Pith. sign in

REVIEW 13 references

Bayesian Comparisons Between Representations

Not yet reviewed by Pith; the record is open.

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

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2411.08739 v3 pith:Z4PQZCKX submitted 2024-11-13 cs.LG q-bio.QM

Heiko H. Schütt This is my paper · ORCID
classification cs.LGq-bio.QM
keywords representationsbayesiancomparisonslinearmetricsdistributionskernelmodel
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
read the original abstract

Which neural networks are similar is a fundamental question for both machine learning and neuroscience. Here, it is proposed to base comparisons on the predictive distributions of linear readouts from intermediate representations. In Bayesian statistics, the prior predictive distribution is a full description of the inductive bias and generalization of a model, making it a great basis for comparisons. This distribution directly gives the evidence a dataset would provide in favor of the model. If we want to compare multiple models to each other, we can use a metric for probability distributions like the Jensen-Shannon distance or the total variation distance. As these are metrics, this induces pseudo-metrics for representations, which measure how well two representations could be distinguished based on a linear read out. For a linear readout with a Gaussian prior on the read-out weights and Gaussian noise, we can analytically compute the (prior and posterior) predictive distributions without approximations. These distributions depend only on the linear kernel matrix of the representations in the model. Thus, the Bayesian metrics connect to both linear read-out based comparisons and kernel based metrics like centered kernel alignment and representational similarity analysis. The new methods are demonstrated with deep neural networks trained on ImageNet-1k comparing them to each other and a small subset of the Natural Scenes Dataset. The Bayesian comparisons are correlated to but distinct from existing metrics. Evaluations vary slightly less across random image samples and yield informative results with full uncertainty information. Thus the proposed Bayesian metrics nicely extend our toolkit for comparing representations.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 9 canonical work pages

  1. [1]

    Alain, G., & Bengio, Y . (2016). Understanding intermediate layers using linear classifier probes. International confer- ence on learning representations (ICLR) . Retrieved from https://openreview.net/forum?id=ryF7rTqgl Allen, E. J., St-Yves, G., Wu, Y ., Breedlove, J. L., Prince, J. S., Dowdle, L. T., . . . Kay, K. (2021, December). A massive 7T fMRI dat...

    work page Pith review arXiv 2016

  2. [2]

    (2014, April)

    Retrieved 2018- 04-03, from https://www.frontiersin.org/ articles/10.3389/neuro.06.004.2008/full doi: 10.3389/neuro.06.004.2008 Krizhevsky, A. (2014, April). One weird trick for par- allelizing convolutional neural networks. arXiv. Re- trieved 2024-05-18, from http://arxiv.org/abs/1404 .5997 (arXiv:1404.5997 [cs]) Krizhevsky, A., Sutskever, I., & Hinton, ...

    arXiv 2012

  3. [3]

    (2017, April)

    doi: 10.51628/001c.27664 Diedrichsen, J., & Kriegeskorte, N. (2017, April). Rep- resentational models: A common framework for understanding encoding, pattern-component, and representational-similarity analysis. PLOS Computa- tional Biology , 13(4), e1005508. Retrieved 2018-06-25, from http://journals.plos.org/ploscompbiol/ article?id=10.1371/journal.pcbi....

    work page doi:10.51628/001c.27664 2021

  4. [6]

    doi: 10.1146/annurev-neuro-080317-061906 Kriegeskorte, N., Mur, M., & Bandettini, P. A. (2008). Representational similarity analysis - connecting the branches of systems neuroscience. Frontiers in Systems Neuroscience ,

    work page doi:10.1146/annurev-neuro-080317-061906 2008

  5. [7]

    doi: 10.1109/TIT.2003.813506 Geirhos, R., Janssen, D. H. J., Schütt, H. H., Rauber, J., Bethge, M., & Wichmann, F. A. (2018, December). Compar- ing deep neural networks against humans: object recogni- tion when the signal gets weaker. arXiv:1706.06969 [cs, q- bio, stat]. Retrieved 2020-11-16, from http://arxiv.org/ abs/1706.06969 (arXiv: 1706.06969) Geirh...

    arXiv 2018

  6. [8]

    doi: 10.1016/j.neuroimage.2010.07.073 Neal, R. M. (1996). Bayesian Learning for Neural Networks (Vol. 118; P . Bickel et al., Eds.). New Y ork, NY: Springer New Y ork. Retrieved 2025-04-13, fromhttp://link.springer .com/10.1007/978-1-4612-0745-0 doi: 10.1007/978-1 -4612-0745-0 Neyshabur, B., Sedghi, H., & Zhang, C. (2020). What is being transferred in tra...

    arXiv 1996

  7. [10]

    (2017, February)

    doi: 10.1371/journal.pone.0223792 Khaligh-Razavi, S.-M., Henriksson, L., Kay, K., & