Omni TM-AE: A Scalable and Interpretable Embedding Model Using the Full Tsetlin Machine State Space

Ahmed K. Kadhim; Lei Jiao; Ole-Christoffer Granmo; Rishad Shafik

REVIEW 4 major objections 5 minor 26 references

Omni TM-AE: A Scalable and Interpretable Embedding Model Using the Full Tsetlin Machine State Space

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Tsetlin Machine embeddings match BERT by using excluded literal states.

desk verdict A genuinely new TM-based embedding construction, but the evaluation doesn't support the headline performance claims; worth a serious referee, expecting heavy revision. read the letter →

arxiv 2505.16386 v1 pith:USESRNRT submitted 2025-05-22 cs.LG

Ahmed K. Kadhim , Lei Jiao , Rishad Shafik , Ole-Christoffer Granmo This is my paper · ORCID

classification cs.LG

keywords Tsetlinmachinewordembeddingautoencoderinterpretabilitysemanticsimilaritytextclassificationdocumentclusteringstatematrix

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

This paper claims that a Tsetlin Machine autoencoder can produce reusable, interpretable word embeddings from a single training phase by exploiting all automaton states, including the sub-threshold literals that earlier TM-based models discarded. The core idea is that a literal excluded from a clause still carries semantic signal, and that summing the signed state differences (feature state minus negation state) across positive clauses yields a dense integer vector per word. The authors report that these embeddings match or beat Word2Vec, GloVe, and FastText on semantic similarity and document clustering, and rank close to or above ELMo and BERT on sentiment classification, while remaining transparent. If correct, this would show that interpretable logical-rule models need not sacrifice performance to black-box neural embeddings.

What carries the argument

The central object is the Tsetlin automaton state matrix of a Coalesced Tsetlin Machine autoencoder: for each clause and each literal (a vocabulary feature or its negation), an integer state in $\{1,\dots,2N\}$ records how strongly that literal is memorized, with $N$ the threshold separating included from excluded literals. The machinery is the signed-state aggregation formula: for each feature $x_i$, sum the states of literal $x_i$ and subtract the states of literal $\neg x_i$ across all positive clauses, then divide by the clause count. This formula turns the whole state matrix—included and excluded literals alike—into a word embedding, and it is what lets the model avoid retraining and remain fully explainable.

What would settle it

Train Omni TM-AE on a corpus, compute word embeddings, then randomly permute or re-randomize the state values of the sub-threshold literals (those below $N$) and recompute the same embeddings and benchmarks. If similarity correlations, classification accuracy, and clustering NMI/ARI do not drop substantially relative to the original, then the claimed 'excluded literal' information is not what drives performance, and the method's contribution collapses. A simpler check is to run the same training configuration with several random seeds and measure how much the sub-threshold state values vary across runs; if they fluctuate wildly while above-threshold literals stay stable, the signed-state embedding is largely noise.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the information previously thrown away in Tsetlin-machine embeddings—the literal states below the clause-inclusion threshold $N$—is not noise but useful semantic context. For a target word, the model trains a Coalesced Tsetlin Machine autoencoder and then builds an embedding vector whose $i$-th component is the average, over all positive-weight clauses, of the automaton state of feature $x_i$ minus the automaton state of its negation $\neg x_i$. This signed-state aggregation recovers a balanced representation of a word's relationship to every vocabulary feature, and because the embedding is computed directly from the trained state matrix, no second training phase is needed. The paper demonstrates the value of this construction across semantic similarity, sentiment classification, and document clustering, where Omni TM-AE performs on par with or better than classical embeddings and is competitive with ELMo and BERT.

Load-bearing premise

The load-bearing premise is that the below-threshold automaton states carry reliable semantic signal rather than noise, so averaging signed state differences over positive clauses produces a meaningful embedding vector.

