Pith. sign in

REVIEW 1 major objections 6 minor 60 references

High-Order Deep Meta-Learning with Category-Theoretic Interpretation

T0 review · 1 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A recursive stack of meta-learners, each generating virtual tasks and soft constraints for the level below, lets neural networks construct, solve, and generalise across hierarchies of tasks.

desk verdict A speculative framework paper with a genuinely interesting recursive architecture but no validation; reject as-is, though the core idea is worth revisiting after real experiments and a tightened formalization. read the letter →

arxiv 2507.02634 v1 pith:COYXDLJR submitted 2025-07-03 cs.LG

David H. Mguni This is my paper · ORCID
classification cs.LG
keywords recursivemeta-learninghigher-ordervirtualtaskssoftconstraintsphysics-informedneuralnetworkscategorytheorycurriculumlearningadversarialgeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes a recursive meta-learning architecture in which each learner trains the learner below it: a meta-learner produces virtual tasks, soft constraints, and evaluation signals for the level beneath, and is itself trained by a higher level in the same way. The central claim is that this stack lets a neural network construct, solve, and generalise across hierarchies of tasks without relying solely on human-generated data, because virtual tasks force lower levels to confront exactly the regions where they are weak. The category-theoretic reading treats learners as functors between task and model categories and meta-learners as transformations between such functors, and is offered as a unifying description of existing meta-learning, data augmentation, meta-regularisation, and curriculum learning. If the claim holds, the framework gives a principled template for autonomous curriculum construction and higher-order knowledge transfer, in which models generate their own instructive problems.

What carries the argument

The central object is the recursive functor stack together with a virtual-task generator. Level 0 is an ordinary parameterised model; level 1's meta-learner is a functor mapping tasks to models, level 2's maps functors to functors, and so on. At each level a soft-constraint model $C_{\mathrm{soft}}(\phi_k)$ emits virtual tasks, penalised in the loss like collocation points in physics-informed neural networks (PINNs), and an adversarial generator $G_\phi$ maximises an exploration score $S_{\mathrm{explore}}$ to find tasks the lower learner finds difficult while a discriminator enforces plausibility. The exploration score rewards sharp difficulty transitions near a reference task or surprisingly stable performance across distant tasks, which is what drives discovery of constraint boundaries.

What would settle it

Run two identical recursive meta-learning stacks on the same real-task distribution, one selecting virtual tasks by the exploration score and the other using random virtual tasks or none, and compare held-out performance across new task families; if the exploration-driven stack does not generalise better, the claim that difficulty-maximising virtual tasks drive constraint discovery and regularisation is falsified.

Watch

Extended reading notes

Core claim

The paper's core claim is that recursively nesting meta-learners, where each level maps lower-level learners to improved learners and simultaneously generates virtual tasks and soft constraints for them, realises learning-to-learn at arbitrary order. Concretely, Algorithm 2 couples a lower-level learner with a generator that proposes synthetic tasks to maximise the learner's difficulty and a discriminator that keeps those tasks on a plausible manifold, and the meta-loss combines the real-task loss with a virtual-task penalty. The paper argues that this loop lets the meta-learner discover constraint-region boundaries and novel tasks, that recursive promotion of each meta-learner to learner creates an abstraction curriculum, and that the categorical formalisation shows a learner is determined by its behaviour across tasks, in the spirit of the Yoneda lemma, which says an object is captured by how all other objects map into it. It further claims that established methods, including standard meta-learning algorithms such as MAML and Reptile, consistency regularisation methods such as SimCLR and FixMatch, meta-regularisation, and curriculum learning, are degenerate cases of the framework.

Load-bearing premise

The load-bearing premise is that virtual tasks chosen to be hardest for the lower-level learner will improve its generalisation to real, unseen tasks; the paper asserts this regularisation effect but supplies no theoretical guarantee and no experiment for it.

