{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,8,1]],"date-time":"2026-08-01T06:19:58Z","timestamp":1785565198358,"version":"3.56.0"},"update-to":[{"DOI":"10.1371\/journal.pcbi.1011316","type":"new_version","label":"New version","source":"publisher","updated":{"date-parts":[[2023,9,12]],"date-time":"2023-09-12T00:00:00Z","timestamp":1694476800000}}],"reference-count":59,"publisher":"Public Library of Science (PLoS)","issue":"8","license":[{"start":{"date-parts":[[2023,8,25]],"date-time":"2023-08-25T00:00:00Z","timestamp":1692921600000},"content-version":"vor","delay-in-days":0,"URL":"http:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"DOI":"10.13039\/100019274","name":"NIH","doi-asserted-by":"publisher","award":["T32MH065214"],"award-info":[{"award-number":["T32MH065214"]}],"id":[{"id":"10.13039\/100019274","id-type":"DOI","asserted-by":"publisher"}]},{"DOI":"10.13039\/100016461","name":"Life Sciences Division, Army Research Office","doi-asserted-by":"publisher","award":["W911NF-16-1-0474"],"award-info":[{"award-number":["W911NF-16-1-0474"]}],"id":[{"id":"10.13039\/100016461","id-type":"DOI","asserted-by":"publisher"}]},{"DOI":"10.13039\/100006502","name":"Defense Sciences Office, DARPA","doi-asserted-by":"publisher","award":["L2M"],"award-info":[{"award-number":["L2M"]}],"id":[{"id":"10.13039\/100006502","id-type":"DOI","asserted-by":"publisher"}]},{"DOI":"10.13039\/100000925","name":"John Templeton Foundation","doi-asserted-by":"publisher","id":[{"id":"10.13039\/100000925","id-type":"DOI","asserted-by":"publisher"}]}],"content-domain":{"domain":["www.ploscompbiol.org"],"crossmark-restriction":false},"short-container-title":["PLoS Comput Biol"],"abstract":"<jats:p>The ability to acquire abstract knowledge is a hallmark of human intelligence and is believed by many to be one of the core differences between humans and neural network models. Agents can be endowed with an inductive bias towards abstraction through meta-learning, where they are trained on a distribution of tasks that share some abstract structure that can be learned and applied. However, because neural networks are hard to interpret, it can be difficult to tell whether agents have learned the underlying abstraction, or alternatively statistical patterns that are characteristic of that abstraction. In this work, we compare the performance of humans and agents in a meta-reinforcement learning paradigm in which tasks are generated from abstract rules. We define a novel methodology for building \u201ctask metamers\u201d that closely match the statistics of the abstract tasks but use a different underlying generative process, and evaluate performance on both abstract and metamer tasks. We find that humans perform better at abstract tasks than metamer tasks whereas common neural network architectures typically perform worse on the abstract tasks than the matched metamers. This work provides a foundation for characterizing differences between humans and machine learning that can be used in future work towards developing machines with more human-like behavior.<\/jats:p>","DOI":"10.1371\/journal.pcbi.1011316","type":"journal-article","created":{"date-parts":[[2023,8,25]],"date-time":"2023-08-25T17:36:51Z","timestamp":1692985011000},"page":"e1011316","update-policy":"https:\/\/doi.org\/10.1371\/journal.pcbi.corrections_policy","source":"Crossref","is-referenced-by-count":2,"title":["Disentangling Abstraction from Statistical Pattern Matching in Human and Machine Learning"],"prefix":"10.1371","volume":"19","author":[{"ORCID":"https:\/\/orcid.org\/0000-0003-1769-5147","authenticated-orcid":true,"given":"Sreejan","family":"Kumar","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Ishita","family":"Dasgupta","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Nathaniel