Pith. sign in

REVIEW 3 major objections 6 minor 29 references

Kryptonite-N: Machine Learning Strikes Back

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read This paper claims that the Kryptonite-N datasets, proposed as counterexamples to universal function approximation, are actually high-dimensional XOR problems with one-third redundant features, solvable by logistic regression with…

desk verdict Real empirical rebuttal for N≤18 undercut by an unverified exact-XOR claim and an unsupported 'any N' extrapolation. read the letter →

arxiv 2412.20588 v2 pith:Z7S4CLIM submitted 2024-12-29 cs.LG cs.AI

classification cs.LGcs.AI
keywords Kryptonite-NXORproblemuniversalapproximationlogisticregressionpolynomialbasisexpansionL1regularizationfeatureredundancyhigh-dimensionalclassification
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 argues that the Kryptonite-N challenge datasets, which were proposed as counterexamples to universal function approximation, are actually generated by a simple predictable rule. Each dataset is a high-dimensional XOR (parity) problem over two-thirds of its features: after standardizing, each feature is discretized by thresholding at its mean, and the label is the XOR of the discretized informative features; the remaining one-third of the features are irrelevant. The authors reverse-engineered this construction from N=9, 12, 15, 18, then showed that logistic regression with interaction-only polynomial basis expansion and L1 regularization reaches the target accuracy on those dimensions, while an oracle using the known informative features reaches above-target accuracy up to N=45. The conclusion is that the datasets are not a refutation of universal approximation but a structured benchmark solvable by standard feature-engineered models.

What carries the argument

The load-bearing construction is the parity label $y = \bigoplus_{i \in F_{\mathrm{info}}} \mathbf{1}\{x_i > \mathrm{mean}(x_i)\}$, together with the observation that polynomial basis expansion of the standardized features creates interaction terms that make this XOR linearly separable. L1 regularization then sparsifies the logistic regression weight vector down to the single monomial that encodes the parity product, which is what allows the model to generalize. The paper also uses the probability-mass-function shape of each feature (burst-like, Gaussian-like, spread-like) as a selection oracle to recover the informative feature set when the expansion is infeasible.

What would settle it

Obtain the N=24 (or N=30) Kryptonite-N dataset and either run the L1-regularized polynomial-expansion logistic regression described in Table 2, or compare the hand-coded XOR construction against the true labels. If the XOR oracle accuracy falls below the reported ~0.975 or the per-feature PMF shapes do not fall into the three observed categories, the claim that the datasets are constructed as XOR with exactly 1/3 redundancy is falsified.

Watch

Extended reading notes

Core claim

The core discovery is that every Kryptonite-N dataset is generated by a high-dimensional XOR operator. With N features, exactly $2N/3$ are informative and are drawn from one of three bimodal distributions (burst-like, Gaussian-like, spread-like); the remaining $N/3$ features are irrelevant. Each feature $x_i$ is discretized as $x_i^{\mathrm{dis}} = \mathbf{1}\{x_i > \bar{x}_i\}$, where $\bar{x}_i$ is the feature mean, and the label is $y = \bigoplus_{i \in F_{\mathrm{info}}} x_i^{\mathrm{dis}}$, the XOR (parity) over the informative features. The evidence for this is that an L1-regularized logistic regression on polynomial-expanded features collapses to a single high-order interaction term, the feature shapes match the informative set exactly, and a hand-coded discretize-then-XOR function reproduces labels at or above 0.958 accuracy on N=9 through N=45.

Load-bearing premise

The paper assumes that the generative rule reverse-engineered from N=9, 12, 15, and 18—XOR over the 2N/3 informative features after thresholding each feature at its mean, with exactly N/3 redundant features—applies identically to the larger Kryptonite-N datasets (N=24, 30, 45) where no polynomial-expansion experiment was run.

