Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Constrained Co-Design for Photonic Bayesian Neural Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper establishes that photonic Bayesian neural networks split hardware constraints into those that hardware-aware training can compensate for and those that require targeted hardware changes, and provides the stability thresholds that

desk verdict Systematic, well-scoped simulation study that gives photonic BNN designers a usable trainable-vs-hard constraint checklist; external validity rests on an unvalidated Gaussian/independent noise surrogate. read the letter →

arxiv 2608.02229 v1 pith:MSJU2OJM submitted 2026-08-03 cs.LG

Hendrik Borras , Xiao Wang , Bernhard Klein , Robin Janssen , Frank Brückerhoff-Plückelmann , Wolfram Pernice , Holger Fröning This is my paper · ORCID
classification cs.LG
keywords photonicBayesianneuralnetworksconstrainedvariationalinferencehardware-awaretraininguncertaintyquantificationstochasticacceleratorco-designout-of-distributiondetection
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

Photonic Bayesian neural networks promise fast, physically stochastic sampling for uncertainty-aware AI, but the hardware is not an ideal sampler: quantization, programming error, and limited mean/variance ranges restrict which probability distributions can actually be implemented. This paper treats the photonic processor as a constrained variational family and asks, for each hardware constraint, whether retraining with the constraint in the loop can recover predictive accuracy and uncertainty quality, or whether the hardware itself must change. By systematically ablating stochasticity location (weights vs activations), modality (additive vs multiplicative), and five continuous constraint types across several datasets and out-of-distribution benchmarks, the authors derive per-variant stability ranges and translate them into concrete co-design guidelines. The central claim is that hardware-aware training recovers performance whenever the required variational family remains representable, and that when it does not, the same sensitivity analysis identifies which single hardware property to fix. Validated under the coupled constraints of a realistic time-wavelength-interleaving photonic architecture, this gives designers an a priori procedure for deciding between software adaptation and hardware intervention.

What carries the argument

The central object is the hardware-constrained stochastic operator h(xi)_psi(mu, sigma, epsilon) = C_psi(mu + sigma*epsilon) for additive stochasticity or C_psi(mu*(1 + sigma*epsilon)) for multiplicative stochasticity, where C_psi collects hardware constraints such as input quantization, signed/unsigned mean bounds, standard-deviation bounds, and programming error. This operator makes the photonic processor a constrained variational family rather than an ideal sampler, and the four variants (weight-additive, weight-multiplicative, activation-additive, activation-multiplicative) form the axes of the ablation study. The quantitative thresholds — the P95 values in Table 2, defined as the constr

What would settle it

A concrete falsifier: measure the actual noise distribution and mean-scale coupling on a real time-wavelength-interleaving photonic accelerator, then re-run the activation-space mapping with hardware-aware training. If training still fails despite extending the mean range to the Table 2 threshold, the claim that the representable mean range is the single hard constraint is wrong. Alternatively, if a non-Gaussian noise surrogate measured from the device shifts the P95 thresholds by more than the bootstrap uncertainties, the Gaussian-surrogate assumption collapses and the a priori partition lose

Watch

Extended reading notes

Core claim

The paper's central discovery is that photonic BNN inference can be modeled as constrained stochastic variational inference, where the hardware's physical limits define a constrained physical variational family that replaces the ideal Gaussian family assumed in training. Under this model, each constraint has a threshold (quantified as the P95 5%-degradation point) beyond which the hardware can no longer represent the required variational distribution; constraints inside the thresholds are trainable-through, while constraints outside are hard representational limits. On the reference time-wavelength-interleaving accelerator, weight-space stochasticity falls inside the trainable regime and har

Load-bearing premise

The whole analysis assumes that the simulated noise and constraint model — Gaussian additive/multiplicative noise, independent samples, input quantization, and programming-error values taken from earlier hardware studies — faithfully represents a real photonic accelerator; the paper explicitly leaves correlated and non-Gaussian effects outside its scope.

