acceptodds
Under review as a conference paper at ICLR 2027

Loss-Guided Tsetlin Machines: Making the Learning Objective Explicit

Abstract

The Tsetlin Machine (TM) has no explicit notion of a loss function. Its learning signal is fixed to an implicit, margin-based update rule (MBUR), with no way to choose an objective function suited to a specific task. This paper introduces Loss-Guided TM, which replaces MBUR with an explicit and swappable loss function. In our proposed method, clause feedback and weight updates are driven by the loss, while keeping the Tsetlin Automaton and clause logic unchanged. The approach incorporates cross entropy (CE), symmetric cross entropy (SCE), and asymmetric loss (ASL). We evaluate Loss-Guided TM across six datasets that span multiclass, multi-label, and binary classification tasks. For each dataset, the best-performing Loss-Guided configuration is significantly better calibrated than the MBUR-based one, reducing calibration error by a factor of to . It further improves accuracy and F1 across all datasets. The gains are greatest for the imbalanced multilabel datasets, where ASL improves macro F1 by up to percentage points compared to MBUR. A qualitative inspection shows that the clause-based global interpretability, examined at the per-class level on image and text data, is preserved when using the Loss-Guided mechanism. These results show that our explicit TM learning objective allows direct targeting of uncertainty calibration, data imbalance, and task structure, opening the TM to the broader literature on loss functions.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Tsetlin Machines, Classification, Multilabel Classification, Interpretability, Loss Function, Objective Learning, Stochastic Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2iFoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work targets the inherent limitation of TM that the absence of an explicit loss function and exclusive reliance on the built-in margin-based heuristic update rule and presents Loss-Guided TM. While preserving the original Tsetlin automata and clause logic intact, the proposed framework uses a pluggable explicit loss to drive clause feedback and weight updates, supporting cross-entropy, symmetric cross-entropy, asymmetric loss and other customized loss formulations. Experiments on six datasets spanning binary, multi-class and multi-label tasks show that compared with the original MBUR mechanism, Loss-Guided TM improves accuracy and F1 scores on most tasks, and significantly reduces calibration error, where the asymmetric loss variant is particularly effective for imbalanced multi-label scenarios. Visualization further verifies that the proposed method does not compromise the native interpretability of TM, and can even learn more reasonable patterns with more stable training dynamics. The main limitation is that the current experiments only adopt limited hyperparameter tuning, without large-scale hyperparameter search.

Critical strengths
  • The original MBUR of TM is a fixed heuristic rule that cannot connect to mature loss function research. This paper implements loss-driven updates without modifying the underlying Tsetlin automata and logical clause structure, expanding the capability of TM to adapt to different task objectives such as imbalance handling and calibration.

  • The paper not only demonstrates on multiple types of datasets that Loss-Guided TM can boost model performance, improve confidence calibration and mitigate training degradation issues, but also verifies through global interpretability visualization on image and text tasks that the introduced loss mechanism does not impair the interpretability of clause rules, and in some scenarios can even learn feature patterns that better fit the task semantics.

Critical weaknesses
  • The paper mainly proves the effectiveness of the proposed method through experimental phenomena, but does not theoretically analyze whether the model can still guarantee the stable learning of logical clauses after using loss gradients to drive Tsetlin automata clause updates, nor does it explain how different loss functions affect the update behavior of the automata. As a modification to the core learning rule of TM, the lack of theoretical support will reduce the trust in the reliability of the method.

  • Both MBUR and Loss-Guided TM in the paper are compared using only one set of hyperparameters, without a complete grid search. This makes it impossible to rule out the possibility that the advantages of Loss-Guided TM merely stem from its more adapted hyperparameter configuration.

Questions for authors
  • The loss‑guided mechanism involves both discrete Tsetlin Automaton feedback based on ΔLk\Delta L_k and gradient‑based weight updates for joint training. Could you further elaborate on the coupling relationship between these two components? For instance, can gradient‑driven weight updates alter clause outputs and thereby interfere with automaton feedback signals? Have you observed any conflicting training behaviours between the two parts?