Editorial extensions

If this is right

  • At $K=1$ with no virtual tasks the framework reduces to standard meta-learning such as MAML and Reptile; with virtual tasks it covers augmentation and consistency methods such as SimCLR and FixMatch; at $K=2$ with learned constraints it covers meta-regularisation; pure recursion without virtual tasks is curriculum learning.
  • Because each meta-learner is promoted to a learner at the next level, the hierarchy implements an abstraction curriculum: level 0 solves task instances, level 1 generalises across tasks within a domain, and level 2 generalises across families of domains.
  • The whole stack composes into a single functor $L_K \circ \cdots \circ L_1$ from the task category to the model category, so higher-order meta-learning can be studied as fixed points of functorial transformations and as transformations between functor categories.
  • The architecture is implementable with ordinary differentiable components, including hypernetworks, generative models, and adversarial generator-discriminator pairs, with nested optimisation carried out by automatic differentiation.
  • In game-theoretic settings the hierarchy can learn equilibrium computation across team, coordination, and potential games, and can generate virtual games that reveal where a base algorithm's convergence behaviour changes qualitatively.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the same loop could be tested as a data-efficiency recipe: replace hand-scheduled curricula with generator-chosen virtual tasks and compare held-out performance on a fixed real-task distribution, which the paper does not do.
  • Beyond the paper, if behaviour-equivalent learners are genuinely interchangeable as the Yoneda-style reasoning suggests, probing a large pretrained model with a diverse bank of probe tasks could predict which fine-tuning strategies will work without any access to its internal weights.
  • Beyond the paper, a clean controlled experiment would hold the real tasks fixed and vary only the exploration objective, for example predictive entropy versus loss-gradient norm versus the proposed contextual score, and check whether the difficulty-maximising choice changes which constraints emerge.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 6 minor

Summary. The manuscript proposes a hierarchical deep meta-learning framework in which each meta-level generates virtual tasks (synthetic points with PINN-style soft penalties) that probe and regularize the learner at the level below, and in which meta-learners are recursively promoted to become learners at the next level. Section 2 specifies the architecture, an exploration criterion (Eq. 1) for generating 'difficult' virtual tasks, and three algorithms (Algorithms 1-3). Section 3 offers a category-theoretic reading, calling learners functors and meta-learners variously natural transformations, functors, and endofunctors, and invoking the Yoneda lemma to argue that learners are characterized by their external behavior and that the framework 'unifies existing meta-learning models'. Section 5 instantiates the ideas on game theory (Nash equilibria over a team/coordination/potential-game hierarchy) and reinforcement learning. The paper contains no theorems and no experiments; the claims of improved generalization, regularization, and unification are asserted rather than established.

Significance. If the framework performed as claimed, the contribution would be significant: recursive higher-order meta-learning with self-generated curricula, a PINN-style soft-constraint mechanism, and a compositional categorical lens is a worthwhile research direction, and the paper's worked examples (games, ODE families, policy-update objectives) indicate plausible use cases. The manuscript is candid in labeling some passages as speculative, and Algorithm 2's adversarial exploration mechanism (Eqs. 1-2) is concrete enough to implement. However, the paper ships no code, no empirical evaluation, no theorem statements, and no tested prediction; the categorical formalization is internally inconsistent (Section 3); and the claimed unification of MAML, Reptile, SimCLR, FixMatch, and curriculum learning is asserted without derivation. The current significance is therefore prospective: the paper outlines a research program rather than establishing results.