Editorial extensions

If this is right

Omni TM-AE would let practitioners build word embeddings in one logical-rule training pass and reuse them across similarity, classification, and clustering without additional training stages.
Its integer embeddings are inexpensive to store and compare, which could make interpretable embeddings practical on edge devices and in low-power hardware.
The approach brings a degree of explainability to the embedding itself: each vocabulary dimension of a word's vector is traceable to a learned logical association, not to opaque dense features.
If the similarity and clustering results generalize beyond the reported benchmarks, TM-based embeddings become a viable transparent alternative to word2vec-style and transformer-based vectors in NLP pipelines.

Reading between the lines

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

The paper does not isolate the contribution of the sub-threshold states; a clean ablation comparing embeddings built only from included literals (states above $N$) versus the full state matrix would directly test the central claim.
The signed-state formula uses only positive clauses and averages by clause count; weighting by clause weight or including negative-clause information might further improve the embeddings, but that is not tested here.
Because the method reduces to computing signed sums over a fixed state matrix, it could be extended to sentence- or document-level embeddings by treating larger text units as target classes, a direction the paper does not pursue.
The stability of sub-threshold states across random seeds and across hyperparameter choices (clause count, $N$, accumulation) is not reported, so a reproducibility study would be needed before relying on the method in practice.

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.

Desk Editor's Note

read the letter

Here's my take on Omni TM-AE. The core idea is real: using the full Tsetlin machine state matrix—including sub-threshold literal states—and computing an embedding as the signed state average over positive clauses is a new construction that extends Bhattarai et al. and Kadhim et al. It also removes the second training phase, which is a practical gain. The paper is clearly written, the algorithm is well-defined, and the interpretability story is plausible.

What the paper does well: it isolates a concrete, novel mechanism (exploiting excluded literals) and demonstrates on RG65 and MEN that the resulting similarity scores are better than the baselines. That is worth noticing.

Where it falls short: the experimental design doesn't support the central claim that this is 'comparable to ELMo and BERT.' The embeddings are 40,000-dimensional while baselines are 100-dimensional, so the comparisons are not apples-to-apples. The classification experiment uses a perturbation protocol that replaces words using the embedding being evaluated, which confounds the classifier accuracy with the embedding's similarity structure; it is not a clean test of embedding quality. There are no error bars or significance tests, and the clustering results are actually mixed—Omni TM-AE has lower NMI than Word2Vec and FastText on most datasets. Additionally, the paper asserts but never isolates the contribution of sub-threshold states; no ablation removes the sub-threshold states or replaces them with noise. The stress-test note about normalization is fair but minor—dividing by total positive clauses rather than by clause occurrence is a design choice, though it can let a single clause dominate.

The reader's verdict of REJECT is defensible for the preprint as written. The empirical claims are not fully supported, and without code or data they are hard to verify. But the core construction is solid and worth a serious referee. I would send this to a specialized venue, not desk-reject it, because the idea is novel and the flaws are in the evaluation, not in the conceptual framework. With a reworked evaluation—dimension-matched baselines, unperturbed classification, and an ablation isolating the sub-threshold signal—this could be a solid contribution.

Referee Report

4 major / 5 minor

Summary. The paper introduces Omni TM-AE, an embedding model derived from the full state matrix of a Tsetlin Machine autoencoder. Unlike prior TM-based embeddings that use only literals whose automaton states exceed a threshold N, the proposed method aggregates signed TA states of all literals, including excluded ones, across positive clauses, producing a vocabulary-sized embedding per target word in a single training phase. The authors evaluate on semantic similarity (WS-353, MTURK, RG65, MEN), sentiment classification (IMDb with a perturbation protocol), and document clustering (20 Newsgroups, Reuters, Yelp, Amazon, AG News), comparing against Word2Vec, FastText, GloVe, ELMo, and BERT. They report competitive or superior results and claim that the method balances performance, scalability, and interpretability.