Editorial extensions

If this is right

  • Hardware-aware training recovers predictive accuracy and uncertainty quality for all constraints inside the stable ranges of Table 2, even when the constraints are applied jointly rather than in isolation.
  • Weight-space stochasticity provides stronger uncertainty quantification, while activation-space stochasticity tolerates substantially more mean-programming error, so the choice of stochasticity location encodes an explicit trade-off.
  • Unsigned (one-sided) mean representations are a hard requirement: they degrade weight-space models and collapse activation-space training, making signed or differential encoding mandatory.
  • Activation-space stochasticity requires roughly five orders of magnitude more representable mean range than weight-space stochasticity, so dynamic range is the limiting hardware property for activation-space mappings.
  • The qualitative trade-offs in the stable ranges transfer from a moderate to a harder benchmark, indicating the guidelines are not artifacts of a single dataset.

Reading between the lines

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

  • If the constrained-variational-family framing is right, the same ablation methodology could be applied to other non-ideal sampling platforms (memristive, spintronic, superconducting) to produce per-platform 'trainable vs hard' maps, turning co-design from case-by-case engineering into a standard characterization step.
  • The paper's P95 partitions suggest a testable prediction: a photonic accelerator with a programmable gain stage that extends the activation mean range should eliminate the need for the targeted mean-representation fix, a hypothesis the authors' own activation-space experiment already hints at.
  • Because distribution-shape mismatch and sample correlations are left outside the study, a concrete next test is whether measured non-Gaussian noise (for example Bose-Einstein statistics) shifts the P95 thresholds; the operator framework predicts that additive, mean-preserving non-Gaussianity is trainable-through while scale-entangled noise is not.
  • The 'representable variational family' criterion could be formalized as a capacity condition — the constrained family's support and variance range containing the unconstrained optimum — giving a theoretical counterpart to the empirical P95 curves.
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 / 5 minor

Summary. The paper studies photonic Bayesian neural network inference as stochastic variational inference over a constrained physical variational family. The authors implement weight- and activation-space stochastic operators (additive and multiplicative) in Pyro with hardware constraints including input quantization, mean bounds, standard-deviation bounds, and mean/standard-deviation programming error. They run per-constraint ablations on LeNet/Dirty-MNIST and ResNet-18/CIFAR-10, fit logistic curves to OOD AUROC, and extract P95 thresholds at 5% degradation. These thresholds are translated into co-design guidelines and then applied jointly to a time-wavelength-interleaving (A-TWI) constraint set on ResNet-18/CIFAR-10 and CINIC-10. The central claim is that hardware-aware training recovers predictive performance and uncertainty quality whenever the required variational family remains representable, whereas violations of representational limits require targeted hardware modifications.

Significance. If the results hold, the paper provides a useful framework for co-designing photonic BNN hardware and training, with an operator-level abstraction that cleanly separates stochasticity location, modality, and hardware constraints. The main strengths are the systematic ablation across four stochastic operator variants, the explicit caveats in Section 6, and the fact that the P95 thresholds are concrete, falsifiable predictions for future hardware measurements. The paper is also transparent about the scope: all experiments are simulation-based in Pyro, no code or data are released, and the coupled validation in Section 5 is a self-consistency check on the same simulator used to derive Table 2. Consequently, the significance is conditional on the fidelity of the Gaussian/independent-sample surrogate to real photonic hardware.

