Corpus

Advances in AI portend a new era of sophisticated disinformation operations. While individual AI systems already create convincing—and at times misleading—information, an imminent development is the emergence of malicious AI swarms. These systems can coordinate covertly, infiltrate communities, evade traditional detectors, and run continuous A/B tests, with round-the-clock persistence. The result can include fabricated grassroots consensus, fragmented shared reality, mass harassment, voter micro-suppression or mobilization, contamination of AI training data, and erosion of institutional trust. With increasing vulnerabilities in democratic processes worldwide, we urge a three-pronged response: (1) platform-side defenses—always-on swarm-detection dashboards, pre-election highfidelity swarm-simulation stress-tests, transparency audits, and optional client-side “AI shields” for users; (2) model-side safeguards—standardized persuasion-risk tests, provenance-authenticating passkeys, and watermarking; and (3) system-level oversight—a UN-backed AI Influence Observatory.

A core component of a successful artificial general intelligence would be the rapid creation and manipulation of grounded compositional abstractions and the demonstration of expertise in the family of recursive hierarchical syntactic objects necessary for the creative use of human language. We evaluated the recently released o3 model (OpenAI; o3-mini-high) and discovered that while it succeeds on some basic linguistic tests relying on linear, surface statistics (e.g., the Strawberry Test), it fails to generalize basic phrase structure rules; it fails with comparative sentences involving semantically illegal cardinality comparisons ('Escher sentences'); its fails to correctly rate and explain acceptability dynamics; and it fails to distinguish between instructions to generate unacceptable semantic vs. unacceptable syntactic outputs. When tasked with generating simple violations of grammatical rules, it is seemingly incapable of representing multiple parses to evaluate against various possible semantic interpretations. In stark contrast to many recent claims that artificial language models are on the verge of replacing the field of linguistics, our results suggest not only that deep learning is hitting a wall with respect to compositionality (Marcus 2022), but that it is hitting [a [stubbornly [resilient wall]]] that cannot readily be surmounted to reach human-like compositional reasoning simply through more compute.

Advances in AI offer the prospect of manipulating beliefs and behaviors on a population-wide level. Large language models and autonomous agents now let influence campaigns reach unprecedented scale and precision. Generative tools can expand propaganda output without sacrificing credibility and inexpensively create falsehoods that are rated as more human-like than those written by humans. Techniques meant to refine AI reasoning, such as chain-of-thought prompting, can just as effectively be used to generate more convincing falsehoods. Enabled by these capabilities, a disruptive threat is emerging: swarms of collaborative, malicious AI agents. Fusing LLM reasoning with multi-agent architectures, these systems are capable of coordinating autonomously, infiltrating communities, and fabricating consensus efficiently. By adaptively mimicking human social dynamics, they threaten democracy. Because the resulting harms stem from design, commercial incentives, and governance, we prioritize interventions at multiple leverage points, focusing on pragmatic mechanisms over voluntary compliance.

The lack of a concrete definition for Artificial General Intelligence (AGI) obscures the gap between today's specialized AI and human-level cognition. This paper introduces a quantifiable framework to address this, defining AGI as matching the cognitive versatility and proficiency of a well-educated adult. To operationalize this, we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition. The framework dissects general intelligence into ten core cognitive domains-including reasoning, memory, and perception-and adapts established human psychometric batteries to evaluate AI systems. Application of this framework reveals a highly "jagged" cognitive profile in contemporary models. While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage. The resulting AGI scores (e.g., GPT-4 at 27%, GPT-5 at 57%) concretely quantify both rapid progress and the substantial gap remaining before AGI.

A core component of a successful artificial general intelligence would be the rapid creation and manipulation of grounded compositional abstractions and the demonstration of expertise in the family of recursive hierarchical syntactic objects necessary for the creative use of human language. We evaluated the recently released o3 model (OpenAI; o3-mini-high) from ChatGPT and discovered that while it succeeds on some basic linguistic tests relying on linear, surface statistics (e.g., the Strawberry Test), it fails to generalize basic phrase structure rules; it fails with comparative sentences involving semantically illegal cardinality comparisons (‘Escher sentences’); it fails to correctly rate and explain acceptability dynamics; and it fails to distinguish between instructions to generate unacceptable semantic vs. unacceptable syntactic outputs. When tasked with generating simple violations of grammatical rules, it is seemingly incapable of representing multiple parses to evaluate against various possible semantic interpretations. We ran all of these prompts multiple times again through the API and provide basic accuracy scores. In stark contrast to many recent claims that artificial language models are on the verge of replacing the field of linguistics, our results suggest not only that deep learning is hitting a wall with respect to compositionality ( Marcus, 2022 ), but that it is hitting [ a [ stubbornly [ resilient wall ]]] that cannot readily be surmounted to reach human-like compositional reasoning simply through more compute.

Large Language Models (LLMs) are recruited in applications that span from clinical assistance and legal support to question answering and education. Their success in specialized tasks has led to the claim that they possess human-like linguistic capabilities related to compositional understanding and reasoning. Yet, reverse-engineering is bound by Moravec's Paradox, according to which easy skills are hard. We systematically assess 7 state-of-the-art models on a novel benchmark. Models answered a series of comprehension questions, each prompted multiple times in two settings, permitting one-word or open-length replies. Each question targets a short text featuring high-frequency linguistic constructions. To establish a baseline for achieving human-like performance, we tested 400 humans on the same prompts. Based on a dataset of n = 26,680 datapoints, we discovered that LLMs perform at chance accuracy and waver considerably in their answers. Quantitatively, the tested models are outperformed by humans, and qualitatively their answers showcase distinctly non-human errors in language understanding. We interpret this evidence as suggesting that, despite their usefulness in various tasks, current AI models fall short of understanding language in a way that matches humans, and we argue that this may be due to their lack of a compositional operator for regulating grammatical and semantic information.

The Winograd Schema Challenge—a set of twin sentences involving pronoun reference disambiguation that seem to require the use of commonsense knowledge—was proposed by Hector Levesque in 2011. By 2019, a number of AI systems, based on large pre-trained transformer-based language models and fine-tuned on these kinds of problems, achieved better than 90% accuracy. In this paper, we review the history of the Winograd Schema Challenge and discuss the lasting contributions of the flurry of research that has taken place on the WSC in the last decade. We discuss the significance of various datasets developed for WSC, and the research community's deeper understanding of the role of surrogate tasks in assessing the intelligence of an AI system.

Machine intelligence is increasingly being linked to claims about sentience, language processing, and an ability to comprehend and transform natural language into a range of stimuli. We analyze the ability of DALL·E 2 to translate natural language prompts into images, since it remains unclear whether it possesses the facility for accurately representing grammatical strategies. We target the performance of DALL·E 2 across 10 phenomena that are pervasive in human language: binding principles and coreference, passives, structural ambiguity, negation, compositionality and word order, quantification, double object constructions, sentence coordination, ellipsis, and comparatives. In contrast to young infants, who master these tasks, DALL·E 2 fails to accurately represent inferred meanings, performing at or near chance. While programs can be trained to recognize vast numbers of words and calculate probabilities of word sequences, these results challenge recent claims concerning artificial understanding of human language. The full set of tested materials and the outputs are made available as a benchmark for future testing.

The Communications website, https://cacm.acm.org, features more than a dozen bloggers in the BLOG@CACM community. In each issue of Communications , we'll publish selected posts or excerpts. twitter Follow us on Twitter at http://twitter.com/blogCACM https://cacm.acm.org/blogs/blog-cacm Gary Marcus on the systems that "pose a real and imminent threat to the fabric of society."

The Communications website, https://cacm.acm.org, features more than a dozen bloggers in the BLOG@CACM community. In each issue of Communications , we'll publish selected posts or excerpts. twitter Follow us on Twitter at http://twitter.com/blogCACM https://cacm.acm.org/blogs/blog-cacm Gary Marcus considers it appropriate that governments are stepping up on artificial intelligence, but finds some of the resulting signals deeply worrisome.