  • Comparisons are conducted only over a limited set of hyperparameters in this work. Without exhaustive grid‑search tuning, how do you rule out the possibility that sufficiently well‑tuned MBUR could narrow or even close the performance gaps in terms of ECE and F‑score against Loss‑Guided TM? Are there any control experiments that support the claim that the observed performance gains stem from the loss‑guided mechanism itself rather than hyperparameter choices?

  • ASL yields substantial F‑score improvements for multi‑label tasks yet produces higher ECE than vanilla MBUR. Could you explain the underlying cause for this trade‑off? Is this phenomenon caused by inherent mismatches between the ASL loss and the probability mapping derived from model voting outputs?

Reviewer_AUP7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that the Tsetlin Machine (TM) lacks an explicit loss function, since its feedback is driven by a fixed margin-based update rule (MBUR, Eq. 2). It proposes Loss-Guided TM, which keeps the Tsetlin Automata and clause logic but changes two parts of training. First, clause feedback is driven by a loss: for each clause, ΔLk\Delta L_k is the change in loss when the clause's output is inverted, the update probability is 1−e−λ∣ΔLk∣1 - e^{-\lambda|\Delta L_k|}, and the sign of ΔLk\Delta L_k together with the clause state selects Type Ia, Ib or II feedback. Second, clause weights become real-valued and are trained by gradient descent on the loss through a softmax or sigmoid over normalized vote sums vi/(C/2)v_i/(C/2). The method is instantiated with cross entropy (CE), symmetric cross entropy (SCE) and asymmetric loss (ASL), and compared with MBUR on MNIST, FashionMNIST, IMDB, AG News, GoEmotions and CelebA over 10 seeds. The authors report lower ECE (3–11× lower), higher accuracy and F1, large macro-F1 gains from ASL on the imbalanced multilabel datasets, and qualitatively preserved or improved global interpretability.

Critical strengths

S1. A simple, general mechanism that makes the loss swappable. Driving clause feedback by the effect of flipping a clause on an arbitrary loss is easy to implement. It keeps the TA learning machinery intact and lets standard losses (CE, SCE, ASL) be plugged into TM training. The ASL results on CelebA and GoEmotions (macro F1 +8.7 and +9.0 points, respectively) suggest this flexibility has practical value for imbalanced multilabel tasks, where TMs have been weak. This matters to my assessment because it is the paper's clearest practical contribution.

S2. Careful statistical reporting across a range of tasks. The evaluation uses 10 paired seeds per condition, Wilcoxon signed-rank tests with Holm–Bonferroni correction and paired effect sizes. It covers multiclass, binary and multilabel tasks on both image and text data, and reports per-class results for the imbalanced datasets. This is more rigorous than much of the TM literature.

Critical weaknesses

W1. The main comparison confounds two changes, and the more conventional change may explain most of the gains.

Loss-Guided TM changes both how clauses receive feedback (loss-driven ΔLk\Delta L_k) and how weights are learned. MBUR uses integer weights updated in ±1\pm 1 steps; Loss-Guided TM uses real-valued weights trained by exact gradient descent on CE, SCE or ASL. Training the output layer with CE through a softmax is logistic regression on clause outputs, and it would be expected on its own to produce much better-calibrated probabilities and smoother learning curves. The paper never separates the two changes. Without ablations such as (a) MBUR clause feedback with gradient-trained real-valued weights and (b) loss-guided clause feedback with integer MBUR-style weights, it is impossible to attribute the calibration, accuracy and stability gains to the paper's actual novelty, the loss-guided clause feedback.

