Melanie Mitchell

Michael L. Littman, Ifeoma Ajunwa, Guy Berger et al.

In September 2021, the "One Hundred Year Study on Artificial Intelligence" project (AI100) issued the second report of its planned long-term periodic assessment of artificial intelligence (AI) and its impact on society. It was written by a panel of 17 study authors, each of whom is deeply rooted in AI research, chaired by Michael Littman of Brown University. The report, entitled "Gathering Strength, Gathering Storms," answers a set of 14 questions probing critical areas of AI development addressing the major risks and dangers of AI, its effects on society, its public perception and the future of the field. The report concludes that AI has made a major leap from the lab to people's lives in recent years, which increases the urgency to understand its potential negative effects. The questions were developed by the AI100 Standing Committee, chaired by Peter Stone of the University of Texas at Austin, consisting of a group of AI leaders with expertise in computer science, sociology, ethics, economics, and other disciplines.

Melanie Mitchell, David C. Krakauer

We survey a current, heated debate in the AI research community on whether large pre-trained language models can be said to "understand" language -- and the physical and social situations language encodes -- in any important sense. We describe arguments that have been made for and against such understanding, and key questions for the broader sciences of intelligence that have arisen in light of these arguments. We contend that a new science of intelligence can be developed that will provide insight into distinct modes of understanding, their strengths and limitations, and the challenge of integrating diverse forms of cognition.

Victor Vikram Odouard, Melanie Mitchell

A long-held objective in AI is to build systems that understand concepts in a humanlike way. Setting aside the difficulty of building such a system, even trying to evaluate one is a challenge, due to present-day AI's relative opacity and its proclivity for finding shortcut solutions. This is exacerbated by humans' tendency to anthropomorphize, assuming that a system that can recognize one instance of a concept must also understand other instances, as a human would. In this paper, we argue that understanding a concept requires the ability to use it in varied contexts. Accordingly, we propose systematic evaluations centered around concepts, by probing a system's ability to use a given concept in many different instantiations. We present case studies of such an evaluations on two domains -- RAVEN (inspired by Raven's Progressive Matrices) and the Abstraction and Reasoning Corpus (ARC) -- that have been used to develop and assess abstraction abilities in AI systems. Our concept-based approach to evaluation reveals information about AI systems that conventional test sets would have left hidden.

Melanie Mitchell, Alessandro B. Palmarini, Arseny Moskvichev

We explore the abstract reasoning abilities of text-only and multimodal versions of GPT-4, using the ConceptARC benchmark [10], which is designed to evaluate robust understanding and reasoning with core-knowledge concepts. We extend the work of Moskvichev et al. [10] by evaluating GPT-4 on more detailed, one-shot prompting (rather than simple, zero-shot prompts) with text versions of ConceptARC tasks, and by evaluating GPT-4V, the multimodal version of GPT-4, on zero- and one-shot prompts using image versions of the simplest tasks. Our experimental results support the conclusion that neither version of GPT-4 has developed robust abstraction abilities at humanlike levels.

Tyler Millhouse, Melanie Moses, Melanie Mitchell

In April of 2022, the Santa Fe Institute hosted a workshop on embodied, situated, and grounded intelligence as part of the Institute's Foundations of Intelligence project. The workshop brought together computer scientists, psychologists, philosophers, social scientists, and others to discuss the science of embodiment and related issues in human intelligence, and its implications for building robust, human-level AI. In this report, we summarize each of the talks and the subsequent discussions. We also draw out a number of key themes and identify important frontiers for future research.

Martha Lewis, Melanie Mitchell

Large language models (LLMs) have performed well on several reasoning benchmarks, including ones that test analogical reasoning abilities. However, it has been debated whether they are actually performing humanlike abstract reasoning or instead employing less general processes that rely on similarity to what has been seen in their training data. Here we investigate the generality of analogy-making abilities previously claimed for LLMs (Webb, Holyoak, & Lu, 2023). We take one set of analogy problems used to evaluate LLMs and create a set of "counterfactual" variants-versions that test the same abstract reasoning abilities but that are likely dissimilar from any pre-training data. We test humans and three GPT models on both the original and counterfactual problems, and show that, while the performance of humans remains high for all the problems, the GPT models' performance declines sharply on the counterfactual set. This work provides evidence that, despite previously reported successes of LLMs on analogical reasoning, these models lack the robustness and generality of human analogy-making.