Editorial extensions

If this is right

  • If the construction holds for every N, Kryptonite-N datasets are solvable in principle by logistic regression with polynomial features and L1 regularization, limited only by memory; the paper reports that it did not run this on N=24, 30, 45 due to RAM, but claims the same performance should follow.
  • The datasets therefore fail as a counterexample to universal function approximation; they become a benchmark for feature engineering and redundancy handling rather than a refutation.
  • Neural networks with one hidden layer (72 neurons) and tuned hyperparameters reach the target accuracy on N=9 through N=18, reinforcing the paper's universal-approximation conclusion.
  • The original GPT-based basis expansion is shown to be flawed in its implementation: it averages causal-masked token embeddings, uses a prompt that asks for classification rather than encoding, and processes each vector in isolation, which the paper likens to fitting noise.
  • L1 regularization is singled out as the right choice for this data because a single interaction deterministically dominates the label, whereas L2 spreads weight across many correlated features.

Reading between the lines

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

  • The 'any N' claim is extrapolated: the polynomial-expansion pipeline was demonstrated only up to N=18, and the oracle/filtering experiments only up to N=45; a direct run on N=24 or N=30 would either confirm the construction or reveal that larger datasets use different parameters.
  • The paper's reverse-engineering recipe—inspect per-feature distributions for bimodality, check for parity-like structure after mean-thresholding, and use L1-selected interaction terms to identify informative features—transfers to other unexplained high-dimensional datasets.
  • If the generative rule is exact, the Kryptonite-N datasets can be regenerated at will, enabling controlled studies of how models degrade as the fraction of redundant features, noise level, or threshold choice changes; the paper does not explore these perturbations.
  • The near-ceiling accuracy of the hand-coded XOR (above 0.958) suggests the construction is essentially noise-free beyond the discretization loss; testing whether the mean threshold is optimal (versus the median or a fitted threshold) would sharpen the recovered generative model.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper claims to refute the Kryptonite-N challenge datasets by showing that they are constructed as high-dimensional XOR problems with exactly 1/3 redundant features, and that logistic regression with sufficient polynomial expansion and L1 regularization can solve them for any dimension N. The authors support this with three lines of evidence: (1) neural networks that exceed the target accuracies for N=9 to N=18, (2) polynomial-expansion logistic regression that exceeds target accuracies for N=9, 12, and 15, and (3) a reverse-engineered generative rule that is then used to construct a handcrafted XOR classifier and a feature-selection-oracle logistic regression for N=9 through N=45.

Significance. If the central claim were fully established, the paper would show that Kryptonite-N is not a fundamental counterexample to machine learning but a structured benchmark solvable by standard feature-engineered models. The empirical demonstrations for N=9-18 are credible and useful, and the identification of three PMF shapes across dimensions is a genuinely interesting observation. However, the paper's own XOR reconstruction does not achieve perfect accuracy, contradicting the exact generative rule asserted in Section 6.1, and the 'any dimension N' claim is supported by experiments only up to N=18 for the main polynomial-logistic-regression method. The theoretical significance is therefore substantially weaker than claimed.