major comments (3)
  1. [§6, Table 1, §5] The central claim that Table 2 thresholds predict a trainable-versus-hard constraint partition is only validated against the same Pyro surrogate from which the thresholds were derived. Section 6 explicitly leaves distribution mismatch and temporal/inter-channel correlations outside the study, and Table 1's A-TWI values are taken from prior hardware work rather than measured end-to-end here. If physical q_phys differs from this surrogate (e.g., Bose-Einstein statistics, heavy tails, mean-scale coupling, or sample correlations), the P95 boundaries need not transfer. A concrete, feasible check would be to record actual A-TWI noise statistics at representative operating points and re-run at least one W-Add and one A-Mul configuration under the measured q_phys.
  2. [§5, Figure 4 and Figure 8] The coupled validation is reported only as bar charts without numerical values, error bars, or confidence intervals. The claims that hardware-aware training 'recovers substantially' and that activation-space training 'remains unstable' cannot be quantitatively assessed, and it is unclear whether the hardware-modified variant reaches the unconstrained baseline. Moreover, the mean-representation modification is not specified: what value of μ_max was used for the relaxed activation-space mapping, and is a range increase from ±1 to values above 2,000 (as implied by Table 2) plausibly a 'targeted' hardware modification? Please report exact accuracy and AUROC values with error bars and list the exact constraint values used in each condition.
  3. [§4, Table 2] The P95 thresholds are defined by an ad-hoc 5% drop from a logistic fitted maximum, with only three repetitions and manual selection from the Pareto-optimal hyperparameter set. Since the entire co-design verdict rests on these thresholds, the sensitivity of the trainable/hard boundary to the 5% criterion, the choice of hyperparameters, and the logistic fit parametrization should be quantified. Without such sensitivity analysis, the Table 2 thresholds could be artifacts of the fitting procedure rather than robust properties of the constrained variational family. A simple perturbation around the 5% threshold and the fitted curve would help establish robustness.
minor comments (5)
  1. [Table 1 and §5] Please clarify whether A-TWI is additive, multiplicative, or both. Table 1 lists both Add and Mul σ ranges for A-TWI, while Section 5 treats them as separate conditions; the reader should know which physical mechanism corresponds to which mode.
  2. [Figures 4 and 8] The labels 'Direct inference on HW' and 'HW-aware training' may overstate the physicality of the experiments, since all results are simulated in Pyro. Consider renaming to 'direct transfer to constrained surrogate' and 'constraint-aware training' to avoid confusion.
  3. [General] No code, data, seeds, or software versions are released, which limits reproducibility of the logistic fits and ablations. A reproducibility package would substantially strengthen the paper.
  4. [Figure 5 caption] The note that unsigned constraints in A-Mul produced NaNs is important supporting evidence for the signedness claim and should be stated in the main text, not only in the figure caption.
  5. [§4.1] The bootstrap procedure for uncertainty estimation is described very tersely. Please specify the resampling units, the number of bootstrap samples beyond 500, and the exact form of the logistic function used for fitting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Table 2 thresholds are empirical fits, and Section 5 tests them on a held-out dataset under coupled constraints; remaining concerns are external validity, not circularity.

full rationale

The paper's central derivation is an empirical ablation-to-validation chain, not a derivation that reduces to its own inputs. Table 2 thresholds (P95 values) are fitted from single-constraint ablations on ResNet-18/CIFAR-10 using the Pyro constraint model. Section 5 then applies these thresholds as predictions to a different dataset (CINIC-10) with all A-TWI constraints applied jointly. This is a genuine transfer test: the joint outcome is not guaranteed by the individual ablations, and the dataset differs from the one used to fit the thresholds. Weight-space A-TWI values are predicted to lie inside the stable range and training recovers; activation-space values are predicted to violate the mean bound and training fails until that bound is relaxed. Both outcomes confirm the fitted thresholds rather than being definitionally entailed by them. The phrase 'required variational family remains representable' is operationalized by the independently fitted stable ranges, so the abstract's conditional claim is an empirical summary, not a tautology. The self-citations to prior hardware work (A-XBAR, A-TWI) supply measured constraint values and support the Gaussian-surrogate assumption; these are external, falsifiable measurements rather than unverified self-support. The paper explicitly concedes its main limitation: 'We also assume sufficiently independent samples and a suitable Gaussian surrogate, leaving distribution mismatch and temporal or inter-channel correlations outside the present study' (Section 6). This is an external-validity caveat—the thresholds may not transfer to real hardware with non-Gaussian or correlated noise—but it is not circularity. Similarly, the A-TWI constraint values are adopted from prior hardware work rather than measured end-to-end here; this affects provenance and generalizability, not internal circularity. The validation is a self-consistency check within one simulation stack, which reduces the strength of the hardware claim, but it does not make any 'prediction' equivalent to its inputs by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced; the 'constrained physical variational family' and 'stochastic neural operator' are abstract modeling concepts, not new hardware components. The central results rest on the fidelity of the Pyro constraint operator, the Gaussian/independence surrogate, and the fitting of P95 thresholds to simulation data.