Large Language Models (LLMs) are recruited in applications that span from clinical assistance and legal support to question answering and education. Their success in specialized tasks has led to the claim that they possess human-like linguistic capabilities related to compositional understanding and reasoning. Yet, reverse-engineering is bound by Moravec's Paradox, according to which easy skills are hard. We systematically assess 7 state-of-the-art models on a novel benchmark. Models answered a series of comprehension questions, each prompted multiple times in two settings, permitting one-word or open-length replies. Each question targets a short text featuring high-frequency linguistic constructions. To establish a baseline for achieving human-like performance, we tested 400 humans on the same prompts. Based on a dataset of n=26,680 datapoints, we discovered that LLMs perform at chance accuracy and waver considerably in their answers. Quantitatively, the tested models are outperformed by humans, and qualitatively their answers showcase distinctly non-human errors in language understanding. We interpret this evidence as suggesting that, despite their usefulness in various tasks, current AI models fall short of understanding language in a way that matches humans, and we argue that this may be due to their lack of a compositional operator for regulating grammatical and semantic information.

Modern Artificial Intelligence applications show great potential for language-related tasks that rely on next-word prediction. The current generation of Large Language Models (LLMs) have been linked to claims about human-like linguistic performance and their applications are hailed both as a step towards artificial general intelligence and as a major advance in understanding the cognitive, and even neural basis of human language. To assess these claims, first we analyze the contribution of LLMs as theoretically informative representations of a target cognitive system vs. atheoretical mechanistic tools. Second, we evaluate the models' ability to see the bigger picture, through top-down feedback from higher levels of processing, which requires grounding in previous expectations and past world experience. We hypothesize that since models lack grounded cognition, they cannot take advantage of these features and instead solely rely on fixed associations between represented words and word vectors. To assess this, we designed and ran a novel 'leet task' (l33t t4sk), which requires decoding sentences in which letters are systematically replaced by numbers. The results suggest that humans excel in this task whereas models struggle, confirming our hypothesis. We interpret the results by identifying the key abilities that are still missing from the current state of development of these models, which require solutions that go beyond increased system scaling.

Generative AI, the most popular current approach to AI, consists of large language models (LLMs) that are trained to produce outputs that are plausible, but not necessarily correct. Although their abilities are often uncanny, they are lacking in aspects of reasoning, leading LLMs to be less than completely trustworthy. Furthermore, their results tend to be both unpredictable and uninterpretable. We lay out 16 desiderata for future AI, and discuss an alternative approach to AI which could theoretically address many of the limitations associated with current approaches: AI educated with curated pieces of explicit knowledge and rules of thumb, enabling an inference engine to automatically deduce the logical entailments of all that knowledge. Even long arguments produced this way can be both trustworthy and interpretable, since the full step-by-step line of reasoning is always available, and for each step the provenance of the knowledge used can be documented and audited. There is however a catch: if the logical language is expressive enough to fully represent the meaning of anything we can say in English, then the inference engine runs much too slowly. That's why symbolic AI systems typically settle for some fast but much less expressive logic, such as knowledge graphs. We describe how one AI system, Cyc, has developed ways to overcome that tradeoff and is able to reason in higher order logic in real time. We suggest that any trustworthy general AI will need to hybridize the approaches, the LLM approach and more formal approach, and lay out a path to realizing that dream.

The first half of this tutorial will make deep nets more accessible to a broader audience, following "Deep Nets for Poets" and "A Gentle Introduction to Fine-Tuning." We will also introduce, gft (general fine tuning), a little language for fine tuning deep nets with short (one line) programs that are as easy to code as regression in statistics packages such as R using glm (general linear models). Based on the success of these methods on a number of benchmarks, one might come away with the impression that deep nets are all we need. However, we believe the glass is half-full: while there is much that can be done with deep nets, there is always more to do. The second half of this tutorial will discuss some of these opportunities.

Between 1965 and 1982, classically symbolic AI was often touted as the most promising basis for a general theory of Cognitive Architecture. By 1986, however, prominent researchers began to herald the emergence of a new paradigm, Connectionism, which would (they argued) eventually displace the classically symbolic methods dominant in AI and Cognitive Science. Since then, a variety of competing positions have emerged, some purely connectionist, some hybrid, and some purely classical. This symposium examines the present and future prospects for contending paradigms in this realm. A variety of arguments, arising from computational, psychological, biological and philosophical bases, are presented.

The Winograd Schema Challenge - a set of twin sentences involving pronoun reference disambiguation that seem to require the use of commonsense knowledge - was proposed by Hector Levesque in 2011. By 2019, a number of AI systems, based on large pre-trained transformer-based language models and fine-tuned on these kinds of problems, achieved better than 90% accuracy. In this paper, we review the history of the Winograd Schema Challenge and discuss the lasting contributions of the flurry of research that has taken place on the WSC in the last decade. We discuss the significance of various datasets developed for WSC, and the research community's deeper understanding of the role of surrogate tasks in assessing the intelligence of an AI system.

The DALL-E 2 system generates original synthetic images corresponding to an input text as caption. We report here on the outcome of fourteen tests of this system designed to assess its common sense, reasoning and ability to understand complex texts. All of our prompts were intentionally much more challenging than the typical ones that have been showcased in recent weeks. Nevertheless, for 5 out of the 14 prompts, at least one of the ten images fully satisfied our requests. On the other hand, on no prompt did all of the ten images satisfy our requests.

Machine intelligence is increasingly being linked to claims about sentience, language processing, and an ability to comprehend and transform natural language into a range of stimuli. We systematically analyze the ability of DALL-E 2 to capture 8 grammatical phenomena pertaining to compositionality that are widely discussed in linguistics and pervasive in human language: binding principles and coreference, passives, word order, coordination, comparatives, negation, ellipsis, and structural ambiguity. Whereas young children routinely master these phenomena, learning systematic mappings between syntax and semantics, DALL-E 2 is unable to reliably infer meanings that are consistent with the syntax. These results challenge recent claims concerning the capacity of such systems to understand of human language. We make available the full set of test materials as a benchmark for future testing.

In line with this year’s Academy’s theme, we propose a panel symposium to discuss how artificial intelligence and human-machine interaction will challenge managers and the organizations and teams they run. In the context of artificial intelligence, we take to heart Herbert Simon’s observation that “(…) a highly automated man-machine system, and the nature of management will surely be conditioned by the character of the system being managed” (Simon, 1960). Will managers become obsolete, because of the unmatched predictive capabilities of AI, and what space will human judgment have in a world of superior machines? Can managers learn to play a role in this transformation? Our panel consists of academics from diverse disciplines (economics, psychology, information systems, strategic management) to discuss, and debate, the questions surrounding AI in the context of human-machine interactions within organizations. The symposium is built upon the participants’ empirical research, but will not be formal paper presentations. The panel discussion seeks to clarify the many great debates over AI and machine learning with the goal of sharpening the agenda of the re-emergent topicality of human-machine interactions. The debate will then be open to broader discussion including the audience.

The project of justifying all the limits and failings of human cognition as inevitable consequences of strategies that are actually "optimal" relative to the limits on computational resources available may have some value, but it is far from a complete explanation. It is inconsistent with both common observation and a large body of experimentation, and it is of limited use in explaining human cognition.

How the cognitive sciences can inform the quest to build systems with the flexibility of the human mind.

Most discussions of connectionism in cognitive science have focused on whether or not simple unstructured networks such as multilayer perceptrons suffice for cognition. Other types of connectionist models have received relatively less attention. The purpose of this symposium is to increase awareness of these other kinds of models, situating them in a context that allows a greater understanding of what kinds of models are good for what kinds of tasks. Our assumption is that in any given domain it is possible build some adequate model, but that the interesting question is how. The emphasis, then, will be on the strengths and weaknesses of competing connectionist architectures and how they can be usefully applied to models of cognition.

Recent research in artificial intelligence and machine learning has largely emphasized general-purpose learning and ever-larger training sets and more and more compute. In contrast, I propose a hybrid, knowledge-driven, reasoning-based approach, centered around cognitive models, that could provide the substrate for a richer, more robust AI than is currently possible.