major comments (3)
  1. [Section 6.1 and Table 4] The XOR verification column in Table 4 reports accuracies of 0.9583 (N=9) through 0.9805 (N=45), not 1.0. Under the exact generative rule stated in Section 6.1 (y = XOR over the thresholded 2N/3 informative features, with the remaining 1/3 redundant), applying the handcrafted XOR with the inferred informative-feature set and mean thresholds should recover the true labels exactly on the test set. The 2-4% shortfall is an internal contradiction: the proposed construction does not reproduce the observed labels even on the datasets from which it was inferred. Possible causes include incorrect threshold location, PMF-shape feature-selection errors, unmodeled label noise, or a different redundancy fraction; in any case, the exact-XOR formulation in Section 6.1 is not established. Since the theoretical argument that a degree-(2N/3) polynomial plus L1 regularization solves any dimension N assumes the discretized informative feature set is known exactly, this failure undermines the central claim.
  2. [Section 6.2 and Tables 2-4] The abstract claims logistic regression with polynomial expansion and L1 regularization can 'solve for any dimension N,' but full polynomial logistic regression is demonstrated only for N=9, 12, and 15 in Table 3, with a single additional L1 result for N=18 in Table 4 and no results for N=24, 30, or 45. Table 2 lists hyperparameters only up to N=15, and Section 5.2 states that 'this approach displayed sub-optimal performance for N greater than 15,' which is contradicted by the N=18 entry in Table 4 (0.9708 ± 0.0013). The extrapolation to arbitrary N rests entirely on the unverified exact construction from Section 6.1, so the 'any dimension N' claim is unsupported by the presented experiments.
  3. [Section 6.2, item 3 and Appendix A.2, Step 6] The logistic-regression-with-feature-selection-oracle (LR with FSO) results are not independent evidence for the XOR conjecture, because the feature-selection oracle is derived from the PMF shapes observed on the same datasets (Appendix A.2 Step 6). This oracle is effectively part of the reverse-engineered construction, so the FSO column in Table 4 merely re-tests the construction on the data used to infer it. Only the LR-with-L1 results (N=9, 12, 15) and the neural-network results (N=9-18) are independent empirical evidence, and those cover only the smaller dimensions.
minor comments (6)
  1. [Section 6.1] The definition of the irrelevant feature set is misprinted as 'Firre = Finfo'; it should be the complement, e.g., 'Firre = F \ Finfo.' As written, it contradicts the immediately preceding sentence.
  2. [Abstract] The phrase 'breaking the notation' should be 'breaking the notion'.
  3. [Section 2] The sentence 'the points fall inkk distinct clusters' contains a stray 'kk' and should read 'fall into k distinct clusters.'
  4. [Section 5.2] The cross-reference 'Appendix ??' should be replaced with the actual appendix or figure number for the accuracy distributions.
  5. [Appendix A.1] The word 'floored' in 'the work using GPT as a basis expansion tool is floored' should be 'flawed.'
  6. [Table 4] The N=18 LR-with-L1 result in Table 4 should be reconciled with the statement in Section 5.2 that the approach displayed sub-optimal performance for N greater than 15; if N=18 is included, the hyperparameters used (degree, lambda) should be reported.

Circularity Check

2 steps flagged · score 4.0 of 10

The XOR 'verification' and LR-with-FSO columns are in-sample fits of a rule reverse-engineered from the same K-N datasets; the NN and cross-validated LR results remain independent empirical evidence.

  1. fitted input called prediction [Section 6.2, item 1 (XOR Conjecture Verification); Appendix A.2 Steps 6-7]
    "As a verification of our conjecture, the second column of Table 4 presents the test accuracy on the entire dataset using our handcrafted feature selection, discretization, and XOR operator functions (XOR)."

    The XOR rule (feature set, per-feature mean threshold, and 2N/3 informative features) was reverse-engineered from the same K-N datasets in Appendix A.2: informative features were chosen by PMF shape observed in the data, the threshold is each feature's dataset mean, and the XOR structure was conjectured after inspecting those fits. Reporting 'test accuracy on the entire dataset' for this handcrafted reconstruction is an in-sample evaluation of a rule fitted to that same data, not an independent prediction. The column therefore cannot verify the conjecture; it only measures how well the fitted rule reproduces the labels it was chosen to match. It is not forced by construction to reach 100% (and in fact does not), but it is not independent evidence either.

  2. fitted input called prediction [Section 6.2, item 3 (LR with FSO); Appendix A.2 Steps 6-7]
    "With Feature Selection Oracle (FSO) based on feature distribution shape we observed in Step 6 of Appendix A.2, we reduced the feature space to 1 dimension only including the multiplication of informative features, formulated as follows: input = Q_{i∈Finfo} x^norm_i."

    The FSO is the PMF-shape-based informative-feature selector discovered in Appendix A.2 from the same datasets. Feeding LR the product of exactly the features that this data-fitted oracle labels informative makes the LR-with-FSO accuracy a function of the fitted oracle. The experiment does not test whether LR can discover the structure; it evaluates the discovered structure itself. This column therefore reduces to an in-sample check of the inferred generative model and adds no independent confirmation of the 'any N' solvability claim. It is the fitted selector, renamed as an oracle, being scored on the data it was fit to.