free parameters (2)
  • Per-configuration training hyperparameters (learning rate, KL annealing factor, prior std, initial variational std) = Not reported numerically; tuned by Optuna (29-37 trials) and manually selected from Pareto set
    Every ablation uses a baseline tuned this way; reported accuracy and AUROC depend on these choices.
  • P95 5%-degradation thresholds per constraint (Table 2) = e.g. input quantization b>=4; signed mean bound |mu_max|>0.003 (W-Mul) to >2100 (A-Add); std and programming-noise bound
    Obtained by fitting logistic curves to simulated OOD AUROC values; these are the central quantitative outputs and are reused to predict joint-constraint behavior.
assumptions (5)
  • standard math SVI with a closed-form Gaussian variational family is a valid way to train BNNs for the studied image tasks.
    Used throughout Section 2.1 and implemented in Pyro; standard variational inference background.
  • domain assumption The additive/multiplicative constrained operator h_psi with quantization, mean/std bounds, and Gaussian programming noise faithfully models real photonic BNN hardware.
    Section 3.1-3.2 defines the operator model; all ablation and validation results inherit this fidelity assumption.
  • domain assumption Physical noise is a suitable Gaussian surrogate and samples are sufficiently independent; distribution-shape mismatch and temporal/inter-channel correlations are negligible.
    Explicitly stated in Section 6 as a limitation: 'We also assume sufficiently independent samples and a suitable Gaussian surrogate, leaving distribution mismatch and temporal or inter-channel correlations outside the present study.'
  • ad hoc to paper A 5% drop in fitted OOD AUROC defines the tolerable stable range of a constraint.
    Section 4: the authors fit logistic curves and identify the constraint level causing a 5% drop; this criterion is chosen by the authors rather than derived or externally benchmarked.
  • domain assumption Stable ranges measured on ResNet-18/CIFAR-10 transfer to other architectures and datasets (e.g., CINIC-10) and to joint constraint combinations.
    Figures 4 and 8 test one transfer; the general guideline in Table 3 assumes broader transferability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Constrained Co-Design for Photonic Bayesian Neural Networks." pith.science (2026). https://pith.science/paper/MSJU2OJM

@misc{pith2026260802229,
  author       = {Pith},
  title        = {Pith review of: Constrained Co-Design for Photonic Bayesian Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MSJU2OJM}},
  note         = {Machine review of arXiv:2608.02229}
}
read the original abstract

Classical neural networks frequently produce overconfident predictions on ambiguous or out-of-distribution (OOD) data, a liability that grows with each AI system deployed in safety-critical real-world scenarios. Bayesian neural networks (BNNs) provide a principled framework for uncertainty-aware prediction by replacing deterministic parameters with probability distributions, but repeated sampling increases latency, memory traffic, and energy consumption. Photonic probabilistic computing offers a promising alternative by exploiting intrinsic optical stochasticity for fast and parallel sampling. However, photonic BNNs are not ideal samplers: analog constraints on quantization, programming error, dynamic range, and representable mean and variance restrict the variational families that can be implemented in hardware. In this work, we study which hardware-imposed constraints limit scalable photonic BNN inference, how these constraints can be represented, and which ranges can be tolerated by photonic BNNs beyond small proof-of-concept networks. We formulate photonic BNN inference as constrained stochastic variational inference and perform a systematic ablation study over stochasticity location, stochasticity modality, quantization, programming error, and mean/variance bounds. From these results, we derive concrete co-design guidelines that distinguish hardware constraints that can be compensated by training from those requiring hardware or architecture intervention. We validate these guidelines under coupled, hardware-realistic constraints on Dirty-MNIST, CIFAR-10, and CINIC-10, using Fashion-MNIST and SVHN as OOD benchmarks, showing that hardware-aware training recovers predictive performance and uncertainty quality whenever the required variational family remains representable, whereas violations of representational limits require targeted hardware modifications.

Figures

Figures reproduced from arXiv: 2608.02229 by the authors.