D.","family":"Daw","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Jonathan. D.","family":"Cohen","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Thomas L.","family":"Griffiths","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"340","published-online":{"date-parts":[[2023,8,25]]},"reference":[{"issue":"2\u20133","key":"pcbi.1011316.ref001","doi-asserted-by":"crossref","first-page":"323","DOI":"10.1016\/0004-3702(92)90021-O","article-title":"A theory of abstraction","volume":"57","author":"F. Giunchiglia","year":"1992","journal-title":"Artificial intelligence"},{"issue":"1","key":"pcbi.1011316.ref002","doi-asserted-by":"crossref","first-page":"i","DOI":"10.1037\/h0093130","article-title":"Quantitative aspects of evolution of concepts: An experimental study","volume":"28","author":"C. L. Hull","year":"1920","journal-title":"Psychological Monographs"},{"key":"pcbi.1011316.ref003","volume-title":"A Study of Thinking","author":"J.S. Bruner","year":"1956"},{"issue":"6","key":"pcbi.1011316.ref004","doi-asserted-by":"crossref","first-page":"1218","DOI":"10.1037\/a0029334","article-title":"A century of Gestalt psychology in visual perception: II. Conceptual and theoretical foundations","volume":"138","author":"J. Wagemans","year":"2012","journal-title":"Psychological Bulletin"},{"issue":"6","key":"pcbi.1011316.ref005","doi-asserted-by":"crossref","first-page":"475","DOI":"10.1016\/j.tics.2021.01.006","article-title":"The perception of relations","volume":"25","author":"A. Hafri","year":"2021","journal-title":"Trends in Cognitive Sciences"},{"key":"pcbi.1011316.ref006","doi-asserted-by":"crossref","first-page":"101717","DOI":"10.1016\/j.pneurobio.2019.101717","article-title":"Structure learning and the posterior parietal cortex","volume":"184","author":"C. Summerfield","year":"2020","journal-title":"Progress in neurobiology"},{"issue":"2","key":"pcbi.1011316.ref007","doi-asserted-by":"crossref","first-page":"490","DOI":"10.1016\/j.neuron.2018.10.002","article-title":"What is a cognitive map? Organizing knowledge for flexible behavior","volume":"100","author":"T. E. Behrens","year":"2018","journal-title":"Neuron"},{"key":"pcbi.1011316.ref008","doi-asserted-by":"crossref","first-page":"111","DOI":"10.1016\/j.cobeha.2019.05.001","article-title":"The value of abstraction","volume":"29","author":"M. K. Ho","year":"2019","journal-title":"Current Opinion in Behavioral Sciences"},{"issue":"4","key":"pcbi.1011316.ref009","doi-asserted-by":"crossref","first-page":"352","DOI":"10.1016\/j.cub.2009.01.036","article-title":"Motor task variation induces structural learning","volume":"19","author":"D. A. Braun","year":"2009","journal-title":"Current Biology"},{"key":"pcbi.1011316.ref010","article-title":"Learning to communicate about shared procedural abstractions","author":"W. P. McCarthy","year":"2021","journal-title":"arXiv preprint arXiv:2107.00077."},{"key":"pcbi.1011316.ref011","first-page":"40","article-title":"Building machines that learn and think like people","author":"B. M. Lake","year":"2017","journal-title":"Behavioral and brain sciences"},{"key":"pcbi.1011316.ref012","author":"S. Dehaene","year":"2021","journal-title":"How we learn: Why brains learn better than any machine\u2026 for now. Penguin"},{"key":"pcbi.1011316.ref013","doi-asserted-by":"crossref","first-page":"101527","DOI":"10.1016\/j.cogpsych.2022.101527","article-title":"A language of thought for the mental representation of geometric shapes","volume":"139","author":"M. Sabl\u00e9-Meyer","year":"2022","journal-title":"Cognitive Psychology"},{"issue":"1","key":"pcbi.1011316.ref014","doi-asserted-by":"crossref","first-page":"79","DOI":"10.1111\/nyas.14619","article-title":"Abstraction and analogy-making in artificial intelligence","volume":"1505","author":"M. Mitchell","year":"2021","journal-title":"Annals