Pith. sign in

REVIEW 3 major objections 6 minor 42 references

A Hardware-oriented Approach for Efficient Bayesian Inference Computation and Deployment

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Discrete Bayesian inference on embedded GPUs can be accelerated by up to 5x without any approximation by merging the tensor contractions that dominate its runtime.

desk verdict A solid, useful optimization paper for embedded discrete Bayesian inference whose central 'numerically identical outputs' claim is threatened by an unmasked log(0) in the appendix; if that is fixed, it earns its place. read the letter →

arxiv 2607.17855 v2 pith:UNQTC3OY submitted 2026-07-20 cs.AI

Nikola Pižurica , Matteo Risso , Nikola Milović , Alessio Burrello , Igor Jovančević , Conor Heins , Miguel de Prado This is my paper
classification cs.AI
keywords BayesianinferencevariationalmessagepassingtensorcontractionGPUaccelerationedgecomputingautotuninghiddenMarkovmodelsPOMDP
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 sets out to show that the computational bottleneck in discrete Bayesian inference—specifically the variational message-passing algorithms used for hidden Markov models and partially observable decision processes—is the repeated evaluation of tensor contractions, and that this bottleneck can be removed by restructuring how those contractions are laid out in memory. The authors propose two complementary merging strategies: zero-padding arrays to a common shape and concatenating them, or flattening high-rank tensors into a single block-diagonal matrix-vector product. On an embedded GPU, across 770 randomly generated POMDP configurations, the optimized variants achieve typical speedups of 2–2.5x and peaks approaching 5x while producing numerically identical outputs to the baseline. A lightweight, trained autotuner selects the fastest variant for a given model, making the acceleration available without per-deployment benchmarking.

What carries the argument

The key mechanism is the transformation of a loop of small tensor contractions into a single large, regular contraction. Axis-aligned merging zero-pads each observation-likelihood array to a common shape and concatenates them along a batch axis, allowing one batched broadcast-multiply-sum; block-diagonal merging flattens the lagging dimensions of each array and arranges the resulting matrices as blocks of one big matrix, collapsing the whole loop into one matrix-vector product. A clustering step groups similarly shaped arrays to limit padding zeros, and batched-coordinate sparse storage optionally reduces memory footprint. These transformations are performed once at initialization, so at inf

What would settle it

Benchmark the optimized implementations on a corpus of real recorded discrete Bayesian workloads, or on models with non-trivial transition dependencies (a hidden state depending on several previous hidden states), and compare average latency ratios; if the typical speedup falls to roughly 1x or below on such realistic inputs, the paper's empirical generalization is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a broad class of discrete variational inference algorithms reduces to tensor contractions between factor potentials and marginal beliefs, and that the irregular shapes of these contractions are what make them slow on GPUs. By merging a loop of small heterogeneous contractions into a single compact, regularly shaped operation—through axis-aligned zero-padded concatenation or block-diagonal reshaping—the authors show that the same mathematical computation can be executed much faster on embedded GPU hardware with zero approximation error. Across 770 randomly sampled POMDP configurations, the fastest variants of variational message passing and marginal message passing a

Load-bearing premise

The benchmark's realism rests on hand-chosen sampling rules for the 770 POMDP configurations—including the requirement that the number of observation modalities exceed the number of hidden factors and that transition dependencies be trivial—so if real deployment workloads differ in shape distributions or dependency structure, the typical 2–2.5x speedups may not carry over.

Editorial extensions

If this is right

  • Embedded Bayesian agents—robots, automotive systems, wearables—could run active inference or filtering at real-time rates on off-the-shelf GPUs, with posterior outputs exactly identical to baseline implementations.
  • Because the optimization targets a generic computational primitive rather than algorithm-specific details, the same merging strategies can be applied to other discrete variational message-passing algorithms beyond the three demonstrated, with comparable speedups expected.
  • The autotuner result implies that deployed systems can skip exhaustive per-model benchmarking; a lightweight tree-based predictor selects the fastest variant in a fraction of a second, with typical regret of 1.6–3.6%.
  • The speedups are lossless, so they compose with any downstream use of the posteriors—planning, decision-making, or model scoring—without changing the numerical results.
  • The memory footprint of the merged representations grows modestly (usually under 1 GB on GPU-class edge hardware), so adopting the acceleration does not typically violate memory budgets.