Martha Lewis, Melanie Mitchell

LLMs have performed well on several reasoning benchmarks, including ones that test analogical reasoning abilities. However, there is debate on the extent to which they are performing general abstract reasoning versus employing non-robust processes, e.g., that overly rely on similarity to pre-training data. Here we investigate the robustness of analogy-making abilities previously claimed for LLMs on three of four domains studied by Webb, Holyoak, and Lu (2023): letter-string analogies, digit matrices, and story analogies. For each domain we test humans and GPT models on robustness to variants of the original analogy problems that test the same abstract reasoning abilities but are likely dissimilar from tasks in the pre-training data. The performance of a system that uses robust abstract reasoning should not decline substantially on these variants. On simple letter-string analogies, we find that while the performance of humans remains high for two types of variants we tested, the GPT models' performance declines sharply. This pattern is less pronounced as the complexity of these problems is increased, as both humans and GPT models perform poorly on both the original and variant problems requiring more complex analogies. On digit-matrix problems, we find a similar pattern but only on one out of the two types of variants we tested. On story-based analogy problems, we find that, unlike humans, the performance of GPT models are susceptible to answer-order effects, and that GPT models also may be more sensitive than humans to paraphrasing. This work provides evidence that LLMs often lack the robustness of zero-shot human analogy-making, exhibiting brittleness on most of the variations we tested. More generally, this work points to the importance of carefully evaluating AI systems not only for accuracy but also robustness when testing their cognitive capabilities.

Samuel G. B. Johnson, Amir-Hossein Karimi, Yoshua Bengio et al.

Although AI has become increasingly smart, its wisdom has not kept pace. In this article, we examine what is known about human wisdom and sketch a vision of its AI counterpart. We analyze human wisdom as a set of strategies for solving intractable problems-those outside the scope of analytic techniques-including both object-level strategies like heuristics [for managing problems] and metacognitive strategies like intellectual humility, perspective-taking, or context-adaptability [for managing object-level strategies]. We argue that AI systems particularly struggle with metacognition; improved metacognition would lead to AI more robust to novel environments, explainable to users, cooperative with others, and safer in risking fewer misaligned goals with human users. We discuss how wise AI might be benchmarked, trained, and implemented.

Claire E. Stevenson, Alexandra Pafford, Han L. J. van der Maas et al.

In people, the ability to solve analogies such as "body : feet :: table : ?" emerges in childhood, and appears to transfer easily to other domains, such as the visual domain "( : ) :: < : ?". Recent research shows that large language models (LLMs) can solve various forms of analogies. However, can LLMs generalize analogy solving to new domains like people can? To investigate this, we had children, adults, and LLMs solve a series of letter-string analogies (e.g., a b : a c :: j k : ?) in the Latin alphabet, in a near transfer domain (Greek alphabet), and a far transfer domain (list of symbols). Children and adults easily generalized their knowledge to unfamiliar domains, whereas LLMs did not. This key difference between human and AI performance is evidence that these LLMs still struggle with robust human-like analogical transfer.

Micah Goldblum, Anima Anandkumar, Richard Baraniuk et al.

The goal of this series is to chronicle opinions and issues in the field of machine learning as they stand today and as they change over time. The plan is to host this survey periodically until the AI singularity paperclip-frenzy-driven doomsday, keeping an updated list of topical questions and interviewing new community members for each edition. In this issue, we probed people's opinions on interpretable AI, the value of benchmarking in modern NLP, the state of progress towards understanding deep learning, and the future of academia.