Figure 1
Figure 1. Schematic implementation of the two Pyro operators used to simulate photonic hardware. Left: Weight space locality; [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Effects of lower bound of constrained Std. Results [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Co-design study for an example accelerator based on the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Ablation study of individual hardware constraints under hardware-aware training. Results are reported for ResNet-18 [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Ablation study of individual hardware constraints applied only during inference to the unconstrained baseline model. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Ablation study of individual hardware constraints under hardware-aware training. Results are reported for LeNet on [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Co-design study for an example accelerator based on the [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith

  1. [1]

    Akiba, T.; Sano, S.; Yanase, T.; Ohta, T.; and Koyama, M. 2019. Optuna: A Next-generation Hyperparameter Optimization Framework. In 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD. Association for Computing Machinery

    2019

  2. [2]

    M.; Kucukelbir, A.; and McAuliffe, J

    Blei, D. M.; Kucukelbir, A.; and McAuliffe, J. D. 2017. Variational inference: A review for statisticians. Journal of the American Statistical Association, 112(518): 859--877

    2017

  3. [3]

    Blundell, C.; Cornebise, J.; Kavukcuoglu, K.; and Wierstra, D. 2015. Weight Uncertainty in Neural Networks. In International Conference on Machine Learning, ICML

    2015

  4. [4]

    Bonnet, D.; Hirtzlin, T.; Majumdar, A.; Dalgaty, T.; Esmanhotto, E.; Meli, V.; Castellani, N.; Martin, S.; Nodin, J.-F.; Bourgeois, G.; Portal, J.-M.; Querlioz, D.; and Vianello, E. 2023. Bringing uncertainty quantification to the extreme-edge with memristor-based Bayesian neural networks. Nature Communications, 14(1): 7530

    2023

  5. [5]

    Borras, H.; Wu, Y.; Klein, B.; and Fröning, H. 2025. Uncertainty-Preserving QBNNs: Multi-Level Quantization of SVI-Based Bayesian Neural Networks for Image Classification. CoRR, abs/2512.10602

    arXiv 2025

  6. [6]

    u ckerhoff-Pl \

    Br \"u ckerhoff-Pl \"u ckelmann, F.; Borras, H.; Klein, B.; Varri, A.; Becker, M.; Dijkstra, J.; Br \"u ckerhoff, M.; Wright, C. D.; Salinga, M.; Bhaskaran, H.; Risse, B.; Fr \"o ning, H.; and Pernice, W. 2024. Probabilistic photonic computing with chaotic light. Nature Communications, 15(1): 10445

    2024

  7. [7]

    U.; Meyer, L.; Ji, X.; Hu, J.; Sun, J.; Klein, B.; Ebert, F.; Dijkstra, J.; McRae, L.; Schmidt, P.; Kippenberg, T

    Brückerhoff-Plückelmann, F.; Borras, H.; Hulyal, S. U.; Meyer, L.; Ji, X.; Hu, J.; Sun, J.; Klein, B.; Ebert, F.; Dijkstra, J.; McRae, L.; Schmidt, P.; Kippenberg, T. J.; Fröning, H.; and Pernice, W. 2025 a . Uncertainty Reasoning with Photonic Bayesian Machines. arXiv:2512.02217

    arXiv 2025

  8. [8]

    P.; Varri, A.; Borras, H.; Klein, B.; Meyer, L.; Wright, C

    Brückerhoff-Plückelmann, F.; Ovvyan, A. P.; Varri, A.; Borras, H.; Klein, B.; Meyer, L.; Wright, C. D.; Bhaskaran, H.; Syed, G. S.; Sebastian, A.; Fröning, H.; and Pernice, W. 2025 b . Probabilistic photonic computing for AI. Nature Computational Science, 5(5): 377–387

    2025

Show all 39 references
  1. [9]

    Z.; and Soljačić, M

    Choi, S.; Salamin, Y.; Roques-Carmes, C.; Dangovski, R.; Luo, D.; Chen, Z.; Horodynski, M.; Sloan, J.; Uddin, S. Z.; and Soljačić, M. 2024. Photonic probabilistic machine learning using quantum vacuum noise. Nature Communications, 15(1): 7760

    2024

  2. [10]

    N.; Crowley, E

    Darlow, L. N.; Crowley, E. J.; Antoniou, A.; and Storkey, A. J. 2018. CINIC-10 is not ImageNet or CIFAR-10. CoRR, abs/1810.03505

    2018 arXiv

  3. [11]

    Defazio, A.; Yang, X.; Mehta, H.; Mishchenko, K.; Khaled, A.; and Cutkosky, A. 2024. The Road Less Scheduled. In Advances in Neural Information Processing Systems, NeurIPS