full rationale

The paper's central demonstration has two distinct parts. Genuinely independent evidence exists: the neural-network results in Table 1 use a 60/20/20 train/validation/test split, and the polynomial-expansion logistic-regression results in Table 3 are 10-fold cross-validated scores for N=9,12,15,18. Those experiments do not reduce to a fit of the discovered rule and are real empirical support for the claim that Kryptonite-N is solvable by standard models. The circularity is concentrated in the 'verification' of the dataset secrets. Section 6.1 states an exact mathematical formulation (y = XOR over 2N/3 informative features thresholded at their dataset means), but the formulation was inferred in Appendix A.2 from the same data, and the XOR column of Table 4 is then reported as test accuracy on that same data. Similarly, the LR-with-FSO column uses the PMF-shape oracle that was itself derived from the data, so it is an in-sample evaluation of the inferred rule rather than a predictive test. A further correctness issue, independent of circularity, is that the handcrafted XOR accuracy in Table 4 is only 95.83-98.05%, not 100%, so the proposed exact construction is not even confirmed in-sample; the theoretical extrapolation to N=24,30,45 (Section 6.2, item 2) is therefore unsupported. No load-bearing self-citation chain is present; the citation to Quinn & Luther is the external target being rebutted. Overall, the core empirical results are independent, but the paper's structural 'verification' partially reduces to fitting the rule it claims to confirm, giving a moderate circularity score of 4.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several fitted hyperparameters (polynomial degree, regularization strength) and on the reverse-engineered structural assumptions (2N/3 informative features, XOR label, mean threshold, PMF-shape-based feature selection). These are not derived from first principles; they are inferred from the datasets or chosen by validation.

free parameters (5)
  • Polynomial degree range per dataset = N-9: (5,7), N-12: (8,10), N-15: (10,14)
    Chosen via GridSearchCV on validation data in Section 5.2; the degree range is a fitted hyperparameter that the logistic regression result depends on.
  • Regularization strength lambda = 0.01 (N-9,N-12), 0.015 (N-15)
    Selected by cross-validation; the L1/L2 choice also changes per N (Table 2).
  • Discretization threshold = feature mean X_i (approximately 0.5)
    Used in the reverse-engineered label formula (Section 6.1) and in the FSO model; chosen by hand based on observed distribution, not derived from an external rule.
  • Number of informative features = 2N/3
    Inferred from the observed datasets in Appendix A.2 Step 6; this structural parameter is used in the claimed construction and in the FSO input.
  • Neural network hyperparameters = learning rates 0.01/0.1, batch sizes 128/256, dropout 0/0.01/0.03/0.07, optimizer SGD/Adam
    Tuned per dataset (Table 5); the NN results up to N=18 depend on these choices.
assumptions (5)
  • standard math The Universal Approximation Theorem guarantees that feedforward networks with non-linear activations can approximate any continuous function on a compact set.
    Invoked in Section 1 and 3.2 to justify using neural networks, but the theorem is about existence, not learnability; the cited Kolmogorov paper does not state it.
  • ad hoc to paper The Kryptonite-N datasets are generated by the reverse-engineered rule: XOR over 2N/3 discretized informative features, with 1/3 redundant features.
    Inferred from N=9,12,15,18 in Appendix A.2 and then assumed to hold for all N, including N=24,30,45 (Section 6.1, Table 4).
  • domain assumption Feature values are drawn from distributions centered near 0.5 such that the feature mean is a valid discretization threshold.
    Needed for the discretization step in Section 6.1; if the true generator used a different threshold, the XOR reconstruction would be approximate, consistent with XOR accuracy around 0.96 rather than 1.0.
  • ad hoc to paper The PMF shape of a feature (burst-like vs gaussian-like vs spread-like) identifies whether it is informative or redundant.
    This is the basis of the 'Feature Selection Oracle' in Section 6.2 and Appendix A.2 Step 6; it was inferred from the observed datasets, not derived.
  • domain assumption The test set is drawn from the same distribution as the training set.
    The models' reported generalization depends on this standard i.i.d. assumption, stated in Section 4.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kryptonite-N: Machine Learning Strikes Back." pith.science (2026). https://pith.science/paper/Z7S4CLIM

@misc{pith2026241220588,
  author       = {Pith},
  title        = {Pith review of: Kryptonite-N: Machine Learning Strikes Back},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z7S4CLIM}},
  note         = {Machine review of arXiv:2412.20588}
}
read the original abstract

Quinn et al propose challenge datasets in their work called ``Kryptonite-N". These datasets aim to counter the universal function approximation argument of machine learning, breaking the notation that machine learning can ``approximate any continuous function" \cite{original_paper}. Our work refutes this claim and shows that universal function approximations can be applied successfully; the Kryptonite datasets are constructed predictably, allowing logistic regression with sufficient polynomial expansion and L1 regularization to solve for any dimension N.