David C. Krakauer, John W. Krakauer, Melanie Mitchell

Emergence is a concept in complexity science that describes how many-body systems manifest novel higher-level properties, properties that can be described by replacing high-dimensional mechanisms with lower-dimensional effective variables and theories. This is captured by the idea "more is different". Intelligence is a consummate emergent property manifesting increasingly efficient -- cheaper and faster -- uses of emergent capabilities to solve problems. This is captured by the idea "less is more". In this paper, we first examine claims that Large Language Models exhibit emergent capabilities, reviewing several approaches to quantifying emergence, and secondly ask whether LLMs possess emergent intelligence.

Claas Beger, Ryan Yi, Shuhao Fu et al.

OpenAI's o3-preview reasoning model exceeded human accuracy on the ARC-AGI benchmark, but does that mean state-of-the-art models recognize and reason with the abstractions that the task creators intended? We investigate models' abstraction abilities on ConceptARC. We evaluate models under settings that vary the input modality (textual vs. visual), whether the model is permitted to use external Python tools, and, for reasoning models, the amount of reasoning effort. In addition to measuring output accuracy, we perform fine-grained evaluation of the natural-language rules that models generate to explain their solutions. This dual evaluation lets us assess whether models solve tasks using the abstractions ConceptARC was designed to elicit, rather than relying on surface-level patterns. Our results show that, while some models using text-based representations match human output accuracy, the best models' rules are often based on surface-level ``shortcuts'' and capture intended abstractions far less often than humans. Thus their capabilities for general abstract reasoning may be overestimated by evaluations based on accuracy alone. In the visual modality, AI models' output accuracy drops sharply, yet our rule-level analysis reveals that models might be underestimated, as they still exhibit a substantial share of rules that capture intended abstractions, but are often unable to correctly apply these rules. In short, our results show that models still lag humans in abstract reasoning, and that using accuracy alone to evaluate abstract reasoning on ARC-like tasks may overestimate abstract-reasoning capabilities in textual modalities and underestimate it in visual modalities. We believe that our evaluation framework offers a more faithful picture of multimodal models' abstract reasoning abilities and a more principled way to track progress toward human-like, abstraction-centered intelligence.

Arseny Moskvichev, Victor Vikram Odouard, Melanie Mitchell

The abilities to form and abstract concepts is key to human intelligence, but such abilities remain lacking in state-of-the-art AI systems. There has been substantial research on conceptual abstraction in AI, particularly using idealized domains such as Raven's Progressive Matrices and Bongard problems, but even when AI systems succeed on such problems, the systems are rarely evaluated in depth to see if they have actually grasped the concepts they are meant to capture. In this paper we describe an in-depth evaluation benchmark for the Abstraction and Reasoning Corpus (ARC), a collection of few-shot abstraction and analogy problems developed by Chollet [2019]. In particular, we describe ConceptARC, a new, publicly available benchmark in the ARC domain that systematically assesses abstraction and generalization abilities on a number of basic spatial and semantic concepts. ConceptARC differs from the original ARC dataset in that it is specifically organized around "concept groups" -- sets of problems that focus on specific concepts and that are vary in complexity and level of abstraction. We report results on testing humans on this benchmark as well as three machine solvers: the top two programs from a 2021 ARC competition and OpenAI's GPT-4. Our results show that humans substantially outperform the machine solvers on this benchmark, showing abilities to abstract and generalize concepts that are not yet captured by AI systems. We believe that this benchmark will spur improvements in the development of AI systems for conceptual abstraction and in the effective evaluation of such systems.

Murray Shanahan, Melanie Mitchell

We characterise the problem of abstraction in the context of deep reinforcement learning. Various well established approaches to analogical reasoning and associative memory might be brought to bear on this issue, but they present difficulties because of the need for end-to-end differentiability. We review developments in AI and machine learning that could facilitate their adoption.