The Winograd Schema Challenge is both a commonsense reasoning and natural language understanding challenge, introduced as an alternative to the Turing test. A Winograd schema is a pair of sentences differing in one or two words with a highly ambiguous pronoun, resolved differently in the two sentences, that appears to require commonsense knowledge to be resolved correctly. The examples were designed to be easily solvable by humans but difficult for machines, in principle requiring a deep understanding of the content of the text and the situation it describes. This paper reviews existing Winograd Schema Challenge benchmark datasets and approaches that have been published since its introduction.

This symposium proposes to discuss and debate the most recent research on the implications of artificial intelligence for organizations and society. The debates on artificial intelligence and its impact on the organization of work touch upon some of the most profound questions of our times and are central to the strategic choices of firms responding to technical change. These debates include: will artificial intelligence replace human labor by code, will decisions vital to firms - such as the recruiting and selection of the best personnel - be assigned to predictive algorithms or human choice, and will AI constitute a type of general purpose technology that benefits adopters and possibly leads the economy to be dominated by relatively few organizations, as currently in the high technology sector. The discussion in the symposium is built upon the presentation of empirical research. The debate will then be open to broader discussion including the audience.

Pater’s (2019) expansive review is a significant contribution toward bridging the disconnect of generative linguistics with connectionism, and as such, it is an important service to the field. But Pater’s efforts for inclusion and reconciliation obscure crucial substantive disagreements on foundational matters. Most connectionist models are antithetical to the algebraic hypothesis that has guided generative linguistics from its inception. They eschew the notions that mental representations have formal constituent structure and that mental operations are structure-sensitive. These representational commitments critically limit the scope of learning and productivity in connectionist models. Moving forward, we see only two options: either those connectionist models are right, and generative linguistics must be radically revised, or they must be replaced by alternatives that are compatible with the algebraic hypothesis. There can be no integration without structured representations.

Pater's (2019) expansive review is a significant contribution toward bridging the disconnect of generative linguistics with connectionism, and as such, it is an important service to the field. But Pater's efforts for inclusion and reconciliation obscure crucial substantive disagreements on foundational matters. Most connectionist models are antithetical to the algebraic hypothesis that has guided generative linguistics from its inception. They eschew the notions that mental representations have formal constituent structure and that mental operations are structure-sensitive. These representational commitments critically limit the scope of learning and productivity in connectionist models. Moving forward, we see only two options: either those connectionist models are right, and generative linguistics must be radically revised, or they must be replaced by alternatives that are compatible with the algebraic hypothesis. There can be no integration without structured representations.

Although deep learning has historical roots going back decades, neither the term "deep learning" nor the approach was popular just over five years ago, when the field was reignited by papers such as Krizhevsky, Sutskever and Hinton's now classic (2012) deep network model of Imagenet. What has the field discovered in the five subsequent years? Against a background of considerable progress in areas such as speech recognition, image recognition, and game playing, and considerable enthusiasm in the popular press, I present ten concerns for deep learning, and suggest that deep learning must be supplemented by other techniques if we are to reach artificial general intelligence.

The concept of innateness is rarely discussed in the context of artificial intelligence. When it is discussed, or hinted at, it is often the context of trying to reduce the amount of innate machinery in a given system. In this paper, I consider as a test case a recent series of papers by Silver et al (Silver et al., 2017a) on AlphaGo and its successors that have been presented as an argument that a "even in the most challenging of domains: it is possible to train to superhuman level, without human examples or guidance", "starting tabula rasa." I argue that these claims are overstated, for multiple reasons. I close by arguing that artificial intelligence needs greater attention to innateness, and I point to some proposals about what that innateness might look like.

Human reasoning is richer than Lake et al. acknowledge, and the emphasis on theories of how images and scenes are synthesized is misleading. For example, the world knowledge used in vision presumably involves a combination of geometric, physical, and other knowledge, rather than just a causal theory of how the image was produced. In physical reasoning, a model can be a set of constraints rather than a physics engine. In intuitive psychology, many inferences proceed without detailed causal generative models. How humans reliably perform such inferences, often in the face of radically incomplete information, remains a mystery.

While acquiring motor skills, animals transform their plastic motor sequences to match desired targets. However, because both the structure and temporal position of individual gestures are adjustable, the number of possible motor transformations increases exponentially with sequence length. Identifying the optimal transformation towards a given target is therefore a computationally intractable problem. Here we show an evolutionary workaround for reducing the computational complexity of song learning in zebra finches. We prompt juveniles to modify syllable phonology and sequence in a learned song to match a newly introduced target song. Surprisingly, juveniles match each syllable to the most spectrally similar sound in the target, regardless of its temporal position, resulting in unnecessary sequence errors, that they later try to correct. Thus, zebra finches prioritize efficient learning of syllable vocabulary, at the cost of inefficient syntax learning. This strategy provides a non-optimal but computationally manageable solution to the task of vocal sequence learning.

We examined mechanisms underlying infants’ ability to\ndetect, extract, and generalize sequential patterns, focusing on\nhow saliency and consistency of distributional information\nguide infant learning of the most “likely” pattern in\naudiovisual sequences. In Experiment 1, we asked if 11- and\n14-month-old infants could learn a “repetition anywhere” rule\n(e.g., ABBC, AABC, ABCC). In Experiment 2 we asked if\n11- and 14-month-olds could generalize a “medial repetition”\nrule when its position is consistent in sequence, and in\nExperiment 3 we asked if 11-month-olds could identify a\nnonadjacent dependency occurring at edge positions. Infants\nwere first habituated to 4-item sequences (shapes + syllables)\ncontaining repetition- and/or position-based structure, and\nwere then tested with “familiar” structure instantiated across\nnew items or combinations of items vs. “novel” (random)\nsequences. We found that 11-month-olds failed to learn the\nrepetition rule both when the structure appeared in initial,\nmedial, or final position (Experiment 1) and when it was\nrestricted to the medial position (Experiment 2). Fourteen-\nmonth-olds learned repetition rules under both conditions.\nFinally, in Experiment 3 11-month-olds succeeded in learning\na nonadjacent dependency in sequences identical to those\nused to test repetition learning in Experiment 2. Our results\nsuggest that infants at 11 months, like adults, are relatively\ninsensitive to patterns in the middle of sequences.

Human readers comprehend vastly more, and in vastly different ways, than any existing comprehension test would suggest. An ideal comprehension test for a story should cover the full range of questions and answers that humans would expect other humans to reasonably learn or infer from a given story. As a step toward these goals we propose a novel test, the crowdsourced comprehension challenge (C 3 ), which is constructed by repeated runs of a three‐person game, the Iterative Crowdsourced Comprehension Game (ICCG). ICCG uses structured crowdsourcing to comprehensively generate relevant questions and supported answers for arbitrary stories, whether fiction or nonfiction, presented across a variety of media such as videos, podcasts, and still images.

The articles in this special issue of AI Magazine include those that propose specific tests and those that look at the challenges inherent in building robust, valid, and reliable tests for advancing the state of the art in AI .

AI has seen great advances of many kinds recently, but there is one critical area where progress has been extremely slow: ordinary commonsense.

The authors reply to a comment on their article How robust are probabilistic models of higher-level cognition?, by G. F. Marcus, et al. The authors debated that multiple, equally probable, Bayesian models could be made for the tasks under consideration, and that Bayesian theories do not restrict which model applies in any given case.