major comments (1)
  1. [Section 2.2, Eq. (1), Algorithms 1-3.] Section 5 (Degenerate Cases): the Abstract's claim that the framework 'unifies existing meta-learning models' is not supported by the examples given. For example, the paper states that with recursion depth K = 1 and no virtual tasks 'our framework reduces to standard meta-learning algorithms such as MAML or Reptile', but no derivation shows that Algorithm 1's K = 1 procedure (sampling a learner from meta-learner parameters, computing a composite loss, and updating the meta-learner) yields the MAML objective, which involves differentiating through inner-loop gradient steps on an initialization; the two are different mechanisms. The claims that SimCLR and FixMatch are recovered as 'rudimentary virtual task generation' and that curriculum learning corresponds to recursion without virtual tasks are analogies, not reductions: none of these methods' objectives is shown to arise as a special case of the loss Lk in Section 2.1. Since unification is a stated central outcome, the paper should either derive each claimed reduction explicitly (specializing Lk and showing equality, up to constants, with the target method's objective) or substantially weaken the claim.
minor comments (6)
  1. [Section 3; Section 2.2.] Section 3 contains the typo 'cateory' for 'category', and Section 2.2 (after Eq. (1)) contains 'are are thresholds'; both should be corrected.
  2. [Section 2.1; Algorithm 1.] The meta-loss is written in three mutually inconsistent forms: Section 2.1 first gives Lk(xi_k) = E_T[L_{k-1}(F_k(F_{k-1}), T)] + lambda*E_tildeT[L_virtual(F_k(F_{k-1}), tildeT)], then gives Lk = E_{T_k}[L_task(f*_{k-1}, T_{k-1}) + lambda*L_virtual(f*_{k-1}, tildeT_{k-1})], and Algorithm 1 writes Lk = L_task + lambda*L_virtual; the arguments (learner vs meta-learner vs base model) and level indices (k vs k-1) should be unified.
  3. [Eq. (1); Section 2.1.] The symbol lambda denotes two different quantities: the balance between real and virtual loss in Section 2.1 and the balance between the two exploration terms in Eq. (1); renaming one of them is needed to avoid ambiguity.
  4. [Section 2.2; Algorithm 2.] Section 2.2 states that the generator G_phi is 'conditioned on historical training feedback', but in Algorithm 2 the generator receives only latent codes z and no conditioning variable; the conditioning mechanism is never specified.
  5. [Section 5; Figure 2; References.] Section 5 refers to 'Figure 2.1', which does not match the figure numbering (the architecture figure is Figure 2, whose caption contains 'Adapts Metak-Learner to Domain'); the reference list contains malformed entries '[F AL17]' and '[KR W24]'.
  6. [Section 1; References.] The paper has no related-work section: prior higher-order meta-learning [BGCK00] and the surveys [VD02, HAMS21, HVRP21] are cited, but the claimed generalization over existing hierarchical meta-learning formulations is never argued against them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an unvalidated proposal whose claims rest on asserted benefits, not on a derivation that reduces to its own inputs.

full rationale

I found no circular step in the derivation chain. The paper proposes a recursive meta-learning architecture: Algorithm 1 defines a multi-level loss with virtual-task penalties, Section 2.2 and Eq. (1) define an exploration score for generating difficult virtual tasks, Eq. (2) defines the generator objective, and Appendix A smooths the score. None of these steps yields a conclusion that was already assumed as an input. No parameter is fitted to data and then reported as a prediction; no theorem is imported from the author's prior work; and there are no load-bearing self-citations. The central claim that virtual tasks chosen for high difficulty improve held-out generalization (Section 2.2) is asserted without proof or experiment, but that is an absence of validation, not circularity: the benefit is not obtained by definitionally equating 'difficult' with 'useful for generalization.' Similarly, the 'unifies existing meta-learning models' claim is supported only by the Degenerate Cases section, which maps MAML, Reptile, SimCLR, FixMatch, and curriculum learning to special settings of the proposed recursion. That is an interpretive relabeling and an overclaim, but it is not a circular derivation: no result is shown to be equivalent to its own premise. The paper's weaknesses are evidential and formal, not circular.

Assumptions & free parameters 3 free parameters · 4 assumptions · 3 invented entities

The framework relies on three clusters of assumptions: that learned networks satisfy functorial composition, that virtual-difficulty exploration improves real generalization, and that task-space metrics exist. None is demonstrated. The entities introduced by the paper are conceptual constructs with no independent empirical support.