of the New York Academy of Sciences"},{"key":"pcbi.1011316.ref015","doi-asserted-by":"crossref","first-page":"1988","DOI":"10.1109\/CVPR.2017.215","volume-title":"In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)","author":"J. Johnson","year":"2017"},{"key":"pcbi.1011316.ref016","article-title":"Human-level reinforcement learning through theory-based modeling, exploration, and planning","author":"P. A. Tsividis","year":"2021","journal-title":"arXiv preprint arXiv:2107.12544"},{"key":"pcbi.1011316.ref017","article-title":"On the measure of intelligence","author":"F. Chollet","year":"2019","journal-title":"arXiv preprint arXiv:1911.01547"},{"key":"pcbi.1011316.ref018","first-page":"5317","article-title":"Raven: A dataset for relational and analogical visual reasoning","author":"C. Zhang","year":"2019","journal-title":"In Proceedings of the IEEE\/CVF conference on computer vision and pattern recognition"},{"key":"pcbi.1011316.ref019","article-title":"The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision","author":"J. Mao","journal-title":"arXiv preprint arXiv:1904.12584"},{"key":"pcbi.1011316.ref020","doi-asserted-by":"crossref","first-page":"24","DOI":"10.1016\/j.cobeha.2019.01.005","article-title":"Doing more with less: meta-reasoning and meta-learning in humans and machines","volume":"29","author":"T. L. Griffiths","year":"2019","journal-title":"Current Opinion in Behavioral Sciences"},{"key":"pcbi.1011316.ref021","article-title":"Meta-Learning in Neural Networks: A Survey","author":"T. M. Hospedales","year":"2021","journal-title":"IEEE transactions on pattern analysis and machine intelligence"},{"key":"pcbi.1011316.ref022","article-title":"Causal reasoning from meta-reinforcement learning","author":"I. Dasgupta","year":"2019","journal-title":"arXiv preprint arXiv:1901.08162"},{"key":"pcbi.1011316.ref023","article-title":"Compositional generalization through meta sequence-to-sequence learning","author":"B. M. Lake","journal-title":"arXiv preprint arXiv:1906.05381"},{"key":"pcbi.1011316.ref024","article-title":"Universal linguistic inductive biases via meta-learning","author":"R. T. McCoy","year":"2020","journal-title":"arXiv preprint arXiv:2006.16324"},{"key":"pcbi.1011316.ref025","first-page":"4218","article-title":"Machine theory of mind. In International conference on machine learning","author":"N. Rabinowitz","year":"2018","journal-title":"ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations"},{"issue":"6","key":"pcbi.1011316.ref026","doi-asserted-by":"crossref","first-page":"860","DOI":"10.1038\/s41593-018-0147-8","article-title":"Prefrontal cortex as a meta-reinforcement learning system","volume":"21","author":"J. X. Wang","year":"2018","journal-title":"Nature neuroscience"},{"key":"pcbi.1011316.ref027","article-title":"ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness","author":"R. Geirhos","year":"2018","journal-title":"arXiv preprint arXiv:1811.12231"},{"issue":"6266","key":"pcbi.1011316.ref028","doi-asserted-by":"crossref","first-page":"1332","DOI":"10.1126\/science.aab3050","article-title":"Human-level concept learning through probabilistic program induction","volume":"350","author":"B. M. Lake","year":"2015","journal-title":"Science"},{"key":"pcbi.1011316.ref029","first-page":"54","article-title":"People infer recursive visual concepts from just a few examples. Computational Brain &","volume":"3","author":"B. M. Lake","year":"2020","journal-title":"Behavior"},{"key":"pcbi.1011316.ref030","article-title":"Fast and flexible: Human program induction in abstract reasoning tasks","author":"A. Johnson","year":"2021","journal-title":"arXiv preprint arXiv:2103.05823"},{"issue":"2","key":"pcbi.1011316.ref031","doi-asserted-by":"crossref","first-page":"139","DOI":"10.1016\/1049-9660(92)90036-3","article-title":"Natural