Pith. sign in

REVIEW 4 major objections 5 minor 187 references

The paper argues that a nine-level 'ladder of learning and reasoning' combined with energy-efficient hardware can cut AI energy use by a thousandfold, replacing today's energy-hungry large language models with nimble domain-specific agents

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 08:09 UTC pith:DAWC6JSF

load-bearing objection A competent vision survey whose 1000× energy-efficiency headline is conceded in the text to be an upper bound; worth engaging as a position paper, not as a research result. the 4 major comments →

arxiv 2510.22052 v2 pith:DAWC6JSF submitted 2025-10-24 cs.AI cs.LG

A Vision Toward Energy-Efficient Domain-Specific Artificial Intelligence Models and Agents

Abhijit Chatterjee , Niraj K. Jha , Jonathan D. Cohen , Thomas L. Griffiths , Hongjing Lu , Diana Marculescu , Ashiqur Rasul , Wenrui Xu
show 1 more author
Keshab K. Parhi
This is my paper
classification cs.AI cs.LG
keywords energy-efficient AIdomain-specific agentsladder of learning and reasoningprospective learninganalogical reasoninghyperdimensional computingmultiplicative energy benefitsbrain-like intelligence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that today's large language models are on the wrong track for general intelligence: they consume enormous energy, hallucinate, and only generalize within the distribution of their training data. It proposes a 'ladder of learning and reasoning' — nine levels from correlation-based machine learning to analogical reasoning — that, if climbed, would let small domain-specific agents reason, plan, and learn continuously in novel situations. The central claim is that combining these techniques with energy-efficient compute paradigms such as hyperdimensional computing, quantization, sparsity, and sublinear attention can multiply energy savings and reach at least 1000× better efficiency than current GPU-based systems. The paper is a vision: it lays out the pieces and sketches the path, but does not yet demonstrate the combined system.

Core claim

On its own terms, the paper proposes that brain-like general intelligence comes from a hierarchy of learning and reasoning capabilities rather than from model scale. The ladder's lower rungs are familiar — correlation-based machine learning, knowledge augmentation, multimodal perception — and its upper rungs are the ones the authors believe unlock efficiency: few-shot continual learning, curiosity, causal and counterfactual reasoning, meta-reasoning, relational and symbolic reasoning, and analogy. Because reasoning from small data and causal structure avoids brute-force training, each rung should reduce energy demand relative to the state of the art; the paper asserts these benefits multiply

What carries the argument

The central organizing object is the 'ladder of learning and reasoning' — nine levels (A–I) spanning from traditional correlation-based machine learning up to analogical reasoning and fluid intelligence. The paper proposes this ladder as the mechanism by which small domain-specific models can generalize to unseen 'unknown-unknown' situations using minimal data, and claims that climbing the ladder yields multiplicative energy benefits. A second key mechanism is 'prospective learning' — continual learning, constraints, curiosity, and causal estimation — posited as the alternative to the i.i.d. assumption behind current training, which the paper identifies as the source of brittleness and hallu

Load-bearing premise

The paper's 1000× target rests on the assumption that the energy savings of the individual techniques multiply when combined; if two techniques save energy by shrinking the same bottleneck, the combined saving falls short of the product, and the target is unsupported.

What would settle it

Build a prototype that combines two of the proposed techniques — e.g., quantized inference on a hyperdimensional classifier, or gradient interleaving with a sparse mixture-of-experts model — and measure energy per inference against each technique alone. If the combined saving is materially less than the product of the individual savings, the multiplicative premise fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the vision holds, a 10–20 billion parameter model specialized to a bounded domain could match or beat current trillion-parameter models on that domain while using a fraction of the energy.
  • Critical applications such as healthcare, robotics, and surveillance, where hallucination blocks deployment, could use reasoning agents grounded in causal structure and analogies rather than web-scraped correlations.
  • Hardware roadmaps would shift priority from bigger matrix engines to process-in-memory, hyperdimensional compute, and configurable systolic arrays for interleaved gradient computations.
  • Training energy, currently tens of GWh for frontier models, would become a minor cost for domain-specific agents that learn continually from small data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The multiplicative-combination premise is the weakest link: many listed techniques (quantization, sparsity, low-rank) attack the same memory-bandwidth bottleneck, so their combined savings may be sub-multiplicative and need separate validation.
  • The ladder suggests a concrete research program: measure energy per correct answer on out-of-distribution reasoning tasks for each rung; if analogical reasoning truly enables zero-shot transfer, it should beat fine-tuned LLMs on such tasks at equal energy.
  • A testable extension would build a domain-specific agent combining the ladder's upper rungs — such as a model that learns abstract symbol-like representations through external-memory binding plus causal reasoning — and compare its energy-accuracy tradeoff against a distilled LLM on a medical or robotics benchmark.
  • If the 1000× target is even partially reached, the economics of AI change: compute costs would become a smaller fraction of AI development, shifting competition toward data-efficient reasoning algorithms rather than scale.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents a vision for energy-efficient, domain-specific AI systems. It proposes a 'ladder of learning and reasoning' ranging from correlation-based learning (rung A) to analogical and fluid intelligence (rung I), and argues that moving up this ladder yields improved generalization and computational efficiency. It then surveys a broad set of techniques: analogical reasoning models (BART, PAM), prospective learning, metareasoning, relational and symbolic reasoning, multimodal perception, knowledge augmentation, hyperdimensional computing, energy-efficient training, quantization/sparsity, mixture of experts, sublinear attention, knowledge-graph superintelligence, Perceiver IO, Titans, and cognitive agent architectures. The central quantitative claim, stated in the abstract and conclusion, is that the described techniques 'will achieve ≥1000× better energy efficiency' relative to state of the art for targeted domain tasks, subject to accuracy, latency, and coverage constraints.