free parameters (3)
  • lambda (constraint weight)
    Appears in the composite meta-loss Lk = Ltask + lambda Lvirtual; chosen by hand or tuned; no sensitivity analysis.
  • beta, gamma (generator objective weights)
    Eq. (2) balances exploration, manifold penalty, and discriminator; values unspecified.
  • alpha1, alpha2, epsilon1, epsilon2, delta (exploration thresholds)
    Eq. (1) and its smooth variants; thresholds suppress uninformative tasks; values unspecified.
assumptions (4)
  • ad hoc to paper Neural learners and meta-learners can be modeled as functors preserving task composition.
    Sections 2 and 3 treat L: T -> M and Fk: Fun(T,M) -> Fun(T,M) as functors without verifying functoriality (composition/identity preservation) for learned networks.
  • domain assumption Maximally difficult virtual tasks improve real-task generalization.
    Section 2.2: the generator maximizes Lexplore and the meta-loss includes lambda Lvirtual; no proof that this transfers to unseen real tasks.
  • domain assumption A meaningful task-space distance d and kernel K exist and can be computed.
    Exploration scores Eq. (1) and Appendix A depend on d(tau, tau') and Kref; the paper leaves their construction open, e.g., 'learned task embedding space'.
  • ad hoc to paper Yoneda lemma transfers to learners: external behavior across tasks fully characterizes a learner.
    Section 3 invokes Yoneda to justify probing black-box learners; the categorical isomorphism is true as math but its analogical application to neural learners is assumed, not established.
invented entities (3)
  • Virtual tasks / virtual points
    purpose: Synthetic problem instances generated by meta-learners to probe lower-level learners and enforce soft constraints.
    Defined in Section 2; no external falsifiable handle outside the framework, and no experiment demonstrates their value.
  • Soft constraint model Csoft(phi_k)
    purpose: Learned distribution over constraints and virtual tasks at each meta-level.
    Introduced in Section 2.1; not grounded in external data or measurable predictions.
  • Virtual point manifold V
    purpose: Space of candidate virtual tasks that the generator explores.
    Section 2.2; purely conceptual construct without independent empirical support.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Order Deep Meta-Learning with Category-Theoretic Interpretation." pith.science (2026). https://pith.science/paper/COYXDLJR

@misc{pith2026250702634,
  author       = {Pith},
  title        = {Pith review of: High-Order Deep Meta-Learning with Category-Theoretic Interpretation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/COYXDLJR}},
  note         = {Machine review of arXiv:2507.02634}
}
read the original abstract

We introduce a new hierarchical deep learning framework for recursive higher-order meta-learning that enables neural networks (NNs) to construct, solve, and generalise across hierarchies of tasks. Central to this approach is a generative mechanism that creates \emph{virtual tasks} -- synthetic problem instances designed to enable the meta-learner to learn \emph{soft constraints} and unknown generalisable rules across related tasks. Crucially, this enables the framework to generate its own informative, task-grounded datasets thereby freeing machine learning (ML) training from the limitations of relying entirely on human-generated data. By actively exploring the virtual point landscape and seeking out tasks lower-level learners find difficult, the meta-learner iteratively refines constraint regions. This enhances inductive biases, regularises the adaptation process, and produces novel, unanticipated tasks and constraints required for generalisation. Each meta-level of the hierarchy corresponds to a progressively abstracted generalisation of problems solved at lower levels, enabling a structured and interpretable learning progression. By interpreting meta-learners as category-theoretic \emph{functors} that generate and condition a hierarchy of subordinate learners, we establish a compositional structure that supports abstraction and knowledge transfer across progressively generalised tasks. The category-theoretic perspective unifies existing meta-learning models and reveals how learning processes can be transformed and compared through functorial relationships, while offering practical design principles for structuring meta-learning. We speculate this architecture may underpin the next generation of NNs capable of autonomously generating novel, instructive tasks and their solutions, thereby advancing ML towards general artificial intelligence.

Figures

Figures reproduced from arXiv: 2507.02634 by the authors.