metamers","volume":"56","author":"M. S. Drew","year":"1992","journal-title":"CVGIP: Image Understanding"},{"key":"pcbi.1011316.ref032","first-page":"10078","article-title":"Metamers of neural networks reveal divergence from human perceptual systems","author":"J. Feather","year":"2019","journal-title":"In NeurIPS"},{"key":"pcbi.1011316.ref033","first-page":"2940","article-title":"Cognitive psychology for deep neural networks: A shape bias case study","author":"S. Ritter","year":"2017","journal-title":"In International conference on machine learning"},{"key":"pcbi.1011316.ref034","article-title":"Probing physics knowledge using tools from developmental psychology","author":"L. Piloto","year":"2018","journal-title":"arXiv preprint arXiv:1804.01128"},{"issue":"34","key":"pcbi.1011316.ref035","article-title":"Does the utility of information influence sampling behavior?","volume":"34","author":"D. Markant","year":"2012","journal-title":"In Proceedings of the annual meeting of the cognitive science society"},{"key":"pcbi.1011316.ref036","volume-title":"Perceptrons: An introduction to computational geometry","author":"M. Minsky","year":"1969"},{"key":"pcbi.1011316.ref037","first-page":"2873","article-title":"Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks","author":"B. Lake","year":"2018","journal-title":"In International conference on machine learning"},{"key":"pcbi.1011316.ref038","article-title":"Evaluating compositionality in sentence embeddings","author":"I. Dasgupta","year":"2018","journal-title":"arXiv preprint arXiv:1802.04302"},{"key":"pcbi.1011316.ref039","article-title":"Bert: Pre-training of deep bidirectional transformers for language understanding","author":"J. Devlin","year":"2018","journal-title":"arXiv preprint arXiv:1810.04805"},{"key":"pcbi.1011316.ref040","first-page":"341","article-title":"Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on","author":"A. A. Efros","year":"2001","journal-title":"Computer graphics and interactive techniques"},{"issue":"6","key":"pcbi.1011316.ref041","doi-asserted-by":"crossref","first-page":"721","DOI":"10.1109\/TPAMI.1984.4767596","article-title":"Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images","author":"S. Geman","year":"1984","journal-title":"IEEE Transactions on pattern analysis and machine intelligence"},{"key":"pcbi.1011316.ref042","article-title":"Rl2: Fast reinforcement learning via slow reinforcement learning","author":"Y. Duan","year":"2016","journal-title":"arXiv preprint arXiv:1611.02779"},{"key":"pcbi.1011316.ref043","article-title":"How to Learn and Represent Abstractions: An Investigation using Symbolic Alchemy","author":"B. AlKhamissi","year":"2021","journal-title":"arXiv preprint arXiv:2112.08360"},{"key":"pcbi.1011316.ref044","article-title":"Rapid Task-Solving in Novel Environments","author":"S. Ritter","year":"2021","journal-title":"In International Conference on Learning Representations"},{"key":"pcbi.1011316.ref045","first-page":"5998","article-title":"Attention is all you need","author":"A. Vaswani","year":"2017","journal-title":"In Advances in neural information processing systems"},{"key":"pcbi.1011316.ref046","article-title":"On Neural Architecture Inductive Biases for Relational Tasks","author":"G. Kerg","year":"2022","journal-title":"arXiv preprint arXiv:2206.05056"},{"key":"pcbi.1011316.ref047","article-title":"An image is worth 16x16 words: Transformers for image recognition at scale","author":"A. Dosovitskiy","journal-title":"arXiv preprint arXiv:2010.11929"},{"key":"pcbi.1011316.ref048","first-page":"1928","article-title":"Asynchronous methods for deep reinforcement learning","author":"V. Mnih","year":"2016","journal-title":"In International conference on machine learning"},{"key":"pcbi.1011316.ref049","article-title":"Emergent symbols through binding in external memory","author":"T. W. Webb","year":"2020","journal-title":"arXiv preprint arXiv:2012.14601."