Tyler Millhouse, Melanie Moses, Melanie Mitchell

In August of 2021, the Santa Fe Institute hosted a workshop on collective intelligence as part of its Foundations of Intelligence project. This project seeks to advance the field of artificial intelligence by promoting interdisciplinary research on the nature of intelligence. The workshop brought together computer scientists, biologists, philosophers, social scientists, and others to share their insights about how intelligence can emerge from interactions among multiple agents--whether those agents be machines, animals, or human beings. In this report, we summarize each of the talks and the subsequent discussions. We also draw out a number of key themes and identify important frontiers for future research.

Tyler Millhouse, Melanie Moses, Melanie Mitchell

In July of 2021, the Santa Fe Institute hosted a workshop on evolutionary computation as part of its Foundations of Intelligence in Natural and Artificial Systems project. This project seeks to advance the field of artificial intelligence by promoting interdisciplinary research on the nature of intelligence. The workshop brought together computer scientists and biologists to share their insights about the nature of evolution and the future of evolutionary computation. In this report, we summarize each of the talks and the subsequent discussions. We also draw out a number of key themes and identify important frontiers for future research.

Jacob M. Springer, Melanie Mitchell, Garrett T. Kenyon

Adversarial examples for neural network image classifiers are known to be transferable: examples optimized to be misclassified by a source classifier are often misclassified as well by classifiers with different architectures. However, targeted adversarial examples -- optimized to be classified as a chosen target class -- tend to be less transferable between architectures. While prior research on constructing transferable targeted attacks has focused on improving the optimization procedure, in this work we examine the role of the source classifier. Here, we show that training the source classifier to be "slightly robust" -- that is, robust to small-magnitude adversarial examples -- substantially improves the transferability of class-targeted and representation-targeted adversarial attacks, even between architectures as different as convolutional neural networks and transformers. The results we present provide insight into the nature of adversarial examples as well as the mechanisms underlying so-called "robust" classifiers.

Tyler Millhouse, Melanie Moses, Melanie Mitchell

In March of 2021, the Santa Fe Institute hosted a workshop as part of its Foundations of Intelligence in Natural and Artificial Systems project. This project seeks to advance the field of artificial intelligence by promoting interdisciplinary research on the nature of intelligence. During the workshop, speakers from diverse disciplines gathered to develop a taxonomy of intelligence, articulating their own understanding of intelligence and how their research has furthered that understanding. In this report, we summarize the insights offered by each speaker and identify the themes that emerged during the talks and subsequent discussions.

Melanie Mitchell

Since its beginning in the 1950s, the field of artificial intelligence has cycled several times between periods of optimistic predictions and massive investment ("AI spring") and periods of disappointment, loss of confidence, and reduced funding ("AI winter"). Even with today's seemingly fast pace of AI breakthroughs, the development of long-promised technologies such as self-driving cars, housekeeping robots, and conversational companions has turned out to be much harder than many people expected. One reason for these repeating cycles is our limited understanding of the nature and complexity of intelligence itself. In this paper I describe four fallacies in common assumptions made by AI researchers, which can lead to overconfident predictions about the field. I conclude by discussing the open questions spurred by these fallacies, including the age-old challenge of imbuing machines with humanlike common sense.

Melanie Mitchell

Conceptual abstraction and analogy-making are key abilities underlying humans' abilities to learn, reason, and robustly adapt their knowledge to new domains. Despite of a long history of research on constructing AI systems with these abilities, no current AI system is anywhere close to a capability of forming humanlike abstractions or analogies. This paper reviews the advantages and limitations of several approaches toward this goal, including symbolic methods, deep learning, and probabilistic program induction. The paper concludes with several proposals for designing challenge tasks and evaluation measures in order to make quantifiable and generalizable progress in this area.