List of Contributors ix Preface Gary Marcus and Jeremy Freeman xi MAPPING THE BRAIN Building Atlases of the Brain 3 Mike Hawrylycz with Chinh Dang, Christof Koch, and Hongkui Zeng Whole Brain Neuroimaging and Virtual Reality 17 Misha B. Ahrens Project MindScope 25 Christof Koch with Clay Reid, Hongkui Zeng, Stefan Mihalas, Mike Hawrylycz, John Philips, Chinh Dang, and Allan Jones The Connectome as a DNA Sequencing Problem 40 Anthony Zador Rosetta Brain 50 George Church with Adam Marblestone and Reza Kalhor COMPUTATION Understanding the Cortex through Grid Cells 67 May-Britt Moser and Edvard I. Moser Recording from Many Neurons Simultaneously: From Measurement to Meaning 78 Krishna V. Shenoy Network Neuroscience 90 Olaf Sporns Large-Scale Neuroscience: From Analytics to Insight 100 Jeremy Freeman SIMULATING THE BRAIN Whole Brain Simulation 111 Sean Hill Building a Behaving Brain 125 Chris Eliasmith LANGUAGE The Neurobiology of Language 139 David Poeppel Translating the Genome in Human Neuroscience 149 Simon E. Fisher Color plates follow p. 160 SKEPTICS Consciousness, Big Science, and Conceptual Clarity 161 Ned Block From Circuits to Behavior: A Bridge Too Far? 177 Matteo Carandini Lessons from Evolution 186 Leah Krubitzer Lessons from the Genome 194 Arthur Caplan with Nathan Kunzler The Computational Brain 205 Gary Marcus IMPLICATIONS Neurotechnology 219 John Donoghue The Miswired Brain, Genes, and Mental Illness 234 Kevin J. Mitchell Neural Dust: An Untethered Approach to Chronic Brain-Machine Interfaces 243 Michel M. Maharbiz with Dongjin Seo, Jose M. Carmena, Jan M. Rabaey, and Elad Alon AFTERWORD Neuroscience in 2064: A Look at the Last Century 255 Christof Koch and Gary Marcus Glossary 271 Index 275

AAAI's 2015 Workshop Program was held Sunday and Monday, January 25–26, 2015, at the Hyatt Regency Austin Hotel in Austin, Texas, USA. The AAAI‐15 workshop program included 16 workshops covering a wide range of topics in artificial intelligence. Most workshops were held on a single day. The titles of the workshops included Algorithm Configuration; Artificial Intelligence and Ethics; Artificial Intelligence Applied to Assistive Technologies and Smart Environments; Artificial Intelligence for Cities; Artificial Intelligence for Transportation: Advice, Interactivity, and Actor Modeling; Beyond the Turing Test; Computational Sustainability; Computer Poker and Imperfect Information; Incentive and Trust in E‐Communities; Knowledge, Skill, and Behavior Transfer in Autonomous Robots; Learning for General Competency in Video Games; Multiagent Interaction without Prior Coordination; Planning, Search, and Optimization; Scholarly Big Data: AI Perspectives, Challenges, and Ideas; Trajectory‐Based Behaviour Analytics; and World Wide Web and Public Health Intelligence.

Music, like speech, is a complex auditory signal that contains structures at multiple timescales, and as such is a potentially powerful entry point into the question of how the brain integrates complex streams of information. Using an experimental design modeled after previous studies that used scrambled versions of a spoken story (Lerner et al., 2011) and a silent movie (Hasson et al., 2008), we investigate whether listeners perceive hierarchical structure in music beyond short (~6 s) time windows and whether there is cortical overlap between music and language processing at multiple timescales. Experienced pianists were presented with an extended musical excerpt scrambled at multiple timescales-by measure, phrase, and section-while measuring brain activity with functional magnetic resonance imaging (fMRI). The reliability of evoked activity, as quantified by inter-subject correlation of the fMRI responses, was measured. We found that response reliability depended systematically on musical structure coherence, revealing a topographically organized hierarchy of processing timescales. Early auditory areas (at the bottom of the hierarchy) responded reliably in all conditions. For brain areas at the top of the hierarchy, the original (unscrambled) excerpt evoked more reliable responses than any of the scrambled excerpts, indicating that these brain areas process long-timescale musical structures, on the order of minutes. The topography of processing timescales was analogous with that reported previously for speech, but the timescale gradients for music and speech overlapped with one another only partially, suggesting that temporally analogous structures-words/measures, sentences/musical phrases, paragraph/sections-are processed separately.

It has been proposed that human physical reasoning consists largely of running "physics engines in the head" in which the future trajectory of the physical system under consideration is computed precisely using accurate scientific theories. In such models, uncertainty and incomplete knowledge is dealt with by sampling probabilistically over the space of possible trajectories ("Monte Carlo simulation"). We argue that such simulation-based models are too weak, in that there are many important aspects of human physical reasoning that cannot be carried out this way, or can only be carried out very inefficiently; and too strong, in that humans make large systematic errors that the models cannot account for. We conclude that simulation-based reasoning makes up at most a small part of a larger system that encompasses a wide range of additional cognitive processes.

Based on a survey of the literature, we attempt to answer Frequently Asked Questions on issues of cortical uniformity vs. non-uniformity, the neural mechanisms of symbolic variable binding, and other issues highlighted in (Marcus, Marblestone and Dean. "The Atoms of Neural Computation". Science. 31 October 2014. Vol 346. Issue 6209).

Does the brain depend on a set of elementary, reusable computations?

One of the most fascinating ideas ever to emerge from cognitive science was the notion that the mind, unlike digital computers, might proceed entirely without recourse to symbol-manipulation. But how tenable is that intriguing idea, 25 years later? Fodor and Pylyshyn challenged it immediately; I argue that so-called “eliminative connectionism” never made as much progress as its authors might hope. Instead, there continues to be good reason to believe that minds have, among other capacities, a neurally realized way of representing symbols, variables, and operations over variables, and have the resources to distinguish types from tokens and to represent ordered pairs and structured units. After a quarter century, advocates of eliminative connectionism have yet to mount an adequate alternative. A more profitable endeavor might be to figure out how to use networks of neurons in systems that unify symbols and statistics, rather than needlessly treating them as antithetical.

In statistical modeling, there are numerous parameters defined mathematically to account for the expected shape of data distributions. In the domain of human language, both principles and parameters are often thought of as innate domain-specific abstractions that connect to many structural properties about language. Linguistic principles correspond to the properties that are invariant across all human languages. The learning path turns out to be crucial for learning the English metrical phonology systems using parameters. An increasingly common way to explore how exactly the human mind learns with parameters is to use computational modeling. Computational modeling gives us very precise control over the language acquisition process, including what hypotheses the child entertains, what data are considered relevant, and how the child changes belief in different hypotheses based on the data. The chapter also looks at a few examples of computational models that investigate language acquisition using parameters.

Biolinguistics involves the study of language from a broad perspective that embraces natural sciences, helping us better to understand the fundamentals of the faculty of language. This Handbook offers the most comprehensive state-of-the-field survey of the subject available. A team of prominent scholars working in a variety of disciplines is brought together to examine language development, language evolution and neuroscience, as well as providing overviews of the conceptual landscape of the field. The Handbook includes work at the forefront of contemporary research devoted to the evidence for a language instinct, the critical period hypothesis, grammatical maturation, bilingualism, the relation between mind and brain and the role of natural selection in language evolution. It will be welcomed by graduate students and researchers in a wide range of disciplines, including linguistics, evolutionary biology and cognitive science.

A central process in music cognition involves the identification of key; however, little is known about how listeners accomplish this task in real time. This study derives from work that suggests overlap between the neural and cognitive resources underlying the analyses of both music and speech and is the first, to our knowledge, to explore the timescales at which the brain infers musical key. We investigated the temporal psychophysics of key-finding over a wide range of tempi using melodic sequences with strong structural cues, where statistical information about overall key profile was ambiguous. Listeners were able to provide robust judgments within specific limits, at rates as high as 400 beats per minute (bpm; ∼7 Hz) and as low as 30 bpm (0.5 Hz), but not outside those bounds. These boundaries on reliable performance show that the process of key-finding is restricted to timescales that are closely aligned with beat induction and speech processing.

An increasingly popular theory holds that the mind should be viewed as a near-optimal or rational engine of probabilistic inference, in domains as diverse as word learning, pragmatics, naive physics, and predictions of the future. We argue that this view, often identified with Bayesian models of inference, is markedly less promising than widely believed, and is undermined by post hoc practices that merit wholesale reevaluation. We also show that the common equation between probabilistic and rational or optimal is not justified.