Reading between the lines

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

  • If the synthetic benchmark's representativeness holds, the same merging logic could plausibly accelerate other tensor-contraction-heavy workloads on embedded GPUs, such as probabilistic programming backends or tensor-network computations, since the underlying primitive is generic.
  • The methodology currently restricts transition dependencies to trivial self-transitions; extending the merging strategies to arbitrary cross-factor transition structures is a natural and testable next step that would broaden the applicability to richer POMDPs.
  • The sparse representation path currently saves memory but not latency due to immature sparse operations; as sparse primitives mature, the sparse variants may close the performance gap with clustering, potentially producing further speedups.
  • A stricter transfer test would benchmark on a different embedded platform or on a corpus of real recorded workloads rather than sampled configurations; the paper's sampling rules are plausible but not derived from actual deployment data.
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

3 major / 6 minor

Summary. The paper proposes a hardware-oriented methodology for accelerating discrete Bayesian inference on embedded GPUs. The key idea is to identify tensor contractions as the dominant primitive in variational message-passing algorithms (FPI, VMP, MMP) and to restructure their memory layout by merging many small, heterogeneous contractions into larger, more regular operations using two strategies: axis-aligned padding/merging and block-diagonal merging. Memory footprint is reduced via optional BCOO sparse representations and a tensor-clustering scheme. The authors instantiate nine implementation variants for each algorithm, benchmark them on an NVIDIA Jetson Orin AGX across 770 randomly generated POMDP configurations against a pymdp/JAX baseline, report speedups up to 5x with typical gains of 2–2.5x, and claim numerically identical outputs. They also train an ML-based autotuner (RF/XGBoost) to select the fastest variant, reporting small cross-validated regret values.

Significance. If the central claims are correct, the paper provides a practical, lossless acceleration of discrete variational message passing on commodity edge GPUs, which is relevant to active inference and Bayesian agents deployed on robots or other edge devices. The paper has several strengths: the merging constructions in the Appendix are mathematically explicit; the evaluation corpus is large and persisted, enabling fair comparison across variants; and the autotuner is evaluated with cross-validated held-out configurations. The parameter-count analysis also gives a useful view of memory trade-offs. However, the log(0) issue in the axis-aligned construction currently undermines the numerical-identity claim, and the representativeness of the synthetic configuration generator is asserted rather than empirically grounded. These are load-bearing issues for the headline claims.

major comments (3)
  1. [Section III-C and Appendix A] The axis-aligned merging construction initializes A-bar with zeros and then computes log(A-bar) · o-bar. For every padded entry, log(0) = -inf. If the corresponding entry of o-bar is also zero (padded observation levels), the product is 0·(-inf) = NaN; if o-bar is nonzero (e.g., hidden dimensions padded beyond a modality's dependency list), the product is -inf. The text does not specify any masking (e.g., jnp.where(o > 0, ...) or a safe log) before the elementwise multiplication or before subsequent contractions. Since the benchmark explicitly includes heterogeneous dependency lengths and observation dimensionalities, this construction as written would contaminate merged log-likelihoods with NaN/-inf and contradict the 'numerically identical outputs' claim. End-to-end variants, which retain the padded representation through marginalization and contract against posterior vectors padded to
  2. [Section V-A] The paper claims the 770 POMDP configurations 'simulate realistic settings,' but the generator is based on manually chosen sampling rules (F≤M, dimensionality caps of 5/10/25, exponential prior over dependency-list lengths, negative correlation between list length and factor size), justified only by 'previous experience' and the self-involved reference [42]. No external workload corpus or comparison to published POMDP models is provided. Because the headline 'typical gains of 2–2.5x' and the autotuner regret are measured within this synthetic distribution, the transfer of these numbers to real edge deployments is unsupported. This is a generalization risk rather than an internal inconsistency, but it is load-bearing for the practical claims. Please either temper the realism claims or validate on an external corpus / several concrete published POMDPs.
  3. [Section V-B] The paper asserts that the optimized variants are 'numerically equivalent' and therefore skips accuracy-based evaluation, but it reports no explicit numerical identity check (e.g., maximum absolute difference, bitwise equality, or N=number of identical outputs on the persisted configurations). Given the log(0) construction concern in Comment 1, this verification is essential. Additionally, latency measurements are reported without repeated-measurement variance or confidence intervals, which is important since the speedup distributions in Fig. 4 have substantial tails. Please include at least a numerical-equivalence test on a subset of configurations and report variance or confidence intervals for the latency ratios.
minor comments (6)
  1. [Appendix A] The elementwise log(A-bar) multiplication should be written with explicit broadcasting and masking semantics; as written, the equations are ambiguous about how zero-padded entries are treated.
  2. [Section V] The exact procedure that produces 770 configurations is not fully reproducible from the text: the number of combinations from the stages should be spelled out, and the sampling seed or persistence mechanism should be described.
  3. [Section V-B] Please report the pymdp version, JAX version, and relevant backend configuration used for the baseline; this is necessary for reproducibility of the speedup ratios.
  4. [Section III-D] The block-diagonal clustering uses an empirically chosen 'elbow' of the artificial-zero curve, but no sensitivity analysis is given. Reporting how the results vary with the elbow choice would strengthen the methodology.
  5. [Section V-B] The autotuner input features are not described; it is unclear how the POMDP specification is encoded for RF/XGBoost. This should be clarified in the text or a table.
  6. [Section V-B] Fig. 5 relates parameter counts to latency, but the statement that differences are 'generally less than 1GB' is not substantiated; either show memory measurements or qualify the inference as approximate.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: core speedup results are measured against an external pymdp baseline, and the autotuner is cross-validated; only a minor non-load-bearing self-citation appears.