Jacob M. Springer, Melanie Mitchell, Garrett T. Kenyon

Neural networks trained on visual data are well-known to be vulnerable to often imperceptible adversarial perturbations. The reasons for this vulnerability are still being debated in the literature. Recently Ilyas et al. (2019) showed that this vulnerability arises, in part, because neural network classifiers rely on highly predictive but brittle "non-robust" features. In this paper we extend the work of Ilyas et al. by investigating the nature of the input patterns that give rise to these features. In particular, we hypothesize that in a neural network trained in a standard way, non-robust features respond to small, "non-semantic" patterns that are typically entangled with larger, robust patterns, known to be more human-interpretable, as opposed to solely responding to statistical artifacts in a dataset. Thus, adversarial examples can be formed via minimal perturbations to these small, entangled patterns. In addition, we demonstrate a corollary of our hypothesis: robust classifiers are more effective than standard (non-robust) ones as a source for generating transferable adversarial examples in both the untargeted and targeted settings. The results we present in this paper provide new insight into the nature of the non-robust features responsible for adversarial vulnerability of neural network classifiers.

Odest Chadwicke Jenkins, Daniel Lopresti, Melanie Mitchell

The history of AI has included several "waves" of ideas. The first wave, from the mid-1950s to the 1980s, focused on logic and symbolic hand-encoded representations of knowledge, the foundations of so-called "expert systems". The second wave, starting in the 1990s, focused on statistics and machine learning, in which, instead of hand-programming rules for behavior, programmers constructed "statistical learning algorithms" that could be trained on large datasets. In the most recent wave research in AI has largely focused on deep (i.e., many-layered) neural networks, which are loosely inspired by the brain and trained by "deep learning" methods. However, while deep neural networks have led to many successes and new capabilities in computer vision, speech recognition, language processing, game-playing, and robotics, their potential for broad application remains limited by several factors. A concerning limitation is that even the most successful of today's AI systems suffer from brittleness-they can fail in unexpected ways when faced with situations that differ sufficiently from ones they have been trained on. This lack of robustness also appears in the vulnerability of AI systems to adversarial attacks, in which an adversary can subtly manipulate data in a way to guarantee a specific wrong answer or action from an AI system. AI systems also can absorb biases-based on gender, race, or other factors-from their training data and further magnify these biases in their subsequent decision-making. Taken together, these various limitations have prevented AI systems such as automatic medical diagnosis or autonomous vehicles from being sufficiently trustworthy for wide deployment. The massive proliferation of AI across society will require radically new ideas to yield technology that will not sacrifice our productivity, our quality of life, or our values.

Erik Conser, Kennedy Hahn, Chandler M. Watson et al.

We revisit a particular visual grounding method: the "Image Retrieval Using Scene Graphs" (IRSG) system of Johnson et al. (2015). Our experiments indicate that the system does not effectively use its learned object-relationship models. We also look closely at the IRSG dataset, as well as the widely used Visual Relationship Dataset (VRD) that is adapted from it. We find that these datasets exhibit biases that allow methods that ignore relationships to perform relatively well. We also describe several other problems with the IRSG dataset, and report on experiments using a subset of the dataset in which the biases and other problems are removed. Our studies contribute to a more general effort: that of better understanding what machine learning methods that combine language and vision actually learn and what popular datasets actually test.

Max H. Quinn, Erik Conser, Jordan M. Witte et al.

We describe a novel architecture for semantic image retrieval---in particular, retrieval of instances of visual situations. Visual situations are concepts such as "a boxing match," "walking the dog," "a crowd waiting for a bus," or "a game of ping-pong," whose instantiations in images are linked more by their common spatial and semantic structure than by low-level visual similarity. Given a query situation description, our architecture---called Situate---learns models capturing the visual features of expected objects as well the expected spatial configuration of relationships among objects. Given a new image, Situate uses these models in an attempt to ground (i.e., to create a bounding box locating) each expected component of the situation in the image via an active search procedure. Situate uses the resulting grounding to compute a score indicating the degree to which the new image is judged to contain an instance of the situation. Such scores can be used to rank images in a collection as part of a retrieval system. In the preliminary study described here, we demonstrate the promise of this system by comparing Situate's performance with that of two baseline methods, as well as with a related semantic image-retrieval system based on "scene graphs."

