaluating Aptness, Reliability, and Loss of Coherence in Solving Calculus Problems and Grading Submissions

AbstractIn this paper, we initially investigate the capabilities of GPT-3 5 and GPT-4 in solving college-level calculus problems, an essential segment of mathematics that remains under-explored so far. Although improving upon earlier versions, GPT-4 attains approximately 65% accuracy for standard problems and decreases to 20% for competition-like scenarios. Overall, the models prove to be unreliable due to common arithmetic errors.Our primary contribution lies then in examining the use of ChatGPT for grading solutions to calculus exercises. Our objectives are to probe an in-context learning task with less emphasis over direct calculations; recognize positive applications of ChatGPT in educational contexts; highlight a potentially emerging facet of AI that could necessitate oversight; and introduce unconventional AI benchmarks, for which models like GPT are untrained. Pertaining to the latter, we uncover a tendency for loss of coherence in extended contexts. Our findings suggest that while the current ChatGPT exhibits comprehension of the grading task and often provides relevant outputs, the consistency of grading is marred by occasional loss of coherence and hallucinations. Intriguingly, GPT-4's overall scores, delivered in mere moments, align closely with human graders, although its detailed accuracy remains suboptimal.This work suggests that, when appropriately orchestrated, collaboration between human graders and LLMs like GPT-4 might combine their unique strengths while mitigating their respective shortcomings In this direction, it is imperative to consider implementing transparency, fairness, and appropriate regulations in the near future.

A seven-layer model with checklists for standardising fairness assessment throughout the AI lifecycle

Problem statement: Standardisation of AI fairness rules and benchmarks is challenging because AI fairness and other ethical requirements depend on multiple factors, such as context, usage case, type of the AI system, and so on. In this paper, we elaborate that the AI system is prone to biases at every stage of its lifecycle, from inception to its use, and that all stages require due attention for mitigating AI bias. We need a standardised approach to handle AI fairness at every stage. Gap analysis: While AI fairness is a hot research topic, a holistic strategy for AI fairness is generally missing. The AI fairness assessment is often developer-centric. Standard processes are required for independent audits for the benefit of other stakeholders. Also, most researchers focus only on a few facets of AI model-building. Peer review shows a focus on biases in the datasets, fairness metrics, and algorithmic bias. In the process, other aspects affecting AI fairness get ignored. The solution proposed: We propose a comprehensive approach in the form of a novel seven-layer model, inspired by the Open System Interconnection (OSI) model, to standardise AI fairness handling. Despite the differences in the various aspects, most AI systems have similar model-building stages. The proposed model splits the AI system lifecycle into seven abstraction layers, each corresponding to a well-defined AI model-building or usage stage. We also provide checklists for each layer and deliberate on potential sources of bias in each layer and their mitigation methodologies. This work will facilitate layer-wise standardisation of AI fairness rules and benchmarking parameters.

Edge AI-Based Tree Trunk Detection for Forestry Monitoring Robotics

Object identification, such as tree trunk detection, is fundamental for forest robotics. Intelligent vision systems are of paramount importance in order to improve robotic perception, thus enhancing the autonomy of forest robots. To that purpose, this paper presents three contributions: an open dataset of 5325 annotated forest images; a tree trunk detection Edge AI benchmark between 13 deep learning models evaluated on four edge-devices (CPU, TPU, GPU and VPU); and a tree trunk mapping experiment using an OAK-D as a sensing device. The results showed that YOLOR was the most reliable trunk detector, achieving a maximum F1 score around 90% while maintaining high scores for different confidence levels; in terms of inference time, YOLOv4 Tiny was the fastest model, attaining 1.93 ms on the GPU. YOLOv7 Tiny presented the best trade-off between detection accuracy and speed, with average inference times under 4 ms on the GPU considering different input resolutions and at the same time achieving an F1 score similar to YOLOR. This work will enable the development of advanced artificial vision systems for robotics in forestry monitoring operations.

DialFRED: Dialogue-Enabled Agents for Embodied Instruction Following

Language-guided Embodied AI benchmarks requiring an agent to navigate an environment and manipulate objects typically allow one-way communication: the human user gives a natural language command to the agent, and the agent can only follow the command passively. We present DialFRED, a dialogue-enabled embodied instruction following benchmark based on the ALFRED benchmark. DialFRED allows an agent to actively ask questions to the human user; the additional information in the user's response is used by the agent to better complete its task. We release a human-annotated dataset with 53K task-relevant questions and answers and an oracle to answer questions. To solve DialFRED, we propose a questioner-performer framework wherein the questioner is pre-trained with the human-annotated data and fine-tuned with reinforcement learning. We make DialFRED publicly available and encourage researchers to propose and evaluate their solutions to building dialog-enabled embodied agents.