Several further issues weaken the comparison:

  • The calibration comparison is structurally unfair. CE-trained models directly optimize a proper scoring rule over probabilistic outputs, while MBUR is not trained to output probabilities at all. The paper does not say how MBUR confidences are computed (presumably by applying softmax or sigmoid to vote sums with some scaling). A 3–11× ECE reduction is therefore unsurprising and largely uninformative.
  • Calibration baselines are missing. The fair comparison is MBUR plus post-hoc calibration (temperature or Platt scaling on the class sums, fitted on held-out data). The paper should also compare with existing TM probability scores and uncertainty methods, namely Helin et al. (2025), "Uncertainty Quantification in the Tsetlin Machine," and the Probabilistic TM of Kuruge et al. (2024), which reports ECE. Neither is cited.
  • The normalization is unexplained. The fixed C/2C/2 normalization acts as a temperature for both training and evaluation, and its effect on ECE is not studied.
  • The added cost is not measured. Computing ΔLk\Delta L_k requires re-evaluating all MM class outputs, including the softmax, for every clause on every sample. That is O(CM)O(CM) floating-point work per sample, for example 80,000×4080{,}000 \times 40 on CelebA. Since efficiency and bit-level hardware friendliness are core selling points of TMs (Section 1), the lack of any runtime, memory or energy comparison with MBUR is a significant omission. This weakness matters because it determines whether the paper's contribution is the loss-guided TM feedback, or simply training the clause weights with a standard differentiable loss.

W2. The novelty and motivation are overstated, and the baseline may not be well configured.

  • MBUR can already be read as a loss. The paper's premise is that the TM has "no explicit notion of a loss function." But MBUR's update probability ∣yiT−clip(vi,−T,T)∣/2T|y_i T - \mathrm{clip}(v_i,-T,T)|/2T is exactly the normalized residual of a clipped squared-error loss, (yiT−vi)2/4T(y_iT - v_i)^2/4T, between the class sum and a target of ±T\pm T. MBUR is better described as an implicit, fixed loss than as having no loss. The contribution is therefore making the loss swappable, which is narrower than the framing in the title and abstract.
  • Closely related work is not cited. This includes:
    • Sharma et al. (2022), "On the Equivalence of the Weighted Tsetlin Machine and the Perceptron," which relates TM weight updates to gradient updates.
    • Regularized Tsetlin Machine work (2024), which discusses making the TM differentiable via a sigmoid.
    • The Regression TM (Abeyrathna et al.), whose feedback explicitly minimizes an error signal.
    • The TM uncertainty and calibration work mentioned in W1. How Loss-Guided TM differs from these should be discussed explicitly.
  • The new feedback rule has no theory behind it. The mapping Pk=1−e−λ∣ΔLk∣P_k = 1 - e^{-\lambda|\Delta L_k|} is itself a heuristic, and the paper offers no analysis that the resulting stochastic TA feedback actually descends the stated loss. One heuristic is replaced by another, with the loss now as its input. This stochastic update rule might be the primary factor behind the mitigation of overfitting, and few more design ablations such as the same rule with MBUR would help isolate why this is even necessary.
  • MBUR may be misconfigured. In most settings the threshold is at or above the clause count: T=40,000T = 40{,}000 with C=8,000C = 8{,}000 on AG News, T=20,000T = 20{,}000 with C=16,000C = 16{,}000 on FashionMNIST, and T=80,000T = 80{,}000 with C=80,000C = 80{,}000 on CelebA. If vote sums rarely approach TT, MBUR's update probability stays near 0.5 regardless of margin, which removes its margin-based behavior. This could explain the unusual "early peak, then train and test decline together" pattern the authors report for MBUR (Section 6), which they themselves note is not ordinary overfitting. The authors acknowledge that MBUR was not tuned beyond "initial trials," and the Loss-Guided hyperparameters were chosen in the same informal way. Without a systematic, equal-budget search for both methods, the results may reflect an under-tuned baseline.
  • The IMDB sweep is internally inconsistent. Configurations v6, v8 and v12 in Table 4 have identical hyperparameters, yet MBUR's accuracy differs (0.862, 0.856, 0.855). The reported configuration, v12, is MBUR's worst of the three. The main-text MBUR result (85.52 ± 0.12) is about six standard deviations below v6's 86.2. Either the reported variance underestimates run-to-run variability, or the configurations differ in ways not described. Furthermore, Table 4 shows IMDB learning rate was 0.001, while Table 1 says 0.01 which is conflicting.