Significance. If the ≥1000× energy-efficiency claim were substantiated, the paper would point toward a genuinely important alternative to brute-force scaling of large language models. The manuscript brings together a credible set of researchers from cognitive science, AI, and hardware, and it identifies several concrete research directions—prospective learning, rational metareasoning, emergent symbol binding, gradient interleaving, and knowledge-graph-grounded curricula—that are individually of interest. The paper also provides a useful organizing framework (the ladder) and a broad survey of the relevant literature. However, the central quantitative claim is not derived, measured, or even given a clear baseline; no energy model, target domain, or evaluation protocol is specified. As a vision statement the paper has merit, but as a research paper the headline claim is currently unsupported.

major comments (4)
  1. [Section V / Abstract] The assertion that the described techniques 'will achieve ≥1000× better energy efficiency' is unsupported. No baseline is defined (state of the art for which task? GPU inference? training? full system?), no target domain is specified, and no calculation, simulation, or measurement is provided. Section II only states that combined use of the techniques 'can, at best, provide overall multiplicative energy benefits.' This is an upper bound, not a lower bound, and cannot justify a '≥1000×' claim. A lower-bound argument or empirical evidence is needed.
  2. [Section II, Fig. 2] Fig. 2 is said to show 'expected energy benefits from techniques at each step of the ladder relative to the state of the art, running on GPUs,' but no quantitative values, derivations, or citations for these benefits are given in the text. Without per-rung benefit estimates, the multiplicative path to 1000× cannot be checked. The paper should either provide explicit benefit ranges with sources or clearly label these as illustrative conjectures, which would preclude the strong 'will achieve' language.
  3. [Section II / Abstract] There is a level-of-analysis mismatch. Fig. 2 and the surveyed techniques concern algorithmic or GPU-level savings, while the abstract and conclusion claim 'system-level energy efficiencies' under 'accuracy, latency, and coverage constraints.' No model converts per-technique savings into system-level energy, and no account is taken of the energy overheads of the proposed methods themselves: grow-and-prune training adds retraining cost, metareasoning adds computation, hyperdimensional computing uses very wide vectors that may erode sparsity/quantization gains, and continual learning has memory and update overheads. The paper needs at least a coarse end-to-end energy model and an interaction analysis.
  4. [Section II] The load-bearing premise that ladder techniques combine multiplicatively is stated without analysis. The techniques are not obviously independent: multiple rungs target the same underlying compute budget (e.g., sparsity, quantization, and low-rank approximation all reduce MAC/memory costs and their benefits may overlap), and algorithmic efficiency gains may not translate to system-level energy savings under latency or accuracy constraints. The paper should discuss whether benefits multiply, add, or interfere, or should refrain from claiming a specific aggregate factor.
minor comments (5)
  1. [Equation (5)] The KL divergence formula in Eq. (5) appears malformed: the first term is missing the expectation, and the sign structure is unusual. Please check against the standard D_KL(π_θ || π_ref) definition.
  2. [General] Several author names and titles have spacing artifacts (e.g., 'V oyelstein', 'V ision', 'P erceiver', 'T ransformers'). A careful proofread is needed.
  3. [Section IV-B] The QwQ-Med-3 result is presented as evidence of domain-specific superintelligence but no performance numbers or comparisons are given. Either provide evaluation results or cite the original paper with concrete figures.
  4. [References] Reference [163] is cited as 'arXiv Preprint arXiv:2507.13966'; please use consistent citation style and, if possible, include a DOI or author list snippet to aid verification.
  5. [Section III-D] The claims about PermDNN and tensor decomposition would benefit from quantitative summaries (e.g., achieved compression ratios or energy reductions) rather than qualitative descriptions.