Figure 1
Figure 1. Recursive higher-order meta-learning pipeline with PINN-inspired soft constraints and virtual data points. Each meta-learner is promoted to a learner at the next level. Soft constraints act via virtual tasks at each level. 2.1 Architecture In practice, this framework can be implemented using standard deep learning tools such as PyTorch [IPK21] or JAX [BFH+21]. Each learner and meta-learner can be parameterised as a … view at source ↗
Figure 2
Figure 2. Architecture of the hierarchical meta-learning framework. Each level corresponds to a progressively higher abstraction over the learning process. The base learner adapts to individual task instances. The meta-learner generalises across tasks within a domain. The meta-meta-learner abstracts across families of domains, learning how to adapt the meta-learning process itself. This constitutes an abstraction curriculum, … view at source ↗
Figure 3
Figure 3. Recursive higher-order meta-learning pipeline with PINN-inspired soft constraints. Each meta-learner is promoted to learner at next level. The meta-learners generate virtual points to regularise lower levels. The soft constraints act via virtual tasks at each level. desirable behaviours of lower-level learners Lk−1. The soft constraint model allows the meta-learner to generate virtual datasets T˜ k−1 = {(xi , yi)} N… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The virtual point landscape. The generative model can be used to explore the landscape and identify subregions where the soft constraints are valid and discover new soft constraints. tion Csoft. This mechanism is analogous to virtual data points in physics-informed neu…
Figure 5
Figure 5. Figure 5: Categorical view of the hierarchical meta-learning framework as a multi-level functor stack. Each level corresponds to functors between categories representing tasks, learning algo￾rithms, constraints, and meta-problems. Dashed arrows indicate meta-meta functors mappin…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 52 canonical work pages

  1. [1]

    Diffusion-convolutional neural networks

    James Atwood and Don Towsley. Diffusion-convolutional neural networks. Advances in neural information processing systems , 29, 2016

    work page 2016

  2. [2]

    Category theory , volume 52

    Steve Awodey. Category theory , volume 52. OUP Oxford, 2010

    work page 2010

  3. [3]

    Bishop and Robert H

    Richard C. Bishop and Robert H. Dorf. Modern control systems . Addison-Wesley Longman Publishing Co., Inc., United States, 2011

    work page 2011

  4. [4]

    Jax: Autograd and xla

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, et al. Jax: Autograd and xla. Astrophysics Source Code Library , pages ascl--2111, 2021

    work page 2021

  5. [5]

    A higher-order approach to meta-learning

    Hilan Bensusan, Christophe G Giraud-Carrier, and Claire Julia Kennedy. A higher-order approach to meta-learning. ILP Work-in-progress reports , 35:44, 2000

    work page 2000

  6. [6]

    A general framework for equivariant neural networks on reductive lie groups

    Ilyes Batatia, Mario Geiger, Jose Munoz, Tess Smidt, Lior Silberman, and Christoph Ortner. A general framework for equivariant neural networks on reductive lie groups. Advances in Neural Information Processing Systems , 36:55260--55284, 2023

    work page 2023

  7. [7]

    Curriculum learning

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning , pages 41--48, 2009

    work page 2009

  8. [8]

    The mechanics of n-player differentiable games

    David Balduzzi, Sebastien Racaniere, James Martens, Jakob Foerster, Karl Tuyls, and Thore Graepel. The mechanics of n-player differentiable games. In International Conference on Machine Learning , pages 354--363. PMLR, 2018

    work page 2018

Show all 60 references
  1. [9]

    Metareg: Towards domain generalization using meta-regularization

    Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using meta-regularization. Advances in neural information processing systems , 31, 2018

    2018

  2. [10]

    Scientific machine learning through physics--informed neural networks: Where we are and what’s next

    Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics--informed neural networks: Where we are and what’s next. Journal of Scientific Computing , 92(3):88, 2022

    2022

  3. [11]

    Settling the complexity of computing two-player nash equilibria

    Xi Chen, Xiaotie Deng, and Shang-Hua Teng. Settling the complexity of computing two-player nash equilibria. J. ACM , 56(3), May 2009

    2009

  4. [12]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning , pages 1597--1607. PmLR, 2020

    2020

  5. [13]

    Variational autoencoder

    Pinheiro Cinelli, Lucas, Matheus Ara \'u jo Marins, Eduardo Ant \'u nio Barros da Silva, and S \'e rgio Lima Netto. Variational autoencoder. In Variational methods for machine learning with applications to deep networks , pages 111--149. Springer, 2021

    2021

  6. [14]

    Group equivariant convolutional networks

    Taco Cohen and Max Welling. Group equivariant convolutional networks. In Maria Florina Balcan and Kilian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning , volume 48 of Proceedings of Machine Learning Research , pages 2990--2999, New...