Significance. If the central claim were rigorously supported, the contribution would be valuable: Omni TM-AE would provide an interpretable, reusable, single-phase embedding from a transparent logical model, potentially bridging interpretability and performance in NLP. The paper also makes a concrete, falsifiable proposal that sub-threshold TA states carry semantic signal. However, the current evidence does not isolate this mechanism, and several evaluation choices confound the comparison; the significance of the contribution is therefore not yet established.

major comments (4)

[§3.4] The central claim that 'previously excluded information' from sub-threshold TA states improves embeddings is supported only by Figure 3, a hand-annotated single clause, with no ablation. Please add a controlled comparison between Omni TM-AE and a variant that uses only above-threshold literals (or randomly permuted sub-threshold states), and report stability of the sub-threshold state signal across clauses, training seeds, and hyperparameter settings. Without such a control, the reported gains cannot be attributed to the proposed mechanism rather than to the high-dimensional projection of the full state matrix.
[§4.2, Tables 1 and 2] All comparisons are dimensionally unfair: Omni TM-AE embeddings are 40,000-dimensional (vocabulary size), while Word2Vec, FastText, and GloVe are 100-dimensional. Because higher-dimensional representations can improve nearest-neighbor and linear-classifier performance for reasons unrelated to semantic quality, please report results after PCA-reducing Omni TM-AE to 100 dimensions, or include 40,000-dimensional baselines (e.g., one-hot or random projections) with the same downstream protocol. This control is necessary to support the claimed improvements over classical embeddings.
[§4.3] The classification evaluation uses a perturbation protocol in which the embedding being evaluated is used to choose word substitutions based on sentiment labels. This creates a test distribution that is adapted to each embedding's own neighborhood structure and is not a standard benchmark. Please evaluate on the original IMDb test set, or use a perturbation dictionary fixed independently of the evaluated embeddings, and apply the identical protocol to all baselines; otherwise the accuracy differences in Table 2 may reflect differences in the perturbation generator rather than embedding quality.
[§4.4, Table 3] The clustering results do not support the abstract's claim of 'substantial improvements' in clustering: Omni TM-AE has lower NMI than Word2Vec and FastText on four of five datasets and only a small ARI advantage overall. Please either temper the claim or provide statistical significance tests and per-dataset analysis that establish a real advantage.

minor comments (5)

[§3.4] In the embedding formula, the normalization divides by the total number of positive clauses rather than by the number of clauses in which the literal actually appears; this means a single aberrant high state can dominate a coordinate. Please clarify or justify this choice, and consider normalizing per-literal occurrence counts.
[§3.4, Figure 3] Figure 3 is described as evidence that excluded literals contain semantic cues, but the annotated words are selected post hoc from a single clause; please provide a systematic evaluation of how often such annotations align with human judgments.
[§4.2, Table 1] Table 1 reports TM-AE as 'N/A' for four datasets; please state explicitly whether this is due to training cost and discuss how the absence of these baselines affects the comparison.
[Appendix A.1, Algorithm 1] Algorithm 1 uses an integer division for e_i, and the embedding vector is declared to be in Z^d; please clarify the rounding convention and whether this affects downstream similarity computations.
[Throughout] Several typos and formatting issues appear, e.g., 'MAX_IT ER' in Section 4.2, 'theTM' in the Introduction, and inconsistent capitalization of 'TM-AE' vs. 'Omni TM-AE'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: Omni TM-AE is a direct post-training aggregation of Tsetlin automaton states, validated against external human-annotated and labeled benchmarks.