Circularity Check

0 steps flagged

No circularity: the paper's central 1000x efficiency claim is an asserted vision, not a derived prediction, and the cited prior work is illustrative rather than load-bearing.

full rationale

The paper is a position/vision paper rather than a derivation. Its central quantitative assertion, 'We believe that the techniques described will achieve ≥1000× better energy efficiency as compared to the state of the art,' is stated in the conclusion (Section V) and abstract as a belief, not as the output of a calculation. Section II's statement that 'Combined use of these techniques can, at best, provide overall multiplicative energy benefits' is an upper-bound remark about how individual benefits might compose; it is not an equation that forces or constructs the 1000× claim. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work to make a choice forced, and no ansatz is smuggled in via citation in a way that carries the central claim. The many self-citations (BART, PAM, ESBN, PermDNN, LayerPipe, QwQ-Med, etc.) are presented as exemplars or supporting techniques; even if one disagrees with their sufficiency, the 1000× claim does not reduce to any of them by construction. The manuscript also contains openly conjectural and limitation language ('We broadly conjecture...', 'The reason behind HDC's failure in certain tasks is still under study', 'A limitation of this model...'), which supports the view that the authors are proposing a research agenda rather than claiming a derived result. Unsupport or lack of evidence is a correctness/epistemic concern, not circularity, under the stated rules. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper contributes no new entities. It relies on the assumed tractability of brain-inspired efficiency, the multiplicative combination of known techniques, and the sufficiency of small domain-specific models. These are the load-bearing assumptions behind the 1000× claim.

axioms (3)
  • domain assumption The human brain's ~20W power consumption is an appropriate and achievable benchmark for AI systems.
    Used in the abstract and introduction to motivate the energy-efficiency target; no argument establishes that engineered systems can reach this efficiency while matching the brain's capabilities.
  • ad hoc to paper The rungs of the 'ladder of learning and reasoning' combine multiplicatively in energy benefit.
    Section II states 'Combined use of these techniques can, at best, provide overall multiplicative energy benefits' without justification or analysis of interactions.
  • domain assumption Domain-specific 10–20B parameter models can achieve the desired intelligence for bounded domains.
    Stated in the abstract as the desired model size; no evidence is given that such models will reach the needed reasoning and planning abilities.

pith-pipeline@v1.3.0-alltime-deepseek · 33409 in / 5743 out tokens · 62505 ms · 2026-08-04T08:09:07.742423+00:00 · methodology

0 comments
read the original abstract

The field of artificial intelligence (AI) has taken a tight hold on broad aspects of society, industry, business, and governance in ways that dictate the prosperity and might of the world's economies. The AI market size is projected to grow from {\$}189 billion in 2023 to {\$}4.8 trillion by 2033. Currently, AI is dominated by large language models (LLMs) that exhibit linguistic and visual intelligence. However, training these models requires a massive amount of data scraped from the web as well as large amounts of energy (50-60 GWh to train GPT-4). Despite these costs, these models often hallucinate, a characteristic that prevents them from being deployed in critical application domains. In contrast, the human brain consumes only 20W of power. What is needed is the next level of AI evolution in which lightweight domain-specific multimodal models, especially compact models with 10--20B parameters for bounded domains, with higher levels of intelligence can reason, plan, and make decisions in dynamic environments with real-time data and prior knowledge, while learning continuously and evolving in ways that enhance future decision-making capability. This will define the next wave of AI, progressing from today's large models, trained with vast amounts of data, to nimble energy-efficient domain-specific agents that can reason and think in a world full of uncertainty. To support such agents, hardware will need to be reimagined to allow system-level energy efficiencies $\geq {1000X}$ over the state of the art for targeted domain tasks, subject to accuracy, latency, and coverage constraints. Such a vision of future AI systems is developed in this work.

Figures

Figures reproduced from arXiv: 2510.22052 by Abhijit Chatterjee, Ashiqur Rasul, Diana Marculescu, Hongjing Lu, Jonathan D. Cohen, Keshab K. Parhi, Niraj K. Jha, Thomas L. Griffiths, Wenrui Xu.