Figures

Figures reproduced from arXiv: 2412.20588 by the authors.

Figure 1
Figure 1. PMF Across Dimensions 10 [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Correlation Matrix of Dimensions and Label [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Correlation Matrix of Dimensions and Label [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Dimensions 1 and 2 11 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Cross Validation Accuracy Score Distribution of Logistic Regression Model with varying degrees of Polynomial basis A.6. Neural Network Architectures [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Training vs Validation Loss for Neural Network Trained on N9 Dataset [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Training vs Validation Loss for Neural Network Trained on N12 Dataset 14 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Training vs Validation Loss for Neural Network Trained on N15 Dataset [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Training vs Validation Loss for Neural Network Trained on N18 Dataset A.8. Neural Network Overfitting Comparison [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Neural Network Trained on N12 Dataset with no Dropout [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: shows the results of training the neural network on the N15 dataset with an ELU activation function rather than a Tanh activation function [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Neural Network Trained on N15 Dataset with 0.01 Dropout [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: shows the results of training the neural network on the N18 dataset with 0.03 dropout compared to the original 0.07 dropout. As can be seen, when compared to figure 9, the network shows more overfitting on the training data [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Network Trained on GPT2 Outputs A.10. Sustainability Analysis Graphs [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: shows the difference in emissions produced for training and inference across datasets on an Apple M3 system. (a) Training Emissions (b) Inference Emissions [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]
Figure 16
Figure 16. Figure 16: Energy Consumed During Training for an Apple M3 System vs x86 system (AMD Ryzen 7 5700X) [PITH_FULL_IMAGE:figures/full_fig_p019_16.png]
Figure 17
Figure 17. Figure 17: shows the difference in energy consumption during inference for an Apple M3 system vs an AMD Ryzen 7 5700X (x86) system. (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p019_17.png]
Figure 18
Figure 18. Figure 18: Emissions Produced During Training for an Apple M3 System vs x86 system (AMD Ryzen 7 5700X) [PITH_FULL_IMAGE:figures/full_fig_p020_18.png]
Figure 19
Figure 19. Figure 19: shows the difference in emissions during inference for an Apple M3 system vs an AMD Ryzen 7 5700X (x86) system. (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p020_19.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 14 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Product Environmental Report, MacBook Air with M3 chip

    Apple . Product Environmental Report, MacBook Air with M3 chip. https://www.apple.com/environment/pdf/products/notebooks/M3_MacBook_Air_PER_March2024.pdf . [Accessed 14 Nov 2024]

  3. [3]

    Adaptive Control Processes: A Guided Tour

    Bellman, R. Adaptive Control Processes: A Guided Tour. Princeton University Press, 1961

  4. [4]

    Bishop, C. M. Pattern recognition and machine learning. Information science and statistics. Springer, New York, NY, 2006 - 2006. ISBN 9780387310732

  5. [5]

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A.,...

  6. [6]

    Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H. W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levska...

  7. [7]

    mlco2/codecarbon: v2.7.4, November 2024

    Courty, B., Schmidt, V., Goyal-Kamal, MarionCoutarel, Blanche, L., Feld, B., inimaz, Lecourt, J., LiamConnell, SabAmine, supatomic, Léval, M., LLORET, P., Cruveiller, A., Saboni, A., ouminasara, Zhao, F., Joshi, A., Bauer, C., Bogroff, A., de Lavoreille, H., Laskaris, N., Phiev, A., Abati, E., rosekelly6400, Blank, D., Wang, Z., Otávio, L., and Catovic, A...