    2016

  7. [15]

    Meta-learning to compositionally generalize

    Henry Conklin, Bailin Wang, Kenny Smith, and Ivan Titov. Meta-learning to compositionally generalize. arXiv preprint arXiv:2106.04252 , 2021

    2021 arXiv

  8. [16]

    Model-agnostic meta-learning for fast adaptation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning , pages 1126--1135. PMLR, 2017

    2017

  9. [17]

    Communication, coordination and nash equilibrium

    Joseph Farrell. Communication, coordination and nash equilibrium. Economics Letters , 27(3):209--214, 1988

    1988

  10. [18]

    Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data

    Marc Finzi, Samuel Stanton, Pavel Izmailov, and Andrew Gordon Wilson. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In International Conference on Machine Learning , pages 3165--3176. PMLR, 2020

    2020

  11. [19]

    Automated curriculum learning for neural networks

    Alex Graves, Marc G Bellemare, Jacob Menick, Remi Munos, and Koray Kavukcuoglu. Automated curriculum learning for neural networks. In international conference on machine learning , pages 1311--1320. Pmlr, 2017

    2017

  12. [20]

    Long short-term memory

    Alex Graves and Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks , pages 37--45, 2012

    2012

  13. [21]

    Self-awareness and the emergence of mind in primates

    Gordon G Gallup Jr. Self-awareness and the emergence of mind in primates. American Journal of Primatology , 2(3):237--248, 1982

    1982

  14. [22]

    Self-awareness and the evolution of social intelligence

    Gordon G Gallup Jr. Self-awareness and the evolution of social intelligence. Behavioural Processes , 42(2-3):239--247, 1998

    1998

  15. [23]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM , 63(11):139--144, 2020

    2020

  16. [24]

    Lie groups, lie algebras, and representations

    Brian C Hall. Lie groups, lie algebras, and representations. In Quantum Theory for Mathematicians , pages 333--366. Springer, 2013

    2013

  17. [25]

    Meta-learning in neural networks: A survey

    Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence , 44(9):5149--5169, 2021

    2021

  18. [26]

    A survey of deep meta-learning

    Mike Huisman, Jan N Van Rijn, and Aske Plaat. A survey of deep meta-learning. Artificial Intelligence Review , 54(6):4483--4541, 2021

    2021

  19. [27]

    On the power of curriculum learning in training deep networks

    Guy Hacohen and Daphna Weinshall. On the power of curriculum learning in training deep networks. In International conference on machine learning , pages 2535--2544. PMLR, 2019

    2019

  20. [28]

    Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. Programming with TensorFlow: solution for edge computing applications , pages 87--104, 2021

    2021

  21. [29]

    Integers, polynomials, and rings: a course in algebra

    Ronald S Irving. Integers, polynomials, and rings: a course in algebra . Springer, 2004

    2004

  22. [30]

    Physics-informed machine learning

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics , 3(6):422--440, 2021

    2021

  23. [31]

    Formalizing the infinity-categorical yoneda lemma

    Nikolai Kudasov, Emily Riehl, and Jonathan Weinberger. Formalizing the infinity-categorical yoneda lemma. In Proceedings of the 13th ACM SIGPLAN International Conference on Certified Programs and Proofs , pages 274--290, 2024

    2024

  24. [32]

    Categories for the working mathematician

    Saunders Mac Lane. Categories for the working mathematician. Graduate Texts in Mathematics , 1978

    1978

  25. [33]

    Basic category theory , volume 143

    Tom Leinster. Basic category theory , volume 143. Cambridge University Press, 2014

    2014

  26. [34]

    The ai scientist: Towards fully automated open-ended scientific discovery

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist: Towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292 , 2024

    2024 arXiv

  27. [35]

    Robot reinforcement learning on the constraint manifold

    Puze Liu, Davide Tateo, Haitham Bou Ammar, and Jan Peters. Robot reinforcement learning on the constraint manifold. In Conference on Robot Learning , pages 1357--1366. PMLR, 2022

    2022

  28. [36]