full rationale

The paper's central claims are empirical comparisons against an external baseline, not derivations from its own assumptions. The merging constructions in Section III-C and Appendix A are explicitly designed to preserve the looped computation by zero-padding and concatenation, so the claimed numerical equivalence is a construction property rather than a fitted prediction. Speedups in Section V-B are latency ratios against pymdp on Jetson Orin AGX, an external implementation, so there is no self-definition or fitted-input-called-prediction loop. The autotuner is trained on measured latencies but evaluated by cross-validated regret on held-out POMDP configurations (Section V-B), which limits fit-to-evaluation circularity. The only self-citation is [42], used in Section V-A to justify the 'realistic' synthetic configuration generator: 'The choice of all numerical values was driven by our previous experience with real-world Bayesian applications and edge-cutting research in the literature [3], [4], [42].' This is a mild external-validity support and is not load-bearing for the mathematical equivalence or for the measured speedups; it would at most affect generalization of the benchmark, not the derivation. The empirical elbow for clustering is an engineering choice, not a result derived from the paper's assumptions. A separate correctness concern about un-masked log(0) in the axis-aligned padded construction could threaten the 'numerically identical outputs' claim, but that is a numerical implementation issue, not circularity, and does not change this verdict.

Assumptions & free parameters 3 free parameters · 6 assumptions · 1 invented entities

The central claim rests on standard tensor-algebra facts plus a set of empirical assumptions about GPU behavior and benchmark representativeness. The only hand-tuned items are the synthetic configuration generator, the autotuner hyperparameters, and the clustering elbow heuristic. No new physical or theoretical entities are postulated.