Natural language expressions are communicated in sequences of articulations, either sounds or gestures. Where sequence is known to matter, language clearly demands some sort of sequence-sensitive structure. The dog bit the man is not equivalent to The man bit the dog, and any system that represented sentences as unordered lists of words simply could not capture that fact. A standard (though not universal) assumption in most theories of linguistics and psycholinguistics is that phrase structure is represented as a tree (Partee, ter Meulen, & Wall, 1990), as in Figure 2.1. Trees are widespread, and their formal properties are well understood (Knuth, 1973). They play a central role not only in linguistics, but also in disciplines such as mathematics, computer science, taxonomy, and even genealogy. For example, computer programmers use trees to represent the nested directory (or folder) structures in which files are stored. In linguistics, tree-theoretic concepts, such as path and distance, often seem to provide important purchase on syntactic phenomena (Kayne, 1984; Pesetsky, 1982). It is an open question, however, whether human beings can actually use trees as a data structure for representing sentences in a stable and reliable fashion. Although it is clear that educated humans can understand at an abstract level what trees are, implement them in computers, and use them as external representational devices, the evidence for their existence qua mental representation has been remarkably indirect, and there has not thus far been any fully satisfying account for how trees as such could be neurally realized. The goal of this chapter is to suggest an alternative: in the course of ordinary linguistic processing, mental representations of sentences may be incomplete, inconsistent, or partial—realized via a system of overlapping but incompletely

In physical reasoning, humans are often able to carry out useful reasoning based on radically incomplete information. One physical domain that it is ubiquitous both in everyday interactions and in many kinds of scientific applications, where reasoning from incomplete information is very common, is the interaction of containers and their contents. We have developed a preliminary knowledge base for qualitative reasoning about containers, expressed in a sorted first-order language of time, geometry, objects, histories, and events. We have demonstrated that the knowledge suffices to justify a number of commonsense physical inferences, based on very incomplete knowledge. 1. Physical Reasoning Based on Radically Incomplete Information In physical reasoning, humans, unlike programs for scientific computation, are often able to carry out useful reasoning based on radically incomplete information. If AI systems are to achieve human levels of reasoning, they must likewise have this ability. Extant automated reasoners based on simulation cannot fully address the challenges of radically incomplete information (Davis & Marcus, 2013); rather such challenges require alternative reasoning techniques specifically designed for incomplete information.

Language is rife with ambiguity. Do children and adults meet this challenge in similar ways? Recent work suggests that while adults resolve syntactic ambiguities by integrating a variety of cues, children are less sensitive to top-down evidence. We test whether this top-down insensitivity is specific to syntax or a general feature of children's linguistic ambiguity resolution by evaluating whether children rely largely or completely on lexical associations to resolve lexical ambiguities (e.g., the word swing primes the baseball meaning of bat) or additionally integrate top-down global plausibility. Using a picture choice task, we compared 4-year-olds' ability to resolve polysemes and homophones with a Bayesian algorithm reliant purely on lexical associations and found that the algorithm's power to predict children's choices was limited. A 2nd experiment confirmed that children override associations and integrate top-down plausibility. We discuss this with regard to models of psycholinguistic development.

Is the human tendency toward musicality better thought of as the product of a specific, evolved instinct or an acquired skill? Developmental and evolutionary arguments are considered, along with issues of domain-specificity. The article also considers the question of why humans might be consistently and intensely drawn to music if musicality is not in fact the product of a specifically evolved instinct.

A recent computational model by Hayes and Wilson (2008) seemingly captures a diverse range of phonotactic phenomena without variables, contrasting with the presumptions of many formal theories. Here, we examine the plausibility of this approach by comparing generalizations of identity restrictions by this architecture and human learners. Whereas humans generalize identity restrictions broadly, to both native and non-native phonemes, the original model and several related variants failed to generalize to non-native phonemes. In contrast, a revised model equipped with variables more closely matches human behavior. These findings suggest that, like syntax, phonological grammars are endowed with algebraic relations among variables that support across-the-board generalizations.

Infants appear to learn abstract rule-like regularities (e.g., la la da follows an AAB pattern) more easily from speech than from a variety of other auditory and visual stimuli (Marcus et al., 2007). We test if that facilitation reflects a specialization to learn from speech alone, or from modality-independent communicative stimuli more generally, by measuring 7.5-month-old infants' ability to learn abstract rules from sign language-like gestures. Whereas infants appear to easily learn many different rules from speech, we found that with sign-like stimuli, and under circumstances comparable to those of Marcus et al. (1999), hearing infants were able to learn an ABB rule, but not an AAB rule. This is consistent with results of studies that demonstrate lower levels of infant rule learning from a variety of other non-speech stimuli, and we discuss implications for accounts of speech-facilitation.

Word-learning likely involves a multiplicity of components, some domain-general, others domain-specific. Against the background of recent studies that suggest that word-learning is domain-specific, we investigated the associative component of word-learning. Seven- and 14-month-old infants viewed a pair of events in which a monkey or a truck moved back and forth, accompanied by a sung syllable or a tone, matched for pitch. Following habituation, infants were presented with displays in which the visual-auditory pairings were preserved or switched, and looked longer at the "switch" events when exposure time was sufficient to learn the intermodal association. At 7 months, performance on speech and tones conditions was statistically identical; at 14 months, infants had begun to favor speech. Thus, the associative component of word-learning does not appear (in contrast to rule-learning, Marcus et al., 2007) to initially privilege speech.

On the eve of his fortieth birthday, scientist Gary Marcus decided to fulfil a lifelong dream and learn to play the guitar, investigating how humans 'make' music and how anyone of any age might master a new skill. In a quest that takes him from Suzuki classes to guitar gods, Marcus discovers the best ways to train your brain and your body. He also brings insight into the question, What counts as a life well lived? A groundbreaking look at the origins and allure of music, Marcus's journey is also an empowering tale of the mind's plasticity and its ability to grow throughout life.

Can the study of individual differences inform debates about modularity and the specialization of function? In this article, we consider the implications of a highly replicated, robust finding known as positive manifold: Individual differences in different cognitive domains tend to be positively intercorrelated. Prima facie, this fact, which has generally been interpreted as reflecting the influence of a domain-general cognitive factor, might be seen as posing a serious challenge to a strong view of modularity. Drawing on a mixture of meta-analysis and computer simulation, we show that positive manifold derives instead largely from between-task neural overlap, suggesting a potential way of reconciling individual differences with some form of modularity.

Abstract We find the theory of neural reuse to be highly plausible, and suggest that human individual differences provide an additional line of argument in its favor, focusing on the well-replicated finding of “positive manifold,” in which individual differences are highly correlated across domains. We also suggest that the theory of neural reuse may be an important contributor to the phenomenon of positive manifold itself.

While substantial evidence suggests that key perception depends in part on the statistical distribution of pitch classes in a given piece, it has also been suggested that structural factors may play an important role. To further the understanding of what such structural factors might be, we conducted two experiments using statistically ambiguous yet structurally unambiguous stimuli that consisted of a uniform distribution of diatonic pitches drawn from the union of two neighboring keys, sequenced to clearly imply one particular key. In Experiment 1, subjects with substantial music theory training were asked to listen to melodic sequences and identify the key. The results indicate that purely structural factors can greatly influence key perception when statistical information about overall key profile is highly ambiguous. In Experiment 2, we examined the temporal psychophysics of structurally-based key induction by systematically varying the tempo of a subset of the statistically ambiguous/structurally unambiguous materials presented in Experiment 1 at rates ranging from 7 BPM to 3400 BPM. Twenty-two musically trained subjects were asked to indicate whether each sequence sounded resolved (ending on an implied tonic) or unresolved (ending on an implied dominant). The results agree strongly with those from Experiment 1 and show a preferred range of tempi in which participants provide robust key judgments (30 BPM 400 BPM). This suggests that structurally-based key-finding can be accurately computed by trained listeners in as little as 150ms per note or 1.2s for an entire sequence of eight notes.

Abstract The mere fact that a particular aspect of mind could offer an adaptive advantage is not enough to show that that property was in fact shaped by that adaptive advantage. Although it is possible that the tendency towards positive illusion is an evolved misbelief, it it also possible that positive illusions could be a by-product of a broader, flawed cognitive mechanism that itself was shaped by accidents of evolutionary inertia.