full rationale

The core embedding formula (Section 3.4, Algorithm 1) is a post-hoc aggregation of the trained TM state matrix: each vocabulary coordinate is the average, over positive clauses, of the signed automaton state of the literal minus the state of its negation. This quantity is not defined in terms of any downstream evaluation target. Semantic similarity is measured against external human judgments (WS-353, MTURK, RG65, MEN), classification uses IMDb labels with independent classifiers, and clustering uses separate labeled corpora (20 News, Reuters, Yelp, Amazon, AG News). None of those labels or human scores are used to construct the embedding, so the evaluations are external rather than forced by construction. The only self-references (Kadhim et al. 2025, Bhattarai et al. 2024, Glimsdal and Granmo 2021) serve as architectural baselines or mechanism citations, not as evidence that sub-threshold states encode semantics; that claim is argued from Figure 3 and tested by the external benchmarks. The perturbation protocol in Section 4.3 does use the Omni embedding to select replacement words, which could bias that particular classification experiment as an evaluation-design issue, but it does not reduce the embedding derivation to the classification target. The paper therefore contains no load-bearing circular step.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The embedding formula itself has no fitted numeric constants; all free parameters are TM hyperparameters chosen by hand and tuned informally. The central axiom is that sub-threshold states carry semantic signal, which the paper asserts but does not independently validate.

free parameters (5)

number_of_state_bits_ta (8) = 8 bits (state range 0..255, threshold N=128)
Sets TM state resolution and inclusion threshold; set by hand, not derived from data.
clauses (32) = 32
Number of CoTM clauses; chosen manually and tuned via t-SNE visualization (Section 4.4).
accumulation (24) and number_of_examples (2000) = 24, 2000
Control construction of positive/negative training examples; chosen by hand.
T (20000) and s (1.0) = T=20000, s=1.0
TM feedback threshold and specificity; standard hyperparameters set manually.
epochs (4) = 4
Training epochs; chosen manually.

assumptions (4)

ad hoc to paper Sub-threshold automaton states encode meaningful semantic association rather than noise.
Core premise of the Omni embedding formula; motivated by Figure 3 but not proven or ablated.
domain assumption Coalesced Tsetlin Machine with clause sharing trains many target words in parallel without harmful interference.
Background from Glimsdal and Granmo (2021); relied on for the single-phase embedding claim.
domain assumption Automaton states are commensurable across different clauses, so averaging signed states over positive clauses is valid.
The embedding formula averages states across clauses; the paper does not justify this aggregation.
domain assumption Binary presence/absence input (no frequency, no order) preserves enough context for semantic similarity.
Standard count-based assumption; inherited from prior TM-AE work.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Omni TM-AE: A Scalable and Interpretable Embedding Model Using the Full Tsetlin Machine State Space." pith.science (2026). https://pith.science/paper/USESRNRT

@misc{pith2026250516386,
  author       = {Pith},
  title        = {Pith review of: Omni TM-AE: A Scalable and Interpretable Embedding Model Using the Full Tsetlin Machine State Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/USESRNRT}},
  note         = {Machine review of arXiv:2505.16386}
}

read the original abstract

The increasing complexity of large-scale language models has amplified concerns regarding their interpretability and reusability. While traditional embedding models like Word2Vec and GloVe offer scalability, they lack transparency and often behave as black boxes. Conversely, interpretable models such as the Tsetlin Machine (TM) have shown promise in constructing explainable learning systems, though they previously faced limitations in scalability and reusability. In this paper, we introduce Omni Tsetlin Machine AutoEncoder (Omni TM-AE), a novel embedding model that fully exploits the information contained in the TM's state matrix, including literals previously excluded from clause formation. This method enables the construction of reusable, interpretable embeddings through a single training phase. Extensive experiments across semantic similarity, sentiment classification, and document clustering tasks show that Omni TM-AE performs competitively with and often surpasses mainstream embedding models. These results demonstrate that it is possible to balance performance, scalability, and interpretability in modern Natural Language Processing (NLP) systems without resorting to opaque architectures.

Figures

Figures reproduced from arXiv: 2505.16386 by the authors.