free parameters (3)
  • POMDP configuration generator hyperparameters = F, M in {5,10,25,125}; dimension upper bounds in {5,10,25}; exponential prior on dependency-list lengths; negative-corre
    Hand-chosen in Section V-A to "simulate realistic settings"; they define the benchmark distribution over which all speedup and autotuner numbers are averaged, so they are load-bearing for the empirical generalization claim.
  • Autotuner model hyperparameters (RF/XGBoost) = Not reported in final form; selected by randomized search over 60 configurations per fold (Table II)
    The autotuner regret figures (1.58-4.78%) come from fitted ML models; the exact chosen hyperparameters are not given, which affects reproducibility but not the central speedup claim.
  • Block-diagonal clustering elbow = Chosen per problem as the elbow of the artificial-zero curve
    Section III-D: "We empirically found that the set of block-diagonal matrices corresponding to the elbow of the curve generally yields the best results." This is a tuning choice with no closed-form selection criterion.
assumptions (6)
  • standard math Coordinate-wise variational updates for discrete factorized models reduce to Einstein summations/tensor contractions over parameter tensors and marginals (Section III-A).
    Standard VI equations [15,16]; the paper's optimization target is this primitive, so soundness of the speedup claim depends on this reduction being the dominant cost.
  • standard math Zero-padding and block-diagonal arrangement reproduce the values of the original contractions exactly (Appendix A).
    Addition of zeros and block-diagonal matrix-vector multiplication preserve the mathematical result; key to the lossless claim.
  • domain assumption Regularly shaped, fused GPU operations are faster than loops over small irregular kernels (Section III-C).
    Empirical hardware model, not proven from first principles; central to why merging helps.
  • domain assumption Trivial B-dependencies cover a significant portion of practical POMDP use cases (Section III-B).
    The implementations support only trivial B-dependencies; if real workloads frequently have nontrivial transitions, the claimed breadth fails.
  • ad hoc to paper The 770 generated POMDP configurations are representative of real edge deployments (Section V-A).
    Justified by "previous experience" and coauthored reference [42], not by an external corpus; this is the weakest load-bearing premise for the generality claim.
  • domain assumption JAX BCOO sparse operations are experimentally usable and their overhead does not negate merging benefits (Sections III-D and V-B).
    The paper relies on JAX's experimental sparse backend and notes that speedups may change as the backend matures.
invented entities (1)
  • None
    purpose: No new physical or theoretical entities are introduced.
    Merged tensors, sparse layouts, and the autotuner are software artifacts, not invented entities with independent falsifiable handles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Hardware-oriented Approach for Efficient Bayesian Inference Computation and Deployment." pith.science (2026). https://pith.science/paper/UNQTC3OY

@misc{pith2026260717855,
  author       = {Pith},
  title        = {Pith review of: A Hardware-oriented Approach for Efficient Bayesian Inference Computation and Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UNQTC3OY}},
  note         = {Machine review of arXiv:2607.17855}
}
read the original abstract

Bayesian inference provides a principled foundation for reasoning under uncertainty, but its computational cost hinders deployment on resource-constrained edge devices. In this paper, we present a hardware-oriented methodology for accelerating discrete Bayesian inference on commercial off-the-shelf embedded GPUs. We identify that the latency of a broad class of variational message-passing algorithms is dominated by tensor contractions. Our approach restructures the memory layout of these operations using two complementary merging strategies that produce compact, regularly-shaped primitives better suited for efficient GPU execution. We then introduce optional sparse array representations and a tensor-clustering scheme to reduce the memory footprint. We instantiate the methodology and produce optimized variants of three message-passing algorithms for Hidden Markov Models (HMMs), namely variational filtering, variational message passing, and marginal message passing. Furthermore, we complement this with a machine-learning-based autotuner that automatically selects the best-performing algorithmic variant for a given generative model specification. Benchmarked on an NVIDIA Jetson Orin AGX across 770 randomly sampled realistic Partially Observable Markov Decision Process (POMDP) configurations, our implementations achieve speedups of up to 5x, with typical gains of 2-2.5x, while producing numerically identical outputs to the baseline implementations.

Figures

Figures reproduced from arXiv: 2607.17855 by the authors.

Figure 1
Figure 1. A visualization of contraction operations over different inputs. Only [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. (a) A visualization of the axis-aligned padding and merging strategy. Gray elements represent zero values that are added for padding. (b) Performing [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Merging tensor contractions into one standard matrix-vector multi [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Distributions of latency ratios (speedups) on an NV Jetson Orin AGX. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: A comparison between parameter counts and latencies of original [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 5 linked inside Pith

  1. [42]

    Towards smart and adaptive agents for active sensing on edge devices,

    D. Vyas, N. Pi ˇzurica, N. Milovi ´c, I. Jovan ˇcevi´c, M. de Prado, and T. Verbelen, “Towards smart and adaptive agents for active sensing on edge devices,” inEuropean Conference on EDGE AI Technologies and Applications (EEAI), 2025

    2025

  2. [1]

    The free-energy principle: a unified brain theory?