Figure 1
Figure 1. Figure 1: General intelligence, adapted from [22]. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ladder of learning and reasoning across unseen data from an unknown distribution (unknown￾unknown). A key research frontier is that of developing foun￾dations for broad generalization [18]–[21] using minimum amounts of additional data. This is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The intelligences are then fused together using higher [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Decision-making transformer [39]: (a) Modality-specific encoders take input from pocket devices and distill it into a representation, (b) a world model [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance on out-of-range extrapolation (Y axis; chance=14%), for [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Diagram of Emergent Symbol Binding Network (ESBN) model [71], [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example of an audio-visual event (AVE) representing the event of an individual speaking. The person’s voice is audible in all of the frames. Only [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: X-VILA schematic diagram [102]. X-VILA augments a pretrained LLM towards new modalities via (i) connecting pretrained encoders to the LLM [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: The growth of weight and activation stashing with number of pipeline [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

187 extracted references · 65 linked inside Pith

  1. [1]

    A comprehensive survey on pretrained foundation models: A history from BERT to ChatGPT,

    C. Zhou, Q. Li, C. Li, J. Yu, Y . Liu, G. Wang, K. Zhang, C. Ji, Q. Yan, L. He, H. Peng, J. Li, J. Wu, Z. Liu, P. Xie, C. Xiong, J. Pei, P. S. Yu, and L. Sun, “A comprehensive survey on pretrained foundation models: A history from BERT to ChatGPT,”arXiv preprint arXiv:2302.09419, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    ChatGPT: Jack of all trades, master of none,

    J. Koco ´n, I. Cichecki, O. Kaszyca, M. Kochanek, D. Szydło, J. Baran, J. Bielaniewicz, M. Gruza, A. Janz, K. Kanclerz, A. Koco´n, B. Koptyra, W. Mieleszczenko-Kowszewicz, P. Miłkowski, M. Oleksy, M. Piasecki, Ł. Radli ´nski, K. Wojtasik, S. Wo ´zniak, and P. Kazienko, “ChatGPT: Jack of all trades, master of none,”arXiv preprint arXiv:2302.10724, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Hierarchi- cal text-conditional image generation with CLIPlatents,

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchi- cal text-conditional image generation with CLIPlatents,”arXiv preprint arXiv:2204.06125, 2022

    Pith/arXiv arXiv 2022

  4. [4]

    Illiterate DALL-E learns to compose,

    G. Singh, F. Deng, and S. Ahn, “Illiterate DALL-E learns to compose,” arXiv preprint arXiv:2110.11405, 2021

    Pith/arXiv arXiv 2021

  5. [5]

    LaMDA: Language models for dialog applications,

    R. Thoppilan, D. De Freitas, J. Hall, N. Shazeer, A. Kulshreshtha, H.- T. Cheng, A. Jin, T. Bos, L. Baker, Y . Du, Y . Li, H. Lee, H. S. Zheng, A. Ghafouri, M. Menegali, Y . Huang, M. Krikun, D. Lepikhin, J. Qin, D. Chen, Y . Xu, Z. Chen, A. Roberts, M. Bosma, V . Zhao, Y . Zhou, C.-C. Chang, I. Krivokon, W. Rusch, M. Pickett, P. Srinivasan, L. Man, K. Me...

    Pith/arXiv arXiv 2022

  6. [6]

    PALM-E: An embodied multimodal language model,

    D. Driess, F. Xia, M. S. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, W. Huang, Y . Chebotar, P. Sermanet, D. Duckworth, S. Levine, V . Vanhoucke, K. Hausman, M. Toussaint, K. Greff, A. Zeng, I. Mordatch, and P. Florence, “PALM-E: An embodied multimodal language model,”arXiv preprint arXiv:2303.03378, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Sparks of artificial general intelligence: Early experiments with gpt-4,

    S. Bubeck, V . Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y . T. Lee, Y . Li, S. Lundberget al., “Sparks of artificial general intelligence: Early experiments with gpt-4,”arXiv preprint arXiv:2303.12712, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azharet al., “Llama: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Deepseek-r1: Incentivizing reasoning capability in LLMs via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in LLMs via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    AI doubling its compute every 3.5 months,

    T. Peng, “AI doubling its compute every 3.5 months,” https://medium.com/syncedreview, 2018

    2018

  11. [11]

    Gpt-4 architecture, datasets, costs and more leaked,

    The Decoder, “Gpt-4 architecture, datasets, costs and more leaked,” 2023, accessed: 2025-09-22. [Online]. Available: https://the-decoder. com/gpt-4-architecture-datasets-costs-and-more-leaked/

    2023

  12. [12]

    2024 report on U.S. data center energy use,

    Lawrence Berkeley National Laboratory, “2024 report on U.S. data center energy use,” https://www.energy.gov/articles/ doe-releases-new-report-evaluating-increase-electricity-demand-data-centers, 2024, u.S. Department of Energy announcement, Dec. 20, 2024. Projects U.S. data center electricity consumption to reach 325–580 TWh (6.7%–12% of U.S. demand) by 2028

    2024

  13. [13]

    GPTs and hallucination: Why do large language models hallucinate?