**Figure 1.** (Left) State-transition diagram for a single Tsetlin automaton, illustrating penalty/reward actions between the ground state=1 “max-forgotten” to the 2N “max-memorizing” state. (Right) Clause construction space: the x-axis indexes literals (vocabulary features and their negations) while the y-axis records their current automaton state. Literals above threshold N are selected to constitute the clause. stages of their… view at source ↗

**Figure 2.** In The TM-AE structure the process of each example in epochs consists of three phases: (1) Preparation, where input vectors are generated from documents and encoded into binary form. (2) Evaluation, where input X is processed through a state matrix, weight matrix, and logical conditions to determine output predictions. and (3) Update, where TAs in a clause adjust their states to include or exclude literals based on … view at source ↗

**Figure 3.** Visualization of a trained clause for the target class “happy,” where the word “fox” is the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: t-SNE visualization of word embeddings generated by the Omni model for 130 words grouped into 13 semantic clusters. The model was trained to generate embeddings for 130 words, manually grouped into 13 semantic clusters, where each group contained words of similar mean…

**Figure 5.** Figure 5: Distribution of literal after training with a large vocabulary of 40,000 tokens (80,000 literals [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 17 canonical work pages

[1]

Darshana Abeyrathna, Ahmed A

K. Darshana Abeyrathna, Ahmed A. O. Abouzeid, Bimal Bhattarai, Charul Giri, Sondre Glimsdal, Ole-Christoffer Granmo, Lei Jiao, Rupsa Saha, Jivitesh Sharma, Svein A. Tunheim, and Xuan Zhang. Building concise logical patterns by constraining Tsetlin machine clause size . In IJCAI, 2023

work page 2023
[2]

Massively parallel and asynchronous Tsetlin machine architecture supporting almost constant-time scaling

Kuruge Darshana Abeyrathna, Bimal Bhattarai, Morten Goodwin, Saeed Rahimi Gorji, Ole-Christoffer Granmo, Lei Jiao, Rupsa Saha, and Rohan K Yadav. Massively parallel and asynchronous Tsetlin machine architecture supporting almost constant-time scaling . In ICML, 2021

work page 2021
[3]

Tsetlin machine embedding: Representing words using logical expressions

Bimal Bhattarai, Ole-Christoffer Granmo, Lei Jiao, Rohan Yadav, and Jivitesh Sharma. Tsetlin machine embedding: Representing words using logical expressions . Findings of EACL, pages 1512--1522, 2024

work page 2024
[4]

Enriching word vectors with subword information

Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information . Transactions of the Association for Computational Linguistics, 5: 0 135--146, 2017

work page 2017
[5]

One billion word benchmark for measuring progress in statistical language modeling

Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony Robinson. One billion word benchmark for measuring progress in statistical language modeling . arXiv preprint arXiv:1312.3005, 2013

arXiv 2013
[6]

Coalesced multi-output Tsetlin machines with clause sharing

Sondre Glimsdal and Ole - Christoffer Granmo. Coalesced multi-output Tsetlin machines with clause sharing . CoRR, abs/2108.07594, 2021. URL https://arxiv.org/abs/2108.07594

arXiv 2021
[7]

The Tsetlin machine - a game theoretic bandit driven approach to optimal pattern recognition with propositional logic

Ole-Christoffer Granmo. The Tsetlin machine - a game theoretic bandit driven approach to optimal pattern recognition with propositional logic . arXiv preprint arXiv:1804.01508, 2018. URL https://arxiv.org/abs/1804.01508

arXiv 2018
[8]

An Optimized Toolbox for Advanced Image Processing with Tsetlin Machine Composites

Ylva Grønningsæter, Halvor S. Smørvik, and Ole-Christoffer Granmo. An optimized toolbox for advanced image processing with Tsetlin machine composites . In Proceedings of the International Symposium on Tsetlin machine (ISTM), 2024. URL https://arxiv.org/abs/2406.00704

work page Pith review arXiv 2024

Show all 26 references

[9]

Interpretable rule-based architecture for GNSS jamming signal classification

Sindhusha Jeeru, Lei Jiao, Per-Arne Andersen, and Ole-Christoffer Granmo. Interpretable rule-based architecture for GNSS jamming signal classification. IEEE Sensors Journal, 2025

2025
[10]