Distributed Reinforcement Learning for Robot Teams: a Review

Recent advances in sensing, actuation, and computation have opened the door to multi-robot systems consisting of hundreds/thousands of robots, with promising applications to automated manufacturing, disaster relief, harvesting, last-mile delivery, port/airport operations, or search and rescue. The community has leveraged model-free multi-agent reinforcement learning (MARL) to devise efficient, scalable controllers for multi-robot systems (MRS). This review aims to provide an analysis of the state-of-the-art in distributed MARL for multi-robot cooperation. Decentralized MRS face fundamental challenges, such as non-stationarity and partial observability. Building upon the “centralized training, decentralized execution” paradigm, recent MARL approaches include independent learning, centralized critic, value decomposition, and communication learning approaches. Cooperative behaviors are demonstrated through AI benchmarks and fundamental real-world robotic capabilities such as multi-robot motion/path planning. This survey reports the challenges surrounding decentralized model-free MARL for multi-robot cooperation and existing classes of approaches. We present benchmarks and robotic applications along with a discussion on current open avenues for research.

Equilibrium Finding in Normal-Form Games via Greedy Regret Minimization

We extend the classic regret minimization framework for approximating equilibria in normal-form games by greedily weighing iterates based on regrets observed at runtime. Theoretically, our method retains all previous convergence rate guarantees. Empirically, experiments on large randomly generated games and normal-form subgames of the AI benchmark Diplomacy show that greedy weights outperforms previous methods whenever sampling is used, sometimes by several orders of magnitude.

Scenario-based AI Benchmark Evaluation of Distributed Cloud/Edge Computing Systems

Distributed cloud/edge (DCE) platform has become popular in recent years. This paper proposes a new AI benchmark suite for assessing the performance of DCE platforms in machine learning (ML) and cognitive science applications. The benchmark suite is custom-designed to satisfy scenario-based performance requirements, namely the model training time, inference speed, model accuracy, job response time, quality of service, and system reliability. These metrics are substantiated by intensive experiments with real-life AI workloads. Our work is specially tailored for supporting massive AI multitasking across distributed resources in the networking environment. Our benchmark experiments were conducted on an AI-oriented AIRS cloud built at the Chinese University of Hong Kong, Shenzhen. We have tested a large number of ML/DL programs to narrow down the inclusion of ten representative AI kernel codes in the benchmark suite. Our benchmark results reveal the advantages of using the DCE systems cost-effectively in smart cities, healthcare, community surveillance, and transportation services. Our technical contributions are in the AIRS cloud architecture, benchmark design, testing, and distributed AI computing requirements. Our work will benefit computer system designers and AI application developers on clouds, edge, and mobile devices, that are supported by 5G mobile networks and AIoT resources.

Dynamic GPU Energy Optimization for Machine Learning Training Workloads

GPUs are widely used to accelerate the training of machine learning workloads. As modern machine learning models become increasingly larger, they require a longer time to train, leading to higher GPU energy consumption. This paper presents GPOEO, an online GPU energy optimization framework for machine learning training workloads. GPOEO dynamically determines the optimal energy configuration by employing novel techniques for online measurement, multi-objective prediction modeling, and search optimization. To characterize the target workload behavior, GPOEO utilizes GPU performance counters. To reduce the performance counter profiling overhead, it uses an analytical model to detect the training iteration change and only collects performance counter data when an iteration shift is detected. GPOEO employs multi-objective models based on gradient boosting and a local search algorithm to find a trade-off between execution time and energy consumption. We evaluate the GPOEO by applying it to 71 machine learning workloads from two AI benchmark suites running on an NVIDIA RTX3080Ti GPU. Compared with the NVIDIA default scheduling strategy, GPOEO delivers a mean energy saving of 16.2% with a modest average execution time increase of 5.1%.

A Neural Network for Decision Making in Real-Time Heuristic Search

Most real-time heuristic search algorithms solve search problems by executing a series of episodes. During each episode the algorithm decides an action for execution. Such a decision is usually made using information gathered by running a bounded, heuristic-search algorithm. In this paper we report on a real-time search algorithm that does not use a search algorithm to choose the next action to be applied. Rather, it uses a neural network whose input is local information about the search graph, comparable to the information that would be used by a bounded search algorithm. We describe a supervised learning approach to training such a network. Our three types of maps from the Moving AI benchmarks, shows that our algorithm is, in some cases, substantially superior to algorithms that have access to the same information about the graph. One of our most important conclusions is that our extended set of features important: indeed, using features beyond the heuristic seems key to achieving good performance.