    K. Friston, “The free-energy principle: a unified brain theory?”Nature Reviews Neuroscience, vol. 11, no. 2, pp. 127–138, 2010

    2010

  3. [2]

    End-to-end pixel- based deep active inference for body perception and action,

    C. Sancaktar, M. A. J. van Gerven, and P. Lanillos, “End-to-end pixel- based deep active inference for body perception and action,” in2020 Joint IEEE 10th International Conference on Development and Learning and Epigenetic Robotics (ICDL-EpiRob). IEEE, 2020, pp. 1–8

    2020

  4. [3]

    Mobile manipulation with active inference for long- horizon rearrangement tasks,

    C. Pezzato, O. C ¸ atal, T. Van de Maele, R. J. Pitliya, and T. Verbelen, “Mobile manipulation with active inference for long- horizon rearrangement tasks,” inNeurIPS 2025 Workshop on Embodied World Models for Decision Making, 2025, poster. [Online]. Available: https://openreview.net/forum?id=iOzxiyQann

    2025

  5. [4]

    Evaluation of “As- Intended

    K. Kidera, T. Miyaguchi, and H. Yanagisawa, “Evaluation of “As- Intended” vehicle dynamics using the active inference framework,” 2025, preprint, v2 revised 29 Aug 2025. [Online]. Available: https://arxiv.org/abs/2506.00035

    arXiv 2025

  6. [5]

    Resolving uncertainty on the fly: Modeling adaptive driving behavior as active inference,

    J. Engstr ¨om, R. Wei, A. D. McDonald, A. Garcia, M. O’Kelly, and L. Johnson, “Resolving uncertainty on the fly: Modeling adaptive driving behavior as active inference,”Frontiers in Neurorobotics, vol. 18, p. 1341750, 2024

    2024

  7. [6]

    Matchbox: Large scale online Bayesian recommendations,

    D. H. Stern, R. Herbrich, and T. Graepel, “Matchbox: Large scale online Bayesian recommendations,” inProceedings of the 18th International Conference on World Wide Web (WWW), 2009, pp. 111–120

    2009

  8. [7]

    Bayesian probabilistic matrix factor- ization using markov chain monte carlo,

    R. Salakhutdinov and A. Mnih, “Bayesian probabilistic matrix factor- ization using markov chain monte carlo,” inProceedings of the 25th International Conference on Machine Learning (ICML). New York, NY , USA: Association for Computing Machinery, 2008, pp. 880–887

    2008

Show all 42 references
  1. [8]

    Recent advances in the application of predictive coding and active inference models within clinical neuroscience,

    R. Smith, P. Badcock, and K. J. Friston, “Recent advances in the application of predictive coding and active inference models within clinical neuroscience,”Psychiatry and Clinical Neurosciences, vol. 75, no. 1, pp. 3–13, 2021

    2021

  2. [9]

    From broken models to treatment selection: Active inference as a tool to guide clinical research and practice,

    R. Smith, S. S. Khalsa, and M. P. Paulus, “From broken models to treatment selection: Active inference as a tool to guide clinical research and practice,”Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 2022

    2022

  3. [10]

    TrueSkill: A Bayesian skill rating system,

    R. Herbrich, T. Minka, and T. Graepel, “TrueSkill: A Bayesian skill rating system,” inAdvances in Neural Information Processing Systems 19 (NIPS 2006), 2006, pp. 569–576

    2006

  4. [11]

    Patterns of scalable Bayesian inference,

    E. Angelino, M. J. Johnson, and R. P. Adams, “Patterns of scalable Bayesian inference,”Foundations and Trends in Machine Learning, vol. 9, no. 2–3, pp. 119–247, 2016

    2016

  5. [12]

    Neuronal mes- sage passing using Mean-field, Bethe, and Marginal approximations,

    T. Parr, D. Markovic, S. J. Kiebel, and K. J. Friston, “Neuronal mes- sage passing using Mean-field, Bethe, and Marginal approximations,” Scientific Reports, vol. 9, no. 1, p. 1889, 2019

    2019

  6. [13]

    Probabilistic inference in the era of tensor networks and differential programming,

    M. Roa-Villescas, X. Gao, S. Stuijk, H. Corporaal, and J.-G. Liu, “Probabilistic inference in the era of tensor networks and differential programming,”Physical Review Research, vol. 6, no. 3, p. 033261, 2024

    2024

  7. [14]

    An introduction to variational methods for graphical models,

    M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, and L. K. Saul, “An introduction to variational methods for graphical models,”Machine Learning, vol. 37, no. 2, pp. 183–233, 1999

    1999

  8. [15]

    Variational inference: A review for statisticians,

    D. M. Blei, A. Kucukelbir, and J. D. McAuliffe, “Variational inference: A review for statisticians,”Journal of the American Statistical Associa- tion, vol. 112, no. 518, pp. 859–877, 2017