Darshana Abeyrathna

Lei Jiao, Xuan Zhang, Ole-Christoffer Granmo, and K. Darshana Abeyrathna . On the convergence of Tsetlin machines for the XOR operator . IEEE Trans. Pattern Anal. Mach. Intell., 45 0 (5): 0 6072--6085, Jan. 2023

2023
[11]

Kadhim, Ole-Christoffer Granmo, Lei Jiao, and Rishad Shafik

Ahmed K. Kadhim, Ole-Christoffer Granmo, Lei Jiao, and Rishad Shafik. Exploring state space and reasoning by elimination in Tsetlin machines . In Proceedings of the International Symposium on Tsetlin machine (ISTM), 2024. URL https://arxiv.org/abs/2407.09162

2024 arXiv
[12]

Kadhim, Lei Jiao, Rishad Shafik, Ole-Christoffer Granmo, and Bimal Bhattarai

Ahmed K. Kadhim, Lei Jiao, Rishad Shafik, Ole-Christoffer Granmo, and Bimal Bhattarai. Scalable multi-phase word embedding using conjunctive propositional clauses , 2025. URL https://arxiv.org/abs/2501.19018

2025
[13]

Maas, Raymond E

Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. Learning word vectors for sentiment analysis . In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142--15...

2011
[14]

REDRESS: Generating compressed models for edge inference using Tsetlin machines

Sidharth Maheshwari, Tousif Rahman, Rishad Shafik, Alex Yakovlev, Ashur Rafiev, Lei Jiao, and Ole-Christoffer Granmo. REDRESS: Generating compressed models for edge inference using Tsetlin machines . IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (9): 0 1...

2023
[15]

Efficient estimation of word representations in vector space

Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013

2013 arXiv
[16]

Glove: Global vectors for word representation

Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation . In EMNLP, pages 1532--1543, 2014

2014
[17]

FedTMOS : Efficient one-shot federated learning with T setlin machine

Shannon How Shi Qi, Jagmohan Chauhan, Geoff V Merrett, and Jonathon Hare. FedTMOS : Efficient one-shot federated learning with T setlin machine. In The Thirteenth International Conference on Learning Representations (ICLR), 2025

2025
[18]

Tsetlin machine for solving contextual bandit problems

Raihan Seraj, Jivitesh Sharma, and Ole-Christoffer Granmo. Tsetlin machine for solving contextual bandit problems . In NeurIPS, 2022

2022
[19]

Drop clause: Enhancing performance, robustness and pattern recognition capabilities of the Tsetlin machine

Jivitesh Sharma, Rohan Yadav, Ole-Christoffer Granmo, and Lei Jiao. Drop clause: Enhancing performance, robustness and pattern recognition capabilities of the Tsetlin machine . In AAAI, 2023

2023
[20]

Tsetlin machine-based image classification FPGA accelerator with on-device training

Svein Anders Tunheim, Lei Jiao, Rishad Shafik, Alex Yakovlev, and Ole-Christoffer Granmo. Tsetlin machine-based image classification FPGA accelerator with on-device training. IEEE Trans. Circuits Syst. I, Reg. Papers, 72 0 (2): 0 830--843, Feb. 2025 a

2025
[21]

An all-digital 65-nm Tsetlin machine image classification accelerator with 8.6 nj per MNIST frame at 60.3 k frames per second

Svein Anders Tunheim, Yujin Zheng, Lei Jiao, Rishad Shafik, Alex Yakovlev, and Ole-Christoffer Granmo. An all-digital 65-nm Tsetlin machine image classification accelerator with 8.6 nj per MNIST frame at 60.3 k frames per second. arXiv preprint arXiv:2501.19347, 2025 b

2025 arXiv
[22]

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
[23]

Human-level interpretable learning for aspect-based sentiment analysis

Rohan K Yadav, Lei Jiao, Ole-Christoffer Granmo, and Morten Goodwin. Human-level interpretable learning for aspect-based sentiment analysis . In AAAI, 2021 a

2021
[24]

Enhancing interpretable clauses semantically using pretrained word representation

Rohan Kumar Yadav, Lei Jiao, Ole-Christoffer Granmo, and Morten Goodwin. Enhancing interpretable clauses semantically using pretrained word representation. In Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 265--274. ...