    J. Waldo and S. Boussard, “GPTs and hallucination: Why do large language models hallucinate?”Queue, vol. 22, no. 4, pp. 19–33, 2024

    2024

  14. [14]

    Retrieval-augmented generation for knowledge-intensive NLP tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-T. Yih, T. Rockt ¨aschel, S. Riedel, and D. Kiela, “Retrieval-augmented generation for knowledge-intensive NLP tasks,”Advances in Neural Information Processing Systems, vol. 33, pp. 9459–9474, 2020

    2020

  15. [15]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,”Advances in Neural Information Processing Systems, vol. 35, pp. 24 824–24 837, 2022

    2022

  16. [16]

    Tree of thoughts: Deliberate problem solving with large language models,

    S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y . Cao, and K. Narasimhan, “Tree of thoughts: Deliberate problem solving with large language models,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  17. [17]

    H. E. Gardner,Frames of Mind: The Theory of Multiple Intelligences. Basic Books, 2011

    2011

  18. [18]

    On the measure of intelligence,

    F. Chollet, “On the measure of intelligence,”arXiv preprint arXiv:1911.01547, 2019

    Pith/arXiv arXiv 1911

  19. [19]

    Towards a measure of general machine intelligence,

    G. Venkatasubramanian, S. Kar, A. Singh, S. Mishra, D. Yadav, and S. Chandak, “Towards a measure of general machine intelligence,” arXiv preprint arXiv:2109.12075, 2021

    Pith/arXiv arXiv 2021

  20. [20]

    General intelligence disentangled via a generality metric for natural and artificial intelligence,

    J. Hern ´andez-Orallo, B. S. Loe, L. Cheke, F. Mart ´ınez-Plumed, and S. ´O h ´Eigeartaigh, “General intelligence disentangled via a generality metric for natural and artificial intelligence,”Scientific reports, vol. 11, no. 1, p. 22822, 2021

    2021

  21. [21]

    Gathering strength, gathering storms: The one hundred year study on artificial intelligence (ai100) 2021 study panel report,

    M. L. Littman, I. Ajunwa, G. Berger, C. Boutilier, M. Currie, F. Doshi-Velez, G. Hadfield, M. C. Horowitz, C. Isbell, H. Kitano, K. Levy, T. Lyons, M. Mitchell, J. Shah, S. Sloman, S. Vallor, and T. Walsh, “Gathering strength, gathering storms: The one hundred year study on artificial intelligence (ai100) 2021 study panel report,”

    2021

  22. [22]

    Prospective learning: Back to the future,

    J. T. V ogelstein, T. Verstynen, K. P. Kording, L. Isik, J. W. Krakauer, R. Etienne-Cummings, E. L. Ogburn, C. E. Priebe, R. Burns, K. Kutten, J. J. Knierim, J. B. Potash, T. Hartung, L. Smirnova, P. Worley, A. Savonenko, I. Phillips, M. I. Miller, R. Vidal, J. Sulam, A. Charles, N. J. Cowan, M. Bichuch, A. Venkataraman, C. Li, N. Thakor, J. M. Kebschull,...

    Pith/arXiv arXiv 2022

  23. [23]

    An introduction to causal inference,

    J. Pearl, “An introduction to causal inference,”International Journal of Biostatistics, vol. 6, no. 2, 2010

    2010

  24. [24]

    A critical anal- ysis of metrics used for measuring progress in artificial intelligence,

    K. Blagec, G. Dorffner, M. Moradi, and M. Samwald, “A critical anal- ysis of metrics used for measuring progress in artificial intelligence,” arXiv preprint arXiv:2008.02577, 2020

    Pith/arXiv arXiv 2008

  25. [25]

    Building machines that learn and think like people,

    B. M. Lake, T. D. Ullman, J. B. Tenenbaum, and S. J. Gershman, “Building machines that learn and think like people,”Behavioral and Brain Sciences, vol. 40, 2017

    2017

  26. [26]

    Children’s performance on a spatial analogies task,

    D. Gentner, “Children’s performance on a spatial analogies task,”Child Development, pp. 1034–1039, 1977

    1977

  27. [27]

    A simple neural network module for relational reasoning,

    A. Santoro, D. Raposo, D. G. Barrett, M. Malinowski, R. Pascanu, P. Battaglia, and T. Lillicrap, “A simple neural network module for relational reasoning,”Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  28. [28]

    J. C. Raven and J. H. Court,Raven’s progressive matrices and vocabulary scales. Oxford Pyschologists Press Oxford, 1998, vol. 759

    1998

  29. [29]

    Abilities: Their structure, growth, and action,

    R. B. Cattell, “Abilities: Their structure, growth, and action,” 1971

    1971

  30. [30]