    2017

  9. [16]

    Variational message passing,

    J. Winn and C. M. Bishop, “Variational message passing,”Journal of Machine Learning Research, vol. 6, pp. 661–694, 2005

    2005

  10. [17]

    Expectation propagation for approximate bayesian infer- ence,

    T. P. Minka, “Expectation propagation for approximate bayesian infer- ence,” inProceedings of the 17th Conference on Uncertainty in Artificial Intelligence (UAI). San Francisco, CA, USA: Morgan Kaufmann, 2001, pp. 362–369

    2001

  11. [18]

    Factor graphs and the sum-product algorithm,

    F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, “Factor graphs and the sum-product algorithm,”IEEE Transactions on Information Theory, vol. 47, no. 2, pp. 498–519, 2001

    2001

  12. [19]

    Active inference on discrete state-spaces: A synthesis,

    L. Da Costa, T. Parr, N. Sajid, S. Veselic, V . Neacsu, and K. Friston, “Active inference on discrete state-spaces: A synthesis,”Journal of Mathematical Psychology, vol. 99, p. 102447, 2020

    2020

  13. [20]

    Variational autoencoders for collaborative filtering,

    D. Liang, R. G. Krishnan, M. D. Hoffman, and T. Jebara, “Variational autoencoders for collaborative filtering,” inProceedings of the 2018 World Wide Web Conference (WWW), 2018, pp. 689–698

    2018

  14. [21]

    Dynamax: A Python package for probabilistic state space modeling with JAX,

    S. W. Linderman, P. Chang, G. Harper-Donnelly, A. Kara, X. Li, G. Duran-Martin, and K. Murphy, “Dynamax: A Python package for probabilistic state space modeling with JAX,”Journal of Open Source Software, vol. 10, no. 108, p. 7069, 2025. [Online]. Available: https://doi.org/10....

    2025 doi

  15. [22]

    borglab/gtsam,

    F. Dellaert and GTSAM Contributors, “borglab/gtsam,” May 2022. [Online]. Available: https://github.com/borglab/gtsam

    2022

  16. [23]

    Active inference and learning,

    K. Friston, T. FitzGerald, F. Rigoli, P. Schwartenbeck, G. Pezzuloet al., “Active inference and learning,”Neuroscience & Biobehavioral Reviews, vol. 68, pp. 862–879, 2016

    2016

  17. [24]

    T. Parr, G. Pezzulo, and K. J. Friston,Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. The MIT Press, 03 2022. [Online]. Available: https://doi.org/10.7551/mitpress/12441.001.0001

    2022 doi

  18. [25]

    A novel adaptive controller for robot manipulators based on active inference,

    C. Pezzato, R. M. G. Ferrari, and C. Hern ´andez, “A novel adaptive controller for robot manipulators based on active inference,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 2973–2980, 2020

    2020

  19. [26]

    Active inference in robotics and artificial agents: Survey and challenges,

    P. Lanillos, C. Meo, C. Pezzato, A. A. Meera, M. Baioumy, W. Ohata, A. Tschantz, B. Millidge, M. Wisse, C. L. Buckley, and J. Tani, “Active inference in robotics and artificial agents: Survey and challenges,”arXiv preprint arXiv:2112.01871, 2021

    2021 arXiv

  20. [27]

    An empirical study of active inference on a humanoid robot,

    G. Oliver, P. Lanillos, and G. Cheng, “An empirical study of active inference on a humanoid robot,”IEEE Transactions on Cognitive and Developmental Systems, vol. 14, no. 2, pp. 462–471, 2022

    2022

  21. [28]

    Robot navigation as hierarchical active inference,

    O. C ¸ atal, S. Wauthier, C. De Boom, T. Verbelen, and B. Dhoedt, “Robot navigation as hierarchical active inference,”Neural Networks, vol. 142, pp. 192–204, 2021

    2021

  22. [29]

    Stan: A probabilistic programming language,

    B. Carpenter, A. Gelman, M. D. Hoffman, D. Lee, B. Goodrich, M. Betancourt, M. A. Brubaker, J. Guo, P. Li, and A. Riddell, “Stan: A probabilistic programming language,”Journal of Statistical Software, vol. 76, no. 1, 2017

    2017

  23. [30]