    2024

  4. [12]

    Depeweg, S.; Hernandez-Lobato, J.-M.; Doshi-Velez, F.; and Udluft, S. 2018. Decomposition of Uncertainty in Bayesian Deep Learning for Efficient and Risk -sensitive Learning . In International Conference on Machine Learning, ICML

    2018

  5. [13]

    Ferianc, M.; Gamper, J.; Chen, C.-Y.; Delaunay, Q.; Li, Y.; and Murray, I. 2021. On the Effects of Quantisation on Model Uncertainty in B ayesian Neural Networks. In 37th Conference on Uncertainty in Artificial Intelligence, UAI

    2021

  6. [14]

    Gal, Y.; and Ghahramani, Z. 2016. Dropout as a Bayesian approximation: representing model uncertainty in deep learning. In International Conference on Machine Learning, ICML

    2016

  7. [15]

    Graves, A. 2011. Practical Variational Inference for Neural Networks. In Advances in Neural Information Processing Systems, NeurIPS

    2011

  8. [16]

    D.; Blei, D

    Hoffman, M. D.; Blei, D. M.; Wang, C.; and Paisley, J. 2013. Stochastic variational inference. Journal of Machine Learning Research, 14(1): 1303–1347

    2013

  9. [17]

    V.; Laga, H.; Boussaid, F.; Buntine, W.; and Bennamoun, M

    Jospin, L. V.; Laga, H.; Boussaid, F.; Buntine, W.; and Bennamoun, M. 2022. Hands-On Bayesian Neural Networks—A Tutorial for Deep Learning Users. IEEE Computational Intelligence Magazine, 17(2): 29--48

    2022

  10. [18]

    Kendall, A.; and Gal, Y. 2017. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? In Advances in Neural Information Processing Systems, NeurIPS

    2017

  11. [19]

    Krizhevsky, A. 2009. Learning multiple layers of features from tiny images. Technical report

    2009

  12. [20]

    Lakshminarayanan, B.; Pritzel, A.; and Blundell, C. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems, NeurIPS

    2017

  13. [21]

    LeCun, Y.; Cortes, C.; and Burges, C. 2010. MNIST handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2

    2010

  14. [22]

    R.; Subedar, M.; Sanghavi, V.; Arunachalam, M.; Tickoo, O.; Iyer, R.; and Kandemir, M

    Lin, J.-L.; Krishnan, R.; Ranipa, K. R.; Subedar, M.; Sanghavi, V.; Arunachalam, M.; Tickoo, O.; Iyer, R.; and Kandemir, M. T. 2023. Quantization for Bayesian Deep Learning: Low-Precision Characterization and Robustness. In IEEE International Symposium on Workload Characteriza...

    2023

  15. [23]

    P.; Kwon, J.; Debusschere, B

    Liu, S.; Xiao, T. P.; Kwon, J.; Debusschere, B. J.; Agarwal, S.; Incorvia, J. A. C.; and Bennett, C. H. 2022. Bayesian neural networks using magnetic tunnel junction-based probabilistic in-memory computing. Frontiers in Nanotechnology, 4

    2022

  16. [24]

    MacKay, D. J. C. 1992. A Practical Bayesian Framework for Backpropagation Networks. Neural Computation, 4(3): 448--472

    1992

  17. [25]

    Mukhoti, J.; Kirsch, A.; van Amersfoort, J.; Torr, P. H. S.; and Gal, Y. 2022. Deep Deterministic Uncertainty : A Simple Baseline . CoRR, abs/2102.11582

    2022 arXiv

  18. [26]

    G.; and Hutter, F

    M\"uller, S. G.; and Hutter, F. 2021. TrivialAugment: Tuning-Free Yet State-of-the-Art Data Augmentation. In IEEE/CVF International Conference on Computer Vision, ICCV

    2021

  19. [27]

    Nado, Z.; Band, N.; Collier, M.; Djolonga, J.; Dusenberry, M.; Farquhar, S.; Filos, A.; Havasi, M.; Jenatton, R.; Jerfel, G.; Liu, J.; Mariet, Z.; Nixon, J.; Padhy, S.; Ren, J.; Rudner, T.; Wen, Y.; Wenzel, F.; Murphy, K.; Sculley, D.; Lakshminarayanan, B.; Snoek, J.; Gal, Y.;...