Additional Concerns

  • W3. The claims do not match the tables.
    • The abstract says Loss-Guided TM "improves accuracy and F1 across all datasets." On MNIST, CE has lower accuracy and F1 than MBUR (98.55 vs. 98.75), and on GoEmotion, CE has lower macro F1 (26.38 vs. 28.09).
    • The "best-performing configuration" is better calibrated only if "best" is chosen per metric. On both multilabel datasets, ASL (the best for F1) has worse ECE than MBUR (0.2396 vs. 0.1601 on GoEmotions; 0.1802 vs. 0.1457 on CelebA). The calibration and F1 headline claims therefore come from different models.
  • No external baselines. Since the weights are now trained with standard losses, logistic regression or a small MLP on the same binarized features is a natural reference point. This is especially true for IMDB and AG News, where linear bag-of-words models are strong. Other TM variants for imbalance (e.g., focused negative sampling) and the inverse-frequency weighted CE that the authors mention trying are also not reported.
  • The interpretability evaluation is purely qualitative and selective.
    • Interpretations are averaged over 10 seeds, which produces a blurred aggregate rather than the logical rules of any one model. Rule-level interpretability is the main reason to use a TM.
    • Judgments such as "better interpretations" are subjective and rely on selected attributes and words.
    • Quantitative measures are absent: clause length or number of included literals, sparsity, rule count, or faithfulness tests such as deletion/insertion.
    • The claim that interpretability is "preserved by construction" holds for the clauses, but with real-valued weights, interpreting a class requires reasoning about weighted sums of clauses. That trade-off should be discussed.
  • Limited scope of the statistics. All variance comes from training seeds on fixed splits. Very large effect sizes (e.g., dz≈18.7d_z \approx 18.7 on IMDB) reflect tiny variance rather than practically large effects. Relatively standard benchmarks also limit the scope; CIFAR-10, a common TM image benchmark, is missing.
Questions for authors
  • Can you report ablations separating loss-guided clause feedback from gradient-trained real-valued weights, i.e., MBUR feedback with gradient-trained weights, and loss-guided feedback with integer ±1\pm 1 weights?
  • How are MBUR's probabilities computed for ECE, ROC-AUC and PR-AUC? How does MBUR compare after post-hoc temperature or Platt scaling fitted on held-out data, and against the TM probability scores of Helin et al. (2025)?
  • How were TT and qq chosen for MBUR, and why is T≥CT \ge C in most configurations? Can you show the distribution of ∣vi∣/T|v_i|/T during training, and MBUR results under a proper search with a budget equal to Loss-Guided TM's?
  • In Table 4, v6, v8 and v12 share hyperparameters but give different MBUR accuracies. Why was v12 chosen, and how does this relate to the 10-seed standard deviation in Table 2? Which η\eta was used for IMDB, 0.01 (Table 1) or 0.001 (Table 4)?
  • What is the training time and memory overhead relative to MBUR, given the per-clause, per-class loss re-evaluation? Is the method compatible with existing bit-level and hardware TM implementations?
  • How does Loss-Guided TM differ from the gradient interpretation of weighted TM updates in Sharma et al. (2022), and from differentiable or regularized TM variants?
  • Can you provide quantitative interpretability measures (clause length, literal counts, rule sparsity, faithfulness tests) for single models rather than seed averages?

Suggestions: Revise the abstract and conclusion to match the per-dataset results. Add linear and small-MLP baselines on the same binarized features. Study the effect of the C/2C/2 normalization (temperature) on calibration. Remove the leftover template sections.

Reviewer_kQKXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Loss-Guided Tsetlin Machines, replacing the conventional margin-based update rule (MBUR) with feedback driven by an explicit classification loss. The method estimates a clause’s influence using the counterfactual change in loss when its output is flipped; the sign and magnitude of that signal guide Tsetlin-automaton feedback, while continuous clause weights are trained with objective gradients. Cross-entropy (CE), symmetric cross-entropy (SCE) and asymmetric loss (ASL) are evaluated on six image and text datasets covering binary, multiclass and multilabel tasks. The authors report ten-seed comparisons, significance tests, improved predictive performance on several datasets and qualitative evidence that readable clauses remain available. The approach is promising, but the evidence does not isolate the contribution of explicit loss guidance from several concurrent changes to the learning procedure.