Sheng Y. Lundquist, Melanie Mitchell, Garrett T. Kenyon

Deep Convolutional Neural Networks (DCNN) require millions of labeled training examples for image classification and object detection tasks, which restrict these models to domains where such datasets are available. In this paper, we explore the use of unsupervised sparse coding applied to stereo-video data to help alleviate the need for large amounts of labeled data. We show that replacing a typical supervised convolutional layer with an unsupervised sparse-coding layer within a DCNN allows for better performance on a car detection task when only a limited number of labeled training examples is available. Furthermore, the network that incorporates sparse coding allows for more consistent performance over varying initializations and ordering of training examples when compared to a fully supervised DCNN. Finally, we compare activations between the unsupervised sparse-coding layer and the supervised convolutional layer, and show that the sparse representation exhibits an encoding that is depth selective, whereas encodings from the convolutional layer do not exhibit such selectivity. These result indicates promise for using unsupervised sparse-coding approaches in real-world computer vision tasks in domains with limited labeled training data.

Anthony D. Rhodes, Jordan Witte, Melanie Mitchell et al.

We devise an algorithm using a Bayesian optimization framework in conjunction with contextual visual data for the efficient localization of objects in still images. Recent research has demonstrated substantial progress in object localization and related tasks for computer vision. However, many current state-of-the-art object localization procedures still suffer from inaccuracy and inefficiency, in addition to failing to provide a principled and interpretable system amenable to high-level vision tasks. We address these issues with the current research. Our method encompasses an active search procedure that uses contextual data to generate initial bounding-box proposals for a target object. We train a convolutional neural network to approximate an offset distance from the target object. Next, we use a Gaussian Process to model this offset response signal over the search space of the target. We then employ a Bayesian active search for accurate localization of the target. In experiments, we compare our approach to a state-of-theart bounding-box regression method for a challenging pedestrian localization task. Our method exhibits a substantial improvement over this baseline regression method.

Anthony D. Rhodes, Max H. Quinn, Melanie Mitchell

A major goal of computer vision is to enable computers to interpret visual situations---abstract concepts (e.g., "a person walking a dog," "a crowd waiting for a bus," "a picnic") whose image instantiations are linked more by their common spatial and semantic structure than by low-level visual similarity. In this paper, we propose a novel method for prior learning and active object localization for this kind of knowledge-driven search in static images. In our system, prior situation knowledge is captured by a set of flexible, kernel-based density estimations---a situation model---that represent the expected spatial structure of the given situation. These estimations are efficiently updated by information gained as the system searches for relevant objects, allowing the system to use context as it is discovered to narrow the search. More specifically, at any given time in a run on a test image, our system uses image features plus contextual information it has discovered to identify a small subset of training images---an importance cluster---that is deemed most similar to the given test image, given the context. This subset is used to generate an updated situation model in an on-line fashion, using an efficient multipole expansion technique. As a proof of concept, we apply our algorithm to a highly varied and challenging dataset consisting of instances of a "dog-walking" situation. Our results support the hypothesis that dynamically-rendered, context-based probability models can support efficient object localization in visual situations. Moreover, our approach is general enough to be applied to diverse machine learning paradigms requiring interpretable, probabilistic representations generated from partially observed data.

Max H. Quinn, Anthony D. Rhodes, Melanie Mitchell

We describe a method for performing active localization of objects in instances of visual situations. A visual situation is an abstract concept---e.g., "a boxing match", "a birthday party", "walking the dog", "waiting for a bus"---whose image instantiations are linked more by their common spatial and semantic structure than by low-level visual similarity. Our system combines given and learned knowledge of the structure of a particular situation, and adapts that knowledge to a new situation instance as it actively searches for objects. More specifically, the system learns a set of probability distributions describing spatial and other relationships among relevant objects. The system uses those distributions to iteratively sample object proposals on a test image, but also continually uses information from those object proposals to adaptively modify the distributions based on what the system has detected. We test our approach's ability to efficiently localize objects, using a situation-specific image dataset created by our group. We compare the results with several baselines and variations on our method, and demonstrate the strong benefit of using situation knowledge and active context-driven localization. Finally, we contrast our method with several other approaches that use context as well as active search for object localization in images.