Decision Making Styles as Deviation from Rational Action: A Super Mario Case Study

In this paper we describe a method of modeling play styles as deviations from approximations of game theoretically rational actions. These deviations are interpreted as containing information about player skill and player decision making style. We hypothesize that this information is useful for differentiating between players and for understanding why human player behavior is attributed intentionality which we argue is a prerequisite for believability. To investigate these hypotheses we describe an experiment comparing 400 games in the Mario AI Benchmark testbed, played by humans, with equivalent games played by an approximately game theoretically rationally playing AI agent. The player actions’ deviations from the rational agent’s actions are subjected to feature extraction, and the resulting features are used to cluster play sessions into expressions of different play styles. We discuss how these styles differ, and how believable agent behavior might be approached by using these styles as an outset for a planning agent. Finally, we discuss the implications of making assumptions about rational game play and the problematic aspects of inferring player intentions from behavior.

Research community dynamics behind popular AI benchmarks

The widespread use of experimental benchmarks in AI research has created competition and collaboration dynamics that are still poorly understood. Here we provide an innovative methodology to explore these dynamics and analyse the way different entrants in these challenges, from academia to tech giants, behave and react depending on their own or others’ achievements. We perform an analysis of 25 popular benchmarks in AI from Papers With Code, with around 2,000 result entries overall, connected with their underlying research papers. We identify links between researchers and institutions (that is, communities) beyond the standard co-authorship relations, and we explore a series of hypotheses about their behaviour as well as some aggregated results in terms of activity, performance jumps and efficiency. We characterize the dynamics of research communities at different levels of abstraction, including organization, affiliation, trajectories, results and activity. We find that hybrid, multi-institution and persevering communities are more likely to improve state-of-the-art performance, which becomes a watershed for many community members. Although the results cannot be extrapolated beyond our selection of popular machine learning benchmarks, the methodology can be extended to other areas of artificial intelligence or robotics, and combined with bibliometric studies. Experimental benchmarks such as ImageNet and Atari games play an important part in advancing artificial intelligence research. An analysis of results and papers linked to 25 popular benchmarks shows that research dynamics beyond conventional co-authorship has developed in this area.

Twenty Years Beyond the Turing Test: Moving Beyond the Human Judges Too

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.

Dual Indicators to Analyze AI Benchmarks: Difficulty, Discrimination, Ability, and Generality

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.

Evolutionary Behavior Tree Approaches for Navigating Platform Games

Computer games are highly dynamic environments, where players are faced with a multitude of potentially unseen scenarios. In this paper, AI controllers are applied to the Mario AI benchmark platform, by using the grammatical evolution system to evolve behavior tree structures. These controllers are either evolved to both deal with navigation and reactiveness to elements of the game or used in conjunction with a dynamic A* approach. The results obtained highlight the applicability of behavior trees as representations for evolutionary computation and their flexibility for incorporation of diverse algorithms to deal with specific aspects of bot control in game environments.

A Panorama of Artificial and Computational Intelligence in Games

This paper attempts to give a high-level overview of the field of artificial and computational intelligence (AI/CI) in games, with particular reference to how the different core research areas within this field inform and interact with each other, both actually and potentially. We identify ten main research areas within this field: NPC behavior learning, search and planning, player modeling, games as AI benchmarks, procedural content generation, computational narrative, believable agents, AI-assisted game design, general game artificial intelligence and AI in commercial games. We view and analyze the areas from three key perspectives: 1) the dominant AI method(s) used under each area; 2) the relation of each area with respect to the end (human) user; and 3) the placement of each area within a human–computer (player-game) interaction perspective. In addition, for each of these areas we consider how it could inform or interact with each of the other areas; in those cases where we find that meaningful interaction either exists or is possible, we describe the character of that interaction and provide references to published studies, if any. We believe that this paper improves understanding of the current nature of the game AI/CI research field and the interdependences between its core areas by providing a unifying overview. We also believe that the discussion of potential interactions between research areas provides a pointer to many interesting future research projects and unexplored subfields.

The Mario AI Benchmark and Competitions

This paper describes the Mario AI benchmark, a game-based benchmark for reinforcement learning algorithms and game AI techniques developed by the authors. The benchmark is based on a public domain clone of Nintendo's classic platform game Super Mario Bros, and completely open source. During the last two years, the benchmark has been used in a number of competitions associated with international conferences, and researchers and students from around the world have contributed diverse solutions to try to beat the benchmark. The paper summarizes these contributions, gives an overview of the state of the art in Mario-playing AIs, and chronicles the development of the benchmark. This paper is intended as the definitive point of reference for those using the benchmark for research or teaching.