Critical strengths

S1. Useful objective-design interface for Tsetlin Machines. Making the loss function explicit while maintaining Boolean clause representations is a clear and practically motivated extension of the Tsetlin Machine design space. Different objectives may be useful for different task structures, particularly multilabel imbalance.

S2. Encouraging results with relatively careful statistics. The evaluation covers six datasets and ten seeds, reporting paired Wilcoxon tests with Holm correction. Reported gains include IMDB accuracy increasing from 85.52% to 88.61% with CE, GoEmotions macro F1 increasing from 28.09% to 37.06% with ASL, and CelebA macro F1 increasing from 55.21% to 63.86% with ASL (Tables 2–3).

Critical weaknesses

W1. The central causal claim is confounded by simultaneous algorithmic changes. Relative to MBUR, the proposed system changes not only the source of clause feedback but also integer/discrete clause weights to continuous gradient-trained weights (Eqs. 7–8), uses normalized logits and sigmoid/softmax probability maps (Eq. 3), changes the clause-feedback schedule and introduces different optimization hyperparameters. Holding clause count and specificity fixed does not isolate these effects. A controlled ablation separating loss-guided feedback, continuous weights, output parameterization and feedback frequency is needed before the observed improvement can be attributed specifically to the explicit learning objective. The absence of throughput and memory comparisons also leaves the cost of per-clause counterfactual loss evaluation unclear, especially in the reported 80,000-clause setting.

W2. Baseline selection, calibration comparisons and implementation details are not sufficiently controlled. The main tables compare 100th-epoch models although Figure 1 suggests MBUR can peak earlier; a validation-selected, comparably tuned and compute-matched comparison may change the conclusions. The probability mapping used to compute MBUR calibration error is insufficiently explicit. In addition, the ASL configurations with the best multilabel macro F1 have worse ECE than MBUR on GoEmotions (0.2396 versus 0.1601) and CelebA (0.1802 versus 0.1457), so the abstract should not imply simultaneous improvement for each best-performing configuration. Table 1 and Appendix Table 4 also disagree on the IMDB learning rate (η = 0.01 versus 0.001), while the scalar form of the derivative in Eq. (7) obscures the full softmax coupling. These issues materially affect attribution and reproducibility.

Questions for authors

Q1. Can the authors provide a factorial ablation separately testing counterfactual loss-guided feedback, continuous gradient-trained clause weights, normalized probability outputs and feedback frequency? Which factor accounts for most of the improvement?

Q2. How exactly are MBUR votes converted into probabilities for ECE? Can the authors use the same score-to-probability mapping and report calibration after the same validation-fitted post-hoc calibration protocol, together with NLL or Brier score?

Q3. Do the reported accuracy, macro-F1 and calibration differences persist when both methods use validation-selected checkpoints, comparable hyperparameter-search budgets and matched training time?

Q4. Which IMDB learning rate produced Table 2 (0.01 or 0.001)? What are the exact softmax CE/SCE gradients and the wall-clock/memory costs of per-clause counterfactual evaluation?

3 comments

We thank the reviewer again for the constructive review. Following the questions, the revised pdf now includes the following, with references in the main text.

  1. Ablations separating the loss-guided clause feedback from the gradient-trained weights (Q1) (comments "Ablation: Integer Weight Updates" and "Ablation: MBUR Feedback with Gradient-Trained Weights"). This provides direct evidence for our answer to Q1 (G1.1), that the clause feedback and the weight updates follow from the same loss and work as one learning mechanism rather than as separable components.
  2. Post-hoc calibration of both models with the same validation-fitted protocol, reporting ECE, NLL, and Brier score (Q2, comment "Post-hoc Calibration").
  3. Runtime of both models (W1, G4).