    Emergence of analogy from relation learning,

    H. Lu, Y . N. Wu, and K. J. Holyoak, “Emergence of analogy from relation learning,”Proceedings of the National Academy of Sciences, vol. 116, no. 10, pp. 4176–4181, 2019

    2019

  31. [31]

    Predicting patterns of similarity among abstract semantic relations,

    N. Ichien, H. Lu, and K. J. Holyoak, “Predicting patterns of similarity among abstract semantic relations,”Journal of Experimental Psychol- ogy: Learning, Memory, and Cognition, 2021

    2021

  32. [32]

    Distributed code for semantic relations predicts neural similarity dur- ing analogical reasoning,

    J. N. Chiang, Y . Peng, H. Lu, K. J. Holyoak, and M. M. Monti, “Distributed code for semantic relations predicts neural similarity dur- ing analogical reasoning,”Journal of Cognitive Neuroscience, vol. 33, no. 3, pp. 377–389, 2021

    2021

  33. [33]

    Probabilistic analogical mapping with semantic relation networks,

    H. Lu, N. Ichien, and K. J. Holyoak, “Probabilistic analogical mapping with semantic relation networks,”Psychological Review, 2022

    2022

  34. [34]

    Incremental learning using a grow-and- prune paradigm with efficient neural networks,

    X. Dai, H. Yin, and N. K. Jha, “Incremental learning using a grow-and- prune paradigm with efficient neural networks,”IEEE Transactions on Emerging Topics in Computing, vol. 10, no. 2, pp. 752–762, 2020

    2020

  35. [35]

    Fully dynamic inference with deep neural networks,

    W. Xia, H. Yin, X. Dai, and N. K. Jha, “Fully dynamic inference with deep neural networks,”IEEE Transactions on Emerging Topics in Computing, 2021

    2021

  36. [36]

    Grow and prune compact, fast, and accurate LSTMs,

    X. Dai, H. Yin, and N. K. Jha, “Grow and prune compact, fast, and accurate LSTMs,”IEEE Transactions on Computers, vol. 69, no. 3, pp. 441–452, 2019

    2019

  37. [37]

    NeST: A neural network synthesis tool based on a grow-and- prune paradigm,

    ——, “NeST: A neural network synthesis tool based on a grow-and- prune paradigm,”IEEE Transactions on Computers, vol. 68, no. 10, pp. 1487–1497, 2019

    2019

  38. [38]

    STEERAGE: Synthesis of neu- ral networks using architecture search and grow-and-prune methods,

    S. Hassantabar, X. Dai, and N. K. Jha, “STEERAGE: Synthesis of neu- ral networks using architecture search and grow-and-prune methods,” arXiv preprint arXiv:1912.05831, 2019

    Pith/arXiv arXiv 1912

  39. [39]

    SCouT: synthetic counterfactuals via spatiotemporal transformers for actionable health- care,

    B. Dedhia, R. Balasubramanian, and N. K. Jha, “SCouT: synthetic counterfactuals via spatiotemporal transformers for actionable health- care,”ACM Transactions on Computing for Healthcare, vol. 4, no. 4, pp. 1–28, 2023. 17

    2023

  40. [40]

    Simulation intelligence: Towards a new generation of scientific methods,

    A. Lavin, H. Zenil, B. Paige, D. Krakauer, J. Gottschlich, T. Mattson, A. Anandkumar, S. Choudry, K. Rocki, A. G. Baydinet al., “Simulation intelligence: Towards a new generation of scientific methods,”arXiv preprint arXiv:2112.03235, 2021

    Pith/arXiv arXiv 2021

  41. [41]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  42. [42]

    Using synthetic controls: Feasibility, data requirements, and methodological aspects,

    A. Abadie, “Using synthetic controls: Feasibility, data requirements, and methodological aspects,”Journal of Economic Literature, vol. 59, no. 2, pp. 391–425, 2021

    2021

  43. [43]

    Synthetic interventions,

    A. Agarwal, D. Shah, and D. Shen, “Synthetic interventions,”arXiv preprint arXiv:2006.07691v4, 2020

    Pith/arXiv arXiv 2006

  44. [44]

    Algorithm selection by rational metareasoning as a model of human strategy selection,

    F. Lieder, D. Plunkett, J. B. Hamrick, S. J. Russell, N. Hay, and T. L. Griffiths, “Algorithm selection by rational metareasoning as a model of human strategy selection,” inAdvances in Neural Information Processing Systems, 2014, pp. 2870–2878

    2014

  45. [45]

    When does bounded-optimal metareasoning favor few cognitive systems?