},{"key":"pcbi.1011316.ref050","first-page":"16691","article-title":"Recurrent model-free rl can be a strong baseline for many pomdps","author":"T. Ni","year":"2022","journal-title":"In the International Conference on Machine Learning"},{"key":"pcbi.1011316.ref051","first-page":"4477","article-title":"Measuring abstract reasoning in neural networks","author":"A. Santoro","year":"2018","journal-title":"In International Conference on Machine Learning"},{"issue":"3","key":"pcbi.1011316.ref052","first-page":"458","article-title":"Statistical learning of higher-order temporal structure from visual shape sequences","volume":"28","author":"J. Fiser","year":"2002","journal-title":"Journal of Experimental Psychology: Learning, Memory, and Cognition"},{"issue":"8","key":"pcbi.1011316.ref053","doi-asserted-by":"crossref","first-page":"1087","DOI":"10.1038\/s41562-021-01259-6","article-title":"Rational arbitration between statistics and rules in human sequence processing","volume":"6","author":"M. Maheu","year":"2022","journal-title":"Nature Human Behaviour"},{"key":"pcbi.1011316.ref054","article-title":"Relational inductive biases, deep learning, and graph networks","author":"P. W. Battaglia","year":"2018","journal-title":"arXiv preprint arXiv:1806.01261"},{"key":"pcbi.1011316.ref055","doi-asserted-by":"crossref","first-page":"835","DOI":"10.1145\/3453483.3454080","article-title":"DreamCoder: bootstrapping inductive program synthesis with wake-sleep library learning. In Proceedings of the 42nd","author":"K. Ellis","year":"2021","journal-title":"ACM SIGPLAN International Conference on Programming Language Design and Implementation"},{"key":"pcbi.1011316.ref056","article-title":"On the opportunities and risks of foundation models","author":"R. Bommasani","year":"2021","journal-title":"arXiv preprint arXiv:2108.07258"},{"issue":"1","key":"pcbi.1011316.ref057","doi-asserted-by":"crossref","first-page":"51","DOI":"10.1002\/rsa.3240010105","article-title":"Asymptotic analysis of a random walk on a hypercube with many dimensions","volume":"1","author":"P. Diaconis","year":"1990","journal-title":"Random Structures & Algorithms"},{"key":"pcbi.1011316.ref058","author":"A Hill","year":"2018","journal-title":"Stable baselines"},{"key":"pcbi.1011316.ref059","article-title":"Algorithms for hyper-parameter optimization","volume":"24","author":"J. Bergstra","year":"2011","journal-title":"Advances in neural information processing systems"}],"updated-by":[{"DOI":"10.1371\/journal.pcbi.1011316","type":"new_version","label":"New version","source":"publisher","updated":{"date-parts":[[2023,9,12]],"date-time":"2023-09-12T00:00:00Z","timestamp":1694476800000}}],"container-title":["PLOS Computational Biology"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/dx.plos.org\/10.1371\/journal.pcbi.1011316","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2023,9,12]],"date-time":"2023-09-12T17:32:56Z","timestamp":1694539976000},"score":1,"resource":{"primary":{"URL":"https:\/\/dx.plos.org\/10.1371\/journal.pcbi.1011316"}},"subtitle":[],"editor":[{"given":"Stefano","family":"Palminteri","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"editor"}]}],"short-title":[],"issued":{"date-parts":[[2023,8,25]]},"references-count":59,"journal-issue":{"issue":"8","published-online":{"date-parts":[[2023,8,25]]}},"URL":"https:\/\/doi.org\/10.1371\/journal.pcbi.1011316","relation":{"new_version":[{"id-type":"doi","id":"10.1371\/journal.pcbi.1011316","asserted-by":"object"}]},"ISSN":["1553-7358"],"issn-type":[{"value":"1553-7358","type":"electronic"}],"subject":[],"published":{"date-parts":[[2023,8,25]]}}}