    Pyro: Deep universal probabilistic programming,

    E. Bingham, J. P. Chen, M. Jankowiak, F. Obermeyer, N. Pradhan, T. Karaletsos, R. Singh, P. Szerlip, P. Horsfall, and N. D. Goodman, “Pyro: Deep universal probabilistic programming,”Journal of Machine Learning Research, vol. 20, no. 28, pp. 1–6, 2019

    2019

  24. [31]

    Composable effects for flexible and accelerated probabilistic programming in NumPyro,

    D. Phan, N. Pradhan, and M. Jankowiak, “Composable effects for flexible and accelerated probabilistic programming in NumPyro,”arXiv preprint arXiv:1912.11554, 2019

    1912 arXiv

  25. [32]

    pymdp: A python library for active inference in discrete state spaces,

    C. Heins, B. Millidge, D. Demekas, B. Klein, K. Friston, I. Couzin, and A. Tschantz, “pymdp: A python library for active inference in discrete state spaces,”arXiv preprint arXiv:2201.03904, 2022

    2022 arXiv

  26. [33]

    PGMax: Factor Graphs for Discrete Probabilistic Graphical Models and Loopy Belief Propagation in JAX,

    G. Zhou, A. Dedieu, N. Kumar, W. Lehrach, S. Kushagra, D. George, and M. L ´azaro-Gredilla, “PGMax: Factor Graphs for Discrete Probabilistic Graphical Models and Loopy Belief Propagation in JAX,” Journal of Machine Learning Research, vol. 25, no. 371, pp. 1–25, 2024. [Online]....

    2024

  27. [34]

    Introducing ActiveInference.jl: A Julia library for simulation and parameter estimation with active inference models,

    S. W. Nehrer, J. Ehrenreich Laursen, C. Heins, K. Friston, C. Mathys, and P. Thestrup Waade, “Introducing ActiveInference.jl: A Julia library for simulation and parameter estimation with active inference models,” Entropy, vol. 27, no. 1, p. 62, 2025

    2025

  28. [35]

    SPM12: Statistical parametric mapping,

    Wellcome Centre for Human Neuroimaging, “SPM12: Statistical parametric mapping,” MATLAB software package, 2014, version 12, FIL Methods Group, UCL. [Online]. Available: https://www.fil.ion.ucl. ac.uk/spm/software/spm12/

    2014

  29. [36]

    Evaluating emerging ai/ml accelerators: Ipu, rdu, and nvidia/amd gpus,

    H. Peng, C. Ding, T. Geng, S. Choudhury, K. Barker, and A. Li, “Evaluating emerging ai/ml accelerators: Ipu, rdu, and nvidia/amd gpus,” inCompanion of the 15th ACM/SPEC International Conference on Performance Engineering, 2024, pp. 14–20

    2024

  30. [37]

    RxInfer: A Julia package for reactive real-time Bayesian inference,

    D. Bagaev, A. Podusenko, and B. de Vries, “RxInfer: A Julia package for reactive real-time Bayesian inference,”Journal of Open Source Software, vol. 8, no. 84, p. 5161, 2023. [Online]. Available: https://doi.org/10.21105/joss.05161 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8,...

    2023 doi

  31. [38]

    Embedded FPGA acceleration of brain-like neural networks: Online learning to scalable inference,

    M. I. Al Hafiz, N. Ravichandran, A. Lansner, P. Herman, and A. Podobas, “Embedded FPGA acceleration of brain-like neural networks: Online learning to scalable inference,” 2025. [Online]. Available: https://arxiv.org/abs/2506.18530

    2025 arXiv

  32. [39]

    Planning and acting in partially observable stochastic domains,

    L. P. Kaelbling, M. L. Littman, and A. R. Cassandra, “Planning and acting in partially observable stochastic domains,”Artificial Intelligence, vol. 101, no. 1–2, pp. 99–134, 1998

    1998

  33. [40]

    Format abstraction for sparse tensor algebra compilers,

    S. Chou, F. Kjolstad, and S. Amarasinghe, “Format abstraction for sparse tensor algebra compilers,”Proceedings of the ACM on Programming Languages, vol. 2, no. OOPSLA, pp. 1–30, Oct. 2018

    2018

  34. [41]

    Pearl,Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference

    J. Pearl,Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. San Mateo, CA, USA: Morgan Kaufmann, 1988

    1988

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.