We hope these address the reviewer's concerns, and we would be grateful to hear whether any points remain, which we are happy to clarify.

Thank you for the additional experiments and clarifications. The validation-fitted calibration results in Appendix A.7 substantially address the concern that the calibration advantage might arise solely from the probability mapping used for MBUR. The runtime measurements in Appendix A.6 and the component ablations in Appendix A.8 also strengthen the empirical evidence.

My remaining concern is primarily the interpretation of the ablations. Table 7 shows that the complete method outperforms the two tested hybrid variants. However, the hybrids inherit hyperparameters from Table 1, and changing the weight-update mechanism may alter the vote scale and effective update dynamics. These experiments therefore do not establish that the components are intrinsically inseparable. Please either qualify the “effective only when used together” conclusion to the evaluated configurations or support the stronger claim with appropriately tuned hybrid comparisons.

Regarding the evaluation protocol, reporting both methods at epoch 100 is consistent, but does not fully address the effect of checkpoint selection when MBUR peaks earlier. Please clarify whether validation-based checkpoint selection changes the comparison. If this evidence is unavailable, the conclusions should explicitly remain limited to the reported final-epoch protocol.

Two clarifications would also improve reproducibility: specify the score transformation applied to MBUR before temperature scaling and NLL computation, and replace Eq. (7) with the general derivative ∂L/∂wik=(2ck/C) ∂L/∂zi\partial L/\partial w_{ik}=(2c_k/C)\,\partial L/\partial z_i. The MNIST exception in the post-hoc calibration results should also be stated explicitly.

Overall, the new results meaningfully strengthen the paper. The remaining issues concern the scope of the causal interpretation and the evaluation protocol.

We thank the reviewer for the follow-up and are glad the new results strengthen the paper. We have revised the PDF accordingly.

  1. Ablations. The conclusion is now qualified to the evaluated configurations, stating that with the hyperparameters of Table 1, neither hybrid matches the complete method.

  2. Checkpoint selection. Validation-based checkpoint selection does not change the comparison or the conclusions. Selecting the epoch with the best validation accuracy (macro F1 for multilabel) for each model, Loss-Guided TM matches or outperforms MBUR on all datasets but MNIST, the same pattern as at the final epoch. Test results at validation selected epoch:
    | Dataset | Metric (%) | Loss-Guided | MBUR | |---|---|---|---| | MNIST | Accuracy | 98.47 ±\pm 0.11 (CE) | 98.69 ±\pm 0.08 | | FashionMNIST | Accuracy | 90.16 ±\pm 0.18 (CE) | 90.17 ±\pm 0.18 | | IMDB | Accuracy | 88.60 ±\pm 0.11 (CE) | 85.95 ±\pm 0.53 | | AG News | Accuracy | 90.32 ±\pm 0.11 (SCE) | 89.84 ±\pm 0.17 | | GoEmotions | Macro F1 | 36.83 ±\pm 0.33 (ASL) | 35.78 ±\pm 0.79 | | CelebA | Macro F1 | 64.17 ±\pm 0.12 (ASL) | 56.61 ±\pm 0.26 |

  3. Score transformation. Temperature and Platt scaling are applied to the class sums, the unclipped viv_i for MBUR and the normalized ziz_i for Loss-Guided TM, and NLL and Brier score are computed on the resulting probabilities. This is now stated in the appendix.

  4. Eq. (7) now uses the general derivative \partial \ell/\partial w_{ik} = (c_k/(C/2)),\partial \ell/\partial z_i , as suggested.

  5. The post-hoc calibration results now state explicitly that on MNIST, calibrated MBUR achieves slightly lower ECE, NLL, and Brier score.

We hope these revisions resolve all remaining points and concerns, and that the reviewer will take them into account in their final assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank You to the Reviewers and Chairs

We thank everyone involved in the review process for their time and constructive feedback. We have addressed all points raised by the reviewers, with additional experiments where feasible. The revised PDF includes these results, and we are happy to clarify any remaining points.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.