2021
[25]

Robust interpretable text classification against spurious correlations using AND-rules with negation

Rohan Kumar Yadav, Jiao Lei, Ole-Christoffer Granmo, and Morten Goodwin. Robust interpretable text classification against spurious correlations using AND-rules with negation . In Proc. 31st Int. Joint Conf. Artif. Intell. (IJCAI). International Joint Conferences on Artificial ...

2022
[26]

On the convergence of Tsetlin machines for the IDENTITY - and NOT Operators

Xuan Zhang, Lei Jiao, Ole-Christoffer Granmo, and Morten Goodwin. On the convergence of Tsetlin machines for the IDENTITY - and NOT Operators . IEEE Trans. Pattern Anal. Mach. Intell., 44 0 (10): 0 6345--6359, Jul. 2022

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Darshana Abeyrathna, Ahmed A

K. Darshana Abeyrathna, Ahmed A. O. Abouzeid, Bimal Bhattarai, Charul Giri, Sondre Glimsdal, Ole-Christoffer Granmo, Lei Jiao, Rupsa Saha, Jivitesh Sharma, Svein A. Tunheim, and Xuan Zhang. Building concise logical patterns by constraining Tsetlin machine clause size . In IJCAI, 2023

work page 2023

[2] [2]

Massively parallel and asynchronous Tsetlin machine architecture supporting almost constant-time scaling

Kuruge Darshana Abeyrathna, Bimal Bhattarai, Morten Goodwin, Saeed Rahimi Gorji, Ole-Christoffer Granmo, Lei Jiao, Rupsa Saha, and Rohan K Yadav. Massively parallel and asynchronous Tsetlin machine architecture supporting almost constant-time scaling . In ICML, 2021

work page 2021

[3] [3]

Tsetlin machine embedding: Representing words using logical expressions

Bimal Bhattarai, Ole-Christoffer Granmo, Lei Jiao, Rohan Yadav, and Jivitesh Sharma. Tsetlin machine embedding: Representing words using logical expressions . Findings of EACL, pages 1512--1522, 2024

work page 2024

[4] [4]

Enriching word vectors with subword information

Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information . Transactions of the Association for Computational Linguistics, 5: 0 135--146, 2017

work page 2017

[5] [5]

One billion word benchmark for measuring progress in statistical language modeling

Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony Robinson. One billion word benchmark for measuring progress in statistical language modeling . arXiv preprint arXiv:1312.3005, 2013

arXiv 2013

[6] [6]

Coalesced multi-output Tsetlin machines with clause sharing

Sondre Glimsdal and Ole - Christoffer Granmo. Coalesced multi-output Tsetlin machines with clause sharing . CoRR, abs/2108.07594, 2021. URL https://arxiv.org/abs/2108.07594

arXiv 2021

[7] [7]

The Tsetlin machine - a game theoretic bandit driven approach to optimal pattern recognition with propositional logic

Ole-Christoffer Granmo. The Tsetlin machine - a game theoretic bandit driven approach to optimal pattern recognition with propositional logic . arXiv preprint arXiv:1804.01508, 2018. URL https://arxiv.org/abs/1804.01508

arXiv 2018

[8] [8]

An Optimized Toolbox for Advanced Image Processing with Tsetlin Machine Composites

Ylva Grønningsæter, Halvor S. Smørvik, and Ole-Christoffer Granmo. An optimized toolbox for advanced image processing with Tsetlin machine composites . In Proceedings of the International Symposium on Tsetlin machine (ISTM), 2024. URL https://arxiv.org/abs/2406.00704

work page Pith review arXiv 2024