    Fictitious play property for games with identical interests

    Dov Monderer and Lloyd S Shapley. Fictitious play property for games with identical interests. Journal of economic theory , 68(1):258--265, 1996

    1996

  29. [37]

    Dov Monderer and Lloyd S. Shapley. Potential games. Games and Economic Behavior , 14(1):124--143, 1996

    1996

  30. [38]

    Reptile: a scalable metalearning algorithm

    Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv preprint arXiv:1803.02999 , 2(3):4, 2018

    2018 arXiv

  31. [39]

    An introduction to convolutional neural networks

    Keiron O'shea and Ryan Nash. An introduction to convolutional neural networks. arXiv preprint arXiv:1511.08458 , 2015

    2015 arXiv

  32. [40]

    A course in game theory

    Martin J Osborne and Ariel Rubinstein. A course in game theory . MIT press, 1994

    1994

  33. [41]

    Yoneda lemma for simplicial spaces

    Nima Rasekh. Yoneda lemma for simplicial spaces. Applied Categorical Structures , 31(4):27, 2023

    2023

  34. [42]

    Metacognition

    Matthew G Rhodes. Metacognition. Teaching of Psychology , 46(2):168--175, 2019

    2019

  35. [43]

    Riedmiller and A

    M. Riedmiller and A. Lernen. Advanced supervised learning in multi-layer perceptrons — from backpropagation to adaptive learning algorithms. Computer Standards & Interfaces , 16(3):265--278, 1994

    1994

  36. [44]

    Karniadakis

    Maziar Raissi, Paris Perdikaris, and George E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics , 378:686--707, 2019

    2019

  37. [45]

    Reinforcement learning

    Richard S Sutton, Andrew G Barto, et al. Reinforcement learning. Journal of Cognitive Neuroscience , 11(1):126--134, 1999

    1999

  38. [46]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems , 33:...

    2020

  39. [47]

    Metacognition

    Norbert Schwarz. Metacognition. 2015

    2015

  40. [48]

    Assessing metacognitive awareness

    Gregory Schraw and Rayne Sperling Dennison. Assessing metacognitive awareness. Contemporary educational psychology , 19(4):460--475, 1994

    1994

  41. [49]

    Metacognitive theories

    Gregory Schraw and David Moshman. Metacognitive theories. Educational psychology review , 7:351--371, 1995

    1995

  42. [50]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 , 2017

    2017 arXiv

  43. [51]

    A perspective view and survey of meta-learning

    Ricardo Vilalta and Youssef Drissi. A perspective view and survey of meta-learning. Artificial intelligence review , 18:77--95, 2002

    2002

  44. [52]

    Relaxing equivariance constraints with non-stationary continuous filters

    Tycho van der Ouderaa, David W Romero, and Mark van der Wilk. Relaxing equivariance constraints with non-stationary continuous filters. Advances in Neural Information Processing Systems , 35:33818--33830, 2022

    2022

  45. [53]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  46. [54]

    Q-learning

    Christopher JCH Watkins and Peter Dayan. Q-learning. Machine learning , 8:279--292, 1992

    1992

  47. [55]

    Wicklund

    Robert A. Wicklund. Objective self-awareness. volume 8 of Advances in Experimental Social Psychology , pages 233--275. Academic Press, 1975

    1975

  48. [56]

    Learning to reinforcement learn

    Jane X Wang, Zeb Kurth-Nelson, Dhruva Tirumala, Hubert Soyer, Joel Z Leibo, Remi Munos, Charles Blundell, Dharshan Kumaran, and Matt Botvinick. Learning to reinforcement learn. arXiv preprint arXiv:1611.05763 , 2016

    2016 arXiv

  49. [57]

    A comprehensive survey of continual learning: Theory, method and application

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

    2024

  50. [58]

    Raissi, M., Perdikaris, P., and Karniadakis, G.E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707

    2019

  51. [59]

    Adámek, J., and Herrlich, H. (1994). Abstract and Concrete Categories: The Joy of Cats. Wiley

    1994

  52. [60]

    Mac Lane, S. (1998). Categories for the Working Mathematician. Springer

    1998

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.