    S. Milli, F. Lieder, and T. L. Griffiths, “When does bounded-optimal metareasoning favor few cognitive systems?” inProc. AAAI Conference on Artificial Intelligence, S. P. Singh and S. Markovitch, Eds., vol. 31. Palo Alto, CA: AAAI Press, 2017

    2017

  46. [46]

    Learning to select computations,

    F. Callaway, S. Gul, P. Kruegera, T. L. Griffiths, and F. Lieder, “Learning to select computations,” inProc. Conference on Uncertainty in Artificial Intelligence, 2018

    2018

  47. [47]

    Rational use of cognitive resources in human planning,

    F. Callaway, B. van Opheusden, S. Gul, P. Das, P. M. Krueger, F. Lieder, and T. L. Griffiths, “Rational use of cognitive resources in human planning,”Nature Human Behaviour, pp. 1–14, Apr. 2022

    2022

  48. [48]

    An automatic method for discovering rational heuristics for risky choice

    F. Lieder, P. M. Krueger, and T. L. Griffiths, “An automatic method for discovering rational heuristics for risky choice.” inProc. Annual Meeting of the Cognitive Science Society, 2017

    2017

  49. [49]

    Burn-in, bias, and the ra- tionality of anchoring,

    F. Lieder, T. Griffiths, and N. Goodman, “Burn-in, bias, and the ra- tionality of anchoring,” inAdvances in Neural Information Processing Systems 25, 2012, pp. 2690–2798

    2012

  50. [50]

    S. J. Russell and E. Wefald,Do the Right Thing: Studies in Limited Rationality. Cambridge, MA: MIT Press, 1991

    1991

  51. [51]

    Principles of metareasoning,

    ——, “Principles of metareasoning,”Artificial Intelligence, vol. 49, no. 1-3, pp. 361–395, 1991

    1991

  52. [52]

    Rationality and intelligence,

    S. J. Russell, “Rationality and intelligence,”Artificial Intelligence, vol. 94, no. 1-2, pp. 57–77, 1997

    1997

  53. [53]

    Computational tradeoffs under bounded resources,

    E. J. Horvitz and S. Zilberstein, “Computational tradeoffs under bounded resources,”Artificial Intelligence, vol. 126, no. 1-2, pp. 1– 4, 2001

    2001

  54. [54]

    Provably bounded-optimal agents,

    S. J. Russell and D. Subramanian, “Provably bounded-optimal agents,” Journal of Artificial Intelligence Research, vol. 2, pp. 575–609, 1994

    1994

  55. [55]

    Selecting computa- tions: Theory and applications,

    N. Hay, S. J. Russell, D. Tolpin, and S. Shimony, “Selecting computa- tions: Theory and applications,” inProc. Conference on Uncertainty in Artificial Intelligence, N. de Freitas and K. Murphy, Eds. Corvallis, OR: AUAI Press, 2012

    2012

  56. [56]

    Using goal-driven deep learning models to understand sensory cortex,

    D. L. Yamins and J. J. DiCarlo, “Using goal-driven deep learning models to understand sensory cortex,”Nature Neuroscience, vol. 19, no. 3, pp. 356–365, 2016

    2016

  57. [57]

    A back-propagation programmed network that simulates response properties of a subset of posterior parietal neurons,

    D. Zipser and R. A. Andersen, “A back-propagation programmed network that simulates response properties of a subset of posterior parietal neurons,”Nature, vol. 331, no. 6158, pp. 679–684, 1988

    1988

  58. [58]

    Vector-based navigation using grid-like representations in artificial agents,

    A. Banino, C. Barry, B. Uria, C. Blundell, T. Lillicrap, P. Mirowski, A. Pritzel, M. J. Chadwick, T. Degris, J. Modayilet al., “Vector-based navigation using grid-like representations in artificial agents,”Nature, vol. 557, no. 7705, pp. 429–433, 2018

    2018

  59. [59]

    The parallel distributed processing approach to semantic cognition,

    J. L. McClelland and T. T. Rogers, “The parallel distributed processing approach to semantic cognition,”Nature Reviews Neuroscience, vol. 4, no. 4, pp. 310–322, 2003

    2003

  60. [60]

    A spiking network model of short-term active memory,

    D. Zipser, B. Kehoe, G. Littlewort, and J. Fuster, “A spiking network model of short-term active memory,”Journal of Neuroscience, vol. 13, no. 8, pp. 3406–3420, 1993

    1993

  61. [61]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  62. [62]

    On the control of automatic processes: A parallel distributed processing account of the Stroop effect

    J. D. Cohen, K. Dunbar, and J. L. McClelland, “On the control of automatic processes: A parallel distributed processing account of the Stroop effect.”Psychological Review, vol. 97, no. 3, p. 332, 1990

    1990

  63. [63]