  8. [8]

    Llmcarbon: Modeling the end-to-end carbon footprint of large language models, 2024

    Faiz, A., Kaneda, S., Wang, R., Osi, R., Sharma, P., Chen, F., and Jiang, L. Llmcarbon: Modeling the end-to-end carbon footprint of large language models, 2024. URL https://arxiv.org/abs/2309.14393

Show all 29 references
  1. [9]

    Generative deep learning : teaching machines to paint, write, compose, and play

    Foster, D. Generative deep learning : teaching machines to paint, write, compose, and play. 2019

  2. [10]

    Deep Learning

    Goodfellow, I., Bengio, Y., and Courville, A. Deep Learning. MIT Press, 2016

  3. [11]

    Hands-on machine learning with scikit-learn, Keras, and tensorflow: Concepts, tools, and techniques to build Intelligent Systems

    Géron, A. Hands-on machine learning with scikit-learn, Keras, and tensorflow: Concepts, tools, and techniques to build Intelligent Systems. O’Reilly Media, 2019

  4. [12]

    Hoerl, A. E. and Kennard, R. W. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12 0 (1): 0 55--67, 1970

  5. [13]

    Jaynes, E. T. Probability Theory: The Logic of Science. Cambridge University Press, 2003

  6. [14]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization, 2017. URL https://arxiv.org/abs/1412.6980

  7. [15]

    Three approaches to the quantitative definition of information

    Kolmogorov, A. Three approaches to the quantitative definition of information. International Journal of Computer Mathematics, 1 0 (1): 0 3--11, 1965

  8. [16]

    S., Viguier, S., and Ligozat, A.-L

    Luccioni, A. S., Viguier, S., and Ligozat, A.-L. Estimating the carbon footprint of bloom, a 176b parameter language model, 2022. URL https://arxiv.org/abs/2211.02001

  9. [17]

    Power hungry processing: Watts driving the cost of ai deployment? In The 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’24

    Luccioni, S., Jernite, Y., and Strubell, E. Power hungry processing: Watts driving the cost of ai deployment? In The 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’24. ACM, June 2024. doi:10.1145/3630106.3658542. URL http://dx.doi.org/10.1145/3630106.3658542

  10. [18]

    and Papert, S

    Minsky, M. and Papert, S. Perceptrons: An Introduction to Computational Geometry. MIT Press, expanded edition edition, 1988

  11. [19]

    Murphy, K. P. Machine learning : a probabilistic perspective. Adaptive computation and machine learning. The MIT Press, Cambridge, Massachusetts, 2012 - 2012. ISBN 9780262018029

  12. [20]

    C02-Converter

    openco2.net . C02-Converter. https://www.openco2.net/en/co2-converter . [Accessed 14 Nov 2024]

  13. [21]

    Scikit-learn: Machine learning in P ython

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. Scikit-learn: Machine learning in P ython. Journal of ...

  14. [22]

    Early Stopping --- But When?, pp.\ 53--67

    Prechelt, L. Early Stopping --- But When?, pp.\ 53--67. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. doi:10.1007/978-3-642-35289-8_5. URL https://doi.org/10.1007/978-3-642-35289-8_5

  15. [23]

    and Luther, L

    Quinn, H. and Luther, L. Kryptonite-n: A simple end to machine learning hype? ICML 2024, 2024

  16. [24]

    and Narasimhan, K

    Radford, A. and Narasimhan, K. Improving language understanding by generative pre-training. 2018. URL https://api.semanticscholar.org/CorpusID:49313245

  17. [25]

    Language models are unsupervised multitask learners

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. Language models are unsupervised multitask learners. 2019. URL https://api.semanticscholar.org/CorpusID:160025533

  18. [26]

    Sgdclassifier, 2024

    Scikit-Learn, D. Sgdclassifier, 2024. URL https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.SGDClassifier.html. Accessed: 2024-11-18

  19. [27]

    Dropout: A simple way to prevent neural networks from overfitting

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15 0 (56): 0 1929--1958, 2014. URL http://jmlr.org/papers/v15/srivastava14a.html

  20. [28]

    Regression shrinkage and selection via the lasso

    Tibshirani, R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 58 0 (1): 0 267--288, 1996

  21. [29]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pp.\ 6000–6010, Red Hook, NY, USA, 201...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.