    2021 arXiv

  20. [28]

    Netzer, Y.; Wang, T.; Coates, A.; Bissacco, A.; Wu, B.; and Ng, A. Y. 2011. Reading Digits in Natural Images with Unsupervised Feature Learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011

    2011

  21. [29]

    V.; Lakshminarayanan, B.; and Snoek, J

    Ovadia, Y.; Fertig, E.; Ren, J.; Nado, Z.; Sculley, D.; Nowozin, S.; Dillon, J. V.; Lakshminarayanan, B.; and Snoek, J. 2019. Can you trust your model's uncertainty? E valuating predictive uncertainty under dataset shift. In 33rd International Conference on Neural Information ...

    2019

  22. [30]

    W.; Park, T.; Bartlett, B.; Williamson, I

    Pai, S.; Sun, Z.; Hughes, T. W.; Park, T.; Bartlett, B.; Williamson, I. A. D.; Minkov, M.; Milanizadeh, M.; Abebe, N.; Morichetti, F.; Melloni, A.; Fan, S.; Solgaard, O.; and Miller, D. A. B. 2023. Experimentally realized in situ backpropagation for deep learning in photonic n...

    2023

  23. [31]

    J.; Mackin, C.; Gallo, M

    Rasch, M. J.; Mackin, C.; Gallo, M. L.; Chen, A.; Fasoli, A.; Odermatt, F.; Li, N.; Nandakumar, S. R.; Narayanan, P.; Tsai, H.; Burr, G. W.; Sebastian, A.; and Narayanan, V. 2023. Hardware-aware training for large-scale and diverse deep learning inference workloads using in-me...

    2023

  24. [32]

    A.; Redwing, J

    Sebastian, A.; Pendurthi, R.; Kozhakhmetov, A.; Trainor, N.; Robinson, J. A.; Redwing, J. M.; and Das, S. 2022. Two-dimensional materials-based probabilistic synapses and reconfigurable neurons for measuring inference uncertainty using Bayesian neural networks. Nature Communic...

    2022

  25. [33]

    Shannon, C. E. 1948. A mathematical theory of communication. The Bell system technical journal, 27(3): 379--423

    1948

  26. [34]

    N.; and Tickoo, O

    Subedar, M.; Krishnan, R.; Kashyap, S. N.; and Tickoo, O. 2021. Quantization of Bayesian Neural Networks and Its Effect on Quality of Uncertainty. In ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning

    2021

  27. [35]

    G.; Onodera, T.; Stein, M

    Wright, L. G.; Onodera, T.; Stein, M. M.; Wang, T.; Schachter, D. T.; Hu, Z.; and McMahon, P. L. 2022. Deep physical neural networks trained with backpropagation. Nature, 601(7894): 549–555

    2022

  28. [36]

    Wu, C.; Yang, X.; Chen, Y.; and Li, M. 2023. Photonic Bayesian Neural Network Using Programmed Optical Noises. IEEE Journal of Selected Topics in Quantum Electronics, 29(2: Optical Computing): 1--6

    2023

  29. [37]

    Xiao, H.; Rasul, K.; and Vollgraf, R. 2017. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms. CoRR, abs/1708.07747

    2017 arXiv

  30. [38]

    G.; Chu, S

    Xu, X.; Tan, M.; Corcoran, B.; Wu, J.; Boes, A.; Nguyen, T. G.; Chu, S. T.; Little, B. E.; Hicks, D. G.; Morandotti, R.; Mitchell, A.; and Moss, D. J. 2021. 11 TOPS photonic convolutional accelerator for optical neural networks. Nature, 589(1): 44--51

    2021

  31. [39]

    P.; Li, N.; and Lee, C

    Zhuge, Y.; Ren, Z.; Xiao, Z.; Zhang, Z.; Liu, X.; Liu, W.; Xu, S.; Ho, C. P.; Li, N.; and Lee, C. 2025. Photonic Bayesian Neural Networks: Leveraging Programmable Noise for Robust and Uncertainty-Aware Computing. Advanced Science, 12(23): 2500525

    2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.