    An integrative theory of prefrontal cortex function,

    E. K. Miller and J. D. Cohen, “An integrative theory of prefrontal cortex function,”Annual Review of Neuroscience, vol. 24, no. 1, pp. 167–202, 2001

    2001

  64. [64]

    Interactions between frontal cortex and basal ganglia in working memory: A computational model,

    M. J. Frank, B. Loughry, and R. C. O’Reilly, “Interactions between frontal cortex and basal ganglia in working memory: A computational model,”Cognitive, Affective, & Behavioral Neuroscience, vol. 1, no. 2, pp. 137–160, 2001

    2001

  65. [65]

    Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory,

    J. L. McClelland, B. L. McNaughton, and R. C. O’Reilly, “Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory,”Psychological Review, vol. 102, no. 3, p. 419, 1995

    1995

  66. [66]

    What learning systems do intelligent agents need? Complementary learning systems theory updated,

    D. Kumaran, D. Hassabis, and J. L. McClelland, “What learning systems do intelligent agents need? Complementary learning systems theory updated,”Trends in Cognitive Sciences, vol. 20, no. 7, pp. 512– 534, 2016

    2016

  67. [67]

    Neural turing machines,

    A. Graves, G. Wayne, and I. Danihelka, “Neural turing machines,” arXiv preprint arXiv:1410.5401, 2014

    Pith/arXiv arXiv 2014

  68. [68]

    Neural episodic control,

    A. Pritzel, B. Uria, S. Srinivasan, A. P. Badia, O. Vinyals, D. Hassabis, D. Wierstra, and C. Blundell, “Neural episodic control,” inProc. International Conference on Machine Learning. PMLR, 2017, pp. 2827–2836

    2017

  69. [69]

    Learning representations that support extrapolation,

    T. Webb, Z. Dulberg, S. Frankland, A. Petrov, R. O’Reilly, and J. Cohen, “Learning representations that support extrapolation,” in Proc. International Conference on Machine Learning. PMLR, 2020, pp. 10 136–10 146

    2020

  70. [70]

    Modelling the development of counting with memory-augmented neural networks,

    Z. Dulberg, T. Webb, and J. Cohen, “Modelling the development of counting with memory-augmented neural networks,”arXiv preprint arXiv:2105.10577, 2021

    Pith/arXiv arXiv 2021

  71. [71]

    Emergent symbols through binding in external memory,

    T. W. Webb, I. Sinha, and J. D. Cohen, “Emergent symbols through binding in external memory,”arXiv preprint arXiv:2012.14601, 2021

    Pith/arXiv arXiv 2012

  72. [72]

    An introduction to variational autoen- coders,

    D. P. Kingma and M. Welling, “An introduction to variational autoen- coders,”arXiv preprint arXiv:1906.02691, 2019

    Pith/arXiv arXiv 1906

  73. [73]

    Unsupervised representation learning with deep convolutional generative adversarial networks,

    A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,” arXiv preprint arXiv:1511.06434, 2015

    Pith/arXiv arXiv 2015

  74. [74]

    Normalization as a canonical neural computation,

    M. Carandini and D. J. Heeger, “Normalization as a canonical neural computation,”Nature Reviews Neuroscience, vol. 13, no. 1, pp. 51–62, 2012

    2012

  75. [75]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,”arXiv preprint arXiv:1607.06450, 2016

    Pith/arXiv arXiv 2016

  76. [76]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inInternational Conference on Machine Learning. PMLR, 2015, pp. 448–456

    2015

  77. [77]

    The proper treatment of symbols in a connectionist architecture. in, E. Deitrich and A. Markman, (Eds.), Cognitive dynamics: Conceptual Change in Human and Machines,

    K. Holyoak and J. Hummel, “The proper treatment of symbols in a connectionist architecture. in, E. Deitrich and A. Markman, (Eds.), Cognitive dynamics: Conceptual Change in Human and Machines,” 2000

    2000

  78. [78]

    Indirection and symbol-like processing in the prefrontal cortex and basal ganglia,

    T. Kriete, D. C. Noelle, J. D. Cohen, and R. C. O’Reilly, “Indirection and symbol-like processing in the prefrontal cortex and basal ganglia,” Proceedings of the National Academy of Sciences, vol. 110, no. 41, pp. 16 390–16 395, 2013

    2013

  79. [79]

    Marcus,The algebraic mind

    G. Marcus,The algebraic mind. Cambridge, MA: MIT Press, 2001

    2001

  80. [80]

    Holographic reduced representations,

    T. Plate, “Holographic reduced representations,”IEEE Transactions on Neural Networks, vol. 6, no. 3, pp. 623–641, 1995

    1995

Showing first 80 references.