The experiments reported here investigated the development of a fundamental component of cognition: to recognize and generalize abstract relations. Infants were presented with simple rule-governed patterned sequences of visual shapes (ABB, AAB, and ABA) that could be discriminated from differences in the position of the repeated element (late, early, or nonadjacent, respectively). Eight-month-olds were found to distinguish patterns on the basis of the repetition, but appeared insensitive to its position in the sequence; 11-month-olds distinguished patterns over the position of the repetition, but appeared insensitive to the nonadjacent repetition. These results suggest that abstract pattern detection may develop incrementally in a process of constructing complex relations from more primitive components.

By 7 months of age, infants are able to learn rules based on the abstract relationships between stimuli (Marcus et al., 1999), but they are better able to do so when exposed to speech than to some other classes of stimuli. In the current experiments we ask whether multimodal stimulus information will aid younger infants in identifying abstract rules. We habituated 5-month-olds to simple abstract patterns (ABA or ABB) instantiated in coordinated looming visual shapes and speech sounds (Experiment 1), shapes alone (Experiment 2), and speech sounds accompanied by uninformative but coordinated shapes (Experiment 3). Infants showed evidence of rule learning only in the presence of the informative multimodal cues. We hypothesize that the additional evidence present in these multimodal displays was responsible for the success of younger infants in learning rules, congruent with both a Bayesian account and with the Intersensory Redundancy Hypothesis.

Abstract Both nature and nurture are important contributors to development, but the two are not mutually exclusive. Understanding their joint contribution will require a balanced look at both contributors. This commentary proposes that the target article by J. P. Spencer et al. (2009) falls short in this regard, largely because it misrepresents the very “nativist” position it seeks to critique.

Cognitive scientists must understand not just what the mind does, but how it does what it does. In this paper, I consider four aspects of cognitive architecture: how the mind develops, the extent to which it is or is not modular, the extent to which it is or is not optimal, and the extent to which it should or should not be considered a symbol-manipulating device (as opposed to, say, an eliminative connectionist network). In each case, I argue that insights from developmental and evolutionary biology can lead to substantive and important compromises in historically vexed debates.

In this article it is argued that evolutionary plausibility must be made an important constraining factor when building theories of language. Recent suggestions that presume that language is necessarily a perfect or optimal system are at odds with this position, evolutionary theory showing us that evolution is a meliorizing agent often producing imperfect solutions. Perfection of the linguistic system is something that must be demonstrated, rather than presumed. Empirically, examples of imperfection are found not only in nature and in human cognition, but also in language — in the form of ambiguity, redundancy, irregularity, movement, locality conditions, and extra-grammatical idioms. Here it is argued that language is neither perfect nor optimal, and shown how theories of language which place these proper-ties at their core run into both conceptual and empirical problems.

A 'kluge' is an engineering term for a makeshift solution, an inelegant construction that somehow works. This is Gary Marcus' analogy for the way the human mind has evolved. Arguing against a whole tradition that praises our human minds as the most perfect result of evolution, Marcus shows how imperfect and ill-adapted our brains really are. They have had to adapt from the environment of our early hominid origins to a complex world in which our penchant for short-term satisfactions is literally fatal. We are prone to rages, addictions and other habits that limit our capacity for rational action in every sphere, from food to politics. This is a breathtaking, witty and revolutionary book.

Abstract Rogers & McClelland's (R&M's) précis represents an important effort to address key issues in concepts and categorization, but few of the simulations deliver what is promised. We argue that the models are seriously underconstrained, importantly incomplete, and psychologically implausible; more broadly, R&M dwell too heavily on the apparent successes without comparable concern for limitations already noted in the literature.

Gary Marcus provides an engaging argument for the two-part view that the mind is a kluge, and that it’s therefore not created by an intelligent designer, a.k.a., God. This argument, while philosophical, is based on empirical evidence culled from psychology. We distill Marcus’s argument, and briefly explain why it’s derailed by at least two objections. In the course of our presentation we summarize the book. What’s a kluge? Two definitions, derived directly from Marcus’s introductory chapter, constitute an answer. The first:

Sequences of speech sounds play a central role in human cognitive life, and the principles that govern such sequences are crucial in determining the syntax and semantics of natural languages. Infants are capable of extracting both simple transitional probabilities and simple algebraic rules from sequences of speech, as demonstrated by studies using ABB grammars (la ta ta, gai mu mu, etc.). Here, we report a striking finding: Infants are better able to extract rules from sequences of nonspeech--such as sequences of musical tones, animal sounds, or varying timbres--if they first hear those rules instantiated in sequences of speech.

This DOI is not currently attached to any metadata records. DOIs can’t actually ever be deleted (they’re persistent), but sometimes our members create DOIs in error. We do have a process to approximate deletion which we follow only in rare cases where the DOI has been genuinely created in error, and most crucially, if the DOI has never been published anywhere online or in print and never otherwise distributed to or communicated with anyone (authors, readers, reviewers, etc.

Editor Gary Marcus has carefully selected brief readings, mostly from popular trade books, that are both relevant and interesting to the introductory student.

Criteria that aim to dichotomize cognition into rules and similarity are destined to fail because rules and similarity are not in genuine conflict. It is possible for a given cognitive domain to exploit rules without similarity, similarity without rules, or both (rules and similarity) at the same time.

Abstract This chapter examines an apparent tension created by recent research on neurological development and genetics on the one hand and cognitive development on the other. It considers what it might mean for intrinsic signals to guide the initial establishment of functional architecture. It argues that an understanding of the mechanisms by which the body develops can inform our understanding of the mechanisms by which the brain develops. It cites the view of developmental neurobiologists Fukuchi-Shimogori and Grove, that the patterning of the part of the brain responsible for our higher functions is coordinated by the same basic mechanisms and signaling protein families used to generate patterning in other embryonic organs. Thus, what's good enough for the body, is good enough for the brain.

MacWhinney's list of proposed mechanisms for language acquisition seems so all-encompassing that it might appear churlish to dwell on a mechanism that's left out, but on his ultimate list of critical mechanisms there is an indeed an important (and intended) omission: innate, linguistically-specialized constraints. Such constraints are indeed mentioned, but if I understand the point of the article correctly, it is to argue that seven cognitive mechanisms, ranging from competition to cue construction, indirect negative evidence and monitoring, collectively obviate the need for any innate machinery that it specialized for language. Why not include room for innate, domain-specific constraints, too?

What Is Thought? . by Eric B. Baum. MIT Press, Cambridge, MA, 2004. 492 pp. $40, £25.95. ISBN 0-262-02548-5. Approaching the problem of the mind from a computer science perspective, Baum argues for a computational explanation of though--one that sees the mind as "a complex but compact program."

“Little by little, the child develops, ” wrote an undergraduate in a friend’s cognitive development class, and so, for the most part, it is. In most respects, a 3-year-old is better than a 2-year-old, and a 2-year-old is better than a 1-year-old. Whether improvement is gradual or steep, it is rare to see children take a step backward. When children do take that rare step backward, researchers stand up and listen. The three target articles in this issue are a look at some of those rare moments. “U ” shaped progressions, in which good performance is followed by worse performance that is finally followed by better performance should remind us (if perhaps in reverse) of the famous riddle of the Sphinx: what at first walks on 4 legs, then 2, and finally 3? We all know that the answer to the Sphinx’s riddle is a human being, progressing from crawling to walking unassisted and finally, late in life, walking with the assistance of the “leg ” of a cane. But what explains the U’s of cognitive development? Namy, Campbell, and Tomasello (this issue) and Cashon and Cohen (this issue) take a standard approach to understanding U-shaped curves: as the product of a mix of different cognitive processes that come on-line at different points in development. Gershkoff-Stowe and Thelen (this issue) take a rather different view, suggesting that we should look outside the head, away from “programs ” and “modules” and out toward the body and the environment. Indeed, in one of the first case studies they discuss, Gershkoff-Stowe and Thelen present a rather convincing story. What at first steps unassisted, then doesn’t, and finally does? A human child, born with the mental machinery for guiding stepping but also with a metabolism that leads her to gain weight faster than she gains muscle mass. The newborn steps because she can. The 6-week-old doesn’t because she can’t; her legs just aren’t strong enough. Even-

“Does grammar start where statistics stop?”, ask M. S. Seidenberg et al. in the title of their Perspective (18 Oct., p. [553][1]). Arguing against a “reconcilist” position in which complex cognitive functions would depend on a mixture of statistical and algebraic (rule) mechanisms ([1][2], [

This commentary discusses the logic of inferring modularity or the lack of modularity from observed patterns of developmental disorders.

In The Algebraic Mind, Gary Marcus attempts to integrate two theories about how the mind works, one that says that the mind is a computer-like manipulator of symbols, and another that says that the mind is a large network of neurons working together in parallel. Resisting the conventional wisdom that says that if the mind is a large neural network it cannot simultaneously be a manipulator of symbols, Marcus outlines a variety of ways in which neural systems could be organized so as to manipulate symbols, and he shows why such systems are more likely to provide an adequate substrate for language and cognition than neural systems that are inconsistent with the manipulation of symbols. Concluding with a discussion of how a neurally realized system of symbol-manipulation could have evolved and how such a system could unfold developmentally within the womb, Marcus helps to set the future agenda of cognitive neuroscience. Bradford Books imprint

In The Algebraic Mind, Gary Marcus attempts to integrate two theories about how the mind works, one that says that the mind is a computer-like manipulator of symbols, and another that says that the mind is a large network of neurons working together in parallel. Resisting the conventional wisdom that says that if the mind is a large neural network it cannot simultaneously be a manipulator of symbols, Marcus outlines a variety of ways in which neural systems could be organized so as to manipulate symbols, and he shows why such systems are more likely to provide an adequate substrate for language and cognition than neural systems that are inconsistent with the manipulation of symbols. Concluding with a discussion of how a neurally realized system of symbol-manipulation could have evolved and how such a system could unfold developmentally within the womb, Marcus helps to set the future agenda of cognitive neuroscience.

Abstract In particular, I propose that humans have both a statistical mechanism which extracts, collects, and tabulates statistical contingencies in the environment, and a rule mechanism which suppresses variation between examples within a category, thereby treating all instances of a category as equal members. Of course, neither mechanism is entirely novel; indeed, much of the debate in cognitive psychology concerns such mechanisms. My goal here is to argue that both mechanisms are necessary. Section 8.2 presents detailed arguments for the existence of both mechanisms within a single narrow linguistic domain, inflection. Section 8.3 proposes that these same two types of mechanisms are used in many other domains of cognition.

Some of the things that adults learn about language, and about the world, are very specific, whereas others are more abstract or rulelike. This article reviews evidence showing that infants, too, can very rapidly acquire both specific and abstract information, and considers the mechanisms that infants might use in doing so.

It is not altogether surprising that McClelland and Plaut, researchers with longstanding interests in providing alternatives to rules, find our recent experiments unconvincing [McClelland, J.L and Plaut, D.C. (1999) Does generalization in infant learning implicate abstract algebra-like rules? Trends Cognit. Sci. 3, 166–168]1McClelland J.L. Plaut D.C. Does generalization in infant learning implicate abstract algebra-like rules?.Trends Cognit. Sci. 1999; 3: 166-168Abstract Full Text Full Text PDF PubMed Google Scholar. But advocates of their cognition-without-rules view might want to look elsewhere to bolster their case, as none of McClelland and Plaut’s objections turns out to be plausible. Before addressing their objections, let me outline what I see as three important points of agreement. First, we all seem to be interested in the study of how cognition could be realized in a neural substrate. Second, we all believe that the study of neural networks can be helpful in this regard. Third, we agree that a basic property of the class of models that McClelland and Plaut advocate is that they depend on the overlap of features. As they put it: generalization in neural networks depends on overlap of representations – that is, the patterns of activity used in the network – to represent items experienced during training and test. For prior learning to generalize to a new stimulus, the representation of the new stimulus must overlap with – that is, activate some units in common with – the representation of the stimuli on which learning is based. This is because learning occurs by the adjustment of connection weights between specific units in a network, and so a new input must activate some of the same units whose weights were influenced by prior experience to benefit from that experience. (1McClelland J.L. Plaut D.C. Does generalization in infant learning implicate abstract algebra-like rules?.Trends Cognit. Sci. 1999; 3: 166-168Abstract Full Text Full Text PDF PubMed Google Scholar, p. 166.) As it turns out, I made almost exactly this point in a recent article2Marcus G.F. Rethinking eliminative connectionism.Cognit. Psychol. 1998; 37: 243-282Crossref PubMed Scopus (192) Google Scholar. Where we seem to disagree is with the implications of this fact about overlap. The problem, as I see it, is that this inability to generalize to non-overlapping items renders a certain class of network models inappropriate for many cognitive tasks, because in many cognitive tasks we are required to generalize to new items that do not overlap with the items that we have seen before. The experiments in our Science article3Marcus G.F. et al.Rule learning in 7-month-old infants.Science. 1999; 283: 77-80Crossref PubMed Scopus (810) Google Scholar, discussed by McClelland and Plaut, were designed to address precisely this point, testing whether infants could generalize an abstract relation, such as the one found in ‘ABA’ grammar consisting of sentences such as ‘li-ti-li’ and ‘ga-ti-ga’, to novel items that didn’t overlap with previous items. Infants were able to do this, discriminating consistent test sentences like ‘wo-fe-wo’ from inconsistent test sentences like ‘wo-fe-fe’. If words are represented as independent items, as they are in, say, Elman’s work on the popular simple recurrent network4Elman J.L. Finding structure in time.Cognit. Sci. 1990; 14: 179-211Crossref Google Scholar, the test items do not overlap with the habituation items, thus the standard version of the simple recurrent network does not succeed in making the discrimination. This is exactly as would be expected, given McClelland and Plaut’s discussion about the importance of overlap. (This is an entirely replicable result, using a wide range of network parameters. Readers who wish to verify this for themselves can look at the sample files we provide at http://www.psych.nyu.edu/∼gary/es.html.) Furthermore, our experiments were designed in such a way that even if the input were encoded as set

In their response to van der Lely et al.1van der Lely H. Rosen S. McClelland A. Evidence for a grammar-specific deficit in children.Curr. Biol. 1998; 8: 1253-1258Abstract Full Text Full Text PDF PubMed Scopus (119) Google Scholar, Tomblin and Pandich appear to doubt the possibility that ‘specific aspects of language content’ are ‘under specific genetic influence’ on the grounds that ‘the only thing genes do is code for proteins or control the expression of other genes that code for proteins. As proteins are highly conserved in nature, it would seem unlikely that there could even be one new protein in biology that has a special role for language let alone the number that would probably be necessary for influencing language’2Tomblin J.B. Pandich J. Lessons from children with specific language impairment.Trends Cognit. Sci. 1999; 3: 283-285Abstract Full Text Full Text PDF PubMed Scopus (24) Google Scholar. The implicit premise seems to be that language could be under specific genetic influence only if humans possessed a number of proteins that cannot be found elsewhere in the animal kingdom. Tomblin and Pandich could well be correct that there are no human-specific proteins, but this would not lead us to infer that humans could not possibly have language. The implicit premise that novel structure depends on novel proteins is demonstrably false. For example, old proteins can be adapted for new uses. The lens of all vertebrate eyes is made up in large part of a class of proteins known as crystallins, which contribute to the ability of the lens to focus. Some of these crystallins probably originally existed for the purpose of protection from heat, and only later did they become involved in the bending of tissue that contributes to focusing; the new function of focusing thus might have depended not on the existence of a new protein, but on a set of mechanisms that co-opted an old set of proteins to new uses3Gerhart J. Kirschner M. Cells, Embryos, and Evolution. Blackwell Science, 1997Google Scholar. By acting as switches, specific ‘master control genes’ can trigger complex, hierarchical cascades of genes that elicit widely varying arrangements of cells, even in the absence of species-particular proteins4Gehring W.J. Master Control Genes in Development and Evolution: The Homeobox Story. Yale University Press, 1998Google Scholar. Van der Lely et al.’s view that aspects of language may be under specific genetic control thus remains viable. It could turn out, for instance, that the language impairments that van der Lely et al. observed reflect the absence of some particular gene that ordinarily triggers a cascade of events that leads to the construction of machinery that uniquely subserves language. As Tooby and Cosmides put it, ‘There are millions of animal species on earth, each with a different set of cognitive programs. The same basic neural tissue embodies all of these programs, and it could support many others as well. Facts about the properties of neurons, neurotransmitters, and cellular development cannot tell you which of these millions of programs the human mind contains. Even if all neural activity is the expression of a uniform process at the cellular level, it is the arrangement of neurons – into bird song templates or web-spinning programs – that matters’5Cosmides L. Tooby J. Beyond intuition and instinct blindness: toward an evolutionary rigorous cognitive science.Cognition. 1994; 50: 41-77Crossref PubMed Scopus (283) Google Scholar.

Humans routinely generalize universal relationships to unfamiliar instances. If we are told "if glork then frum," and "glork," we can infer "frum"; any name that serves as the subject of a sentence can appear as the object of a sentence. These universals are pervasive in language and reasoning. One account of how they are generalized holds that humans possess mechanisms that manipulate symbols and variables; an alternative account holds that symbol-manipulation can be eliminated from scientific theories in favor of descriptions couched in terms of networks of interconnected nodes. Can these "eliminative" connectionist models offer a genuine alternative? This article shows that eliminative connectionist models cannot account for how we extend universals to arbitrary items. The argument runs as follows. First, if these models, as currently conceived, were to extend universals to arbitrary instances, they would have to generalize outside the space of training examples. Next, it is shown that the class of eliminative connectionist models that is currently popular cannot learn to extend universals outside the training space. This limitation might be avoided through the use of an architecture that implements symbol manipulation.

Constructivism is the Piagetian notion that learning leads the child to develop new types of representations. For example, on the Piagetian view, a child is born without knowing that objects persist in time even when they are occluded; through a process of learning, the child comes to know that objects persist in time. The trouble with this view has always been the lack of a concrete, computational account of how a learning mechanism could lead to such a change. Recently, however, in a book entitled Rethinking Innateness. Elman et al. (Elman, J.L., Bates, E., Johnson, M.H., Karmiloff-Smith, A., Parisi, D., Plunkett, K., 1996. Rethinking Innateness: A Connectionist Perspective on Development. Cambridge, MA: MIT Press) have claimed that connectionist models might provide an account of the development of new kinds of representations that would not depend on the existence of innate representations. I show that the models described in Rethinking Innateness depend on innately assumed representations and that they do not offer a genuine alternative to nativism. Moreover, I present simulation results which show that these models are incapable of deriving genuine abstract representations that are not presupposed. I then give a formal account of why the models fail to generalize in the ways that humans do. Thus, connectionism, at least in its current form, does not provide any support for constructivism. I conclude by sketching a possible alternative.

Connectionist networks excel at extracting statistical regularities but have trouble extracting higher-order relationships. Clark & Thornton suggest that a solution to this problem might come from Elman (1993), but I argue that the success of Elman's single recurrent network is illusory, and show that it cannot in fact represent abstract relationships that can be generalized to novel instances, undermining Clark & Thornton's key arguments.

This paper brings a quantitative study of children's noun plural overregularizations (foots, mans) to bear on recent comparisons of connectionist and symbolic models of language. The speech of 10 English-speaking children (aged 1;3 to 5;2) from the CHILDES database (MacWhinney & Snow, 1985, 1990) were analysed. The rate of noun overregularization is low, mean = 8.5%, demonstrating that children prefer correct to overregularized forms. Rates of noun overregularization are not significantly different from their rates of past tense overregularization, and noun plurals, like verb past tenses, follow a U-shaped developmental curve in which correct irregulars precede the first overregularized forms. These facts suggest that plural and past tense overregularizations are caused by similar underlying processes. The results pose challenges to connectionist models, but are consistent with Marcus et al.'s (1992) blocking-and-retrieval-failure model in which regulars are generated by a default rule while irregulars are retrieved from the lexicon.

A chartered accountant says the problem of fraud within medical practices is far from rare, even though prevention is relatively simple. Gary Marcus provides advice to physicians about ways to spot fraud within their practices and things they can do to prevent it. No matter how small your practice or how long you have had the same employees, says Marcus, given the right circumstances anyone might steal from you.

What is the input to the mental system that computes inflected forms like walked, came, dogs, and men? Recent connectionist models feed a word's phonological features into a single network, allowing it to generalize both regular and irregular phonological patterns, like stop-stopped, step-stepped and fling-flung, cling-clung. But for adults, phonological input is insufficient: verbs derived from nouns like ring the city always have regular past tense forms (ringed), even if they are phonologically identical to irregular verbs (ring the bell). Similarly, nouns based on names, like two Mickey Mouses, and compounds based on possessing rather than being their root morpheme, such as two sabertooths, take regular plurals, even when they are homophonous with irregular nouns like mice and teeth. In four experiments, testing 70 three- to ten-year-old children, we found that children are sensitive to such nonphonological information: they were more likely to produce regular inflected forms for forms like to ring ('to put a ring on') and snaggletooth (a kind of animal doll with big teeth) than for their homophonous irregular counterparts, even when these counterparts were also extended in meaning. Children's inflectional systems thus seem to be like adults': irregular forms are tied to the lexicon but regular forms are computed by a default rule, and words are represented as morphological tree structures reflecting their derivation from basic word roots. Such structures, which determine how novel complex words are derived and interpreted, also govern whether words with irregular sound patterns will be regularized: a word can be irregular only if its structure contains an irregular root in 'head' position, allowing the lexically stored irregular information to percolate up to apply to the word as a whole. In all other cases, the inflected form is compouted by a default regular rule. This proposal fits the facts better than alternatives appealing to ambiguity reduction or semantic similarity to a word's central sense. The results, together with an analysis of adult speech to children, suggest that morphological structure and a distinction between mechanisms for regular and irregular inflection may be inherent to the design of children's language systems.

An abstract is not available for this content so a preview has been provided. Please use the Get access link above for information on how to access this content.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1993.

Children extend regular grammatical patterns to irregular words, resulting in overregularizations like comed, often after a period of correct performance ("U-shaped development"). The errors seem paradigmatic of rule use, hence bear on central issues in the psychology of rules: how creative rule application interacts with memorized exceptions in development, how overgeneral rules are unlearned in the absence of parental feedback, and whether cognitive processes involve explicit rules or parallel distributed processing (connectionist) networks. We remedy the lack of quantitative data on overregularization by analyzing 11,521 irregular past tense utterances in the spontaneous speech of 83 children. Our findings are as follows. (1) Overregularization errors are relatively rare (median 2.5% of irregular past tense forms), suggesting that there is no qualitative defect in children's grammars that must be unlearned. (2) Overregularization occurs at a roughly constant low rate from the 2s into the school-age years, affecting most irregular verbs. (3) Although overregularization errors never predominate, one aspect of their purported U-shaped development was confirmed quantitatively: an extended period of correct performance precedes the first error. (4) Overregularization does not correlate with increases in the number or proportion of regular verbs in parental speech, children's speech, or children's vocabularies. Thus, the traditional account in which memory operates before rules cannot be replaced by a connectionist alternative in which a single network displays rotelike or rulelike behavior in response to changes in input statistics. (5) Overregularizations first appear when children begin to mark regular verbs for tense reliably (i.e., when they stop saying Yesterday I walk). (6) The more often a parent uses an irregular form, the less often the child overregularizes it. (7) Verbs are protected from overregularization by similar-sounding irregulars, but they are not attracted to overregularization by similar-sounding regulars, suggesting that irregular patterns are stored in an associative memory with connectionist properties, but that regulars are not. We propose a simple explanation. Children, like adults, mark tense using memory (for irregulars) and an affixation rule that can generate a regular past tense form for any verb. Retrieval of an irregular blocks the rule, but children's memory traces are not strong enough to guarantee perfect retrieval. When retrieval fails, the rule is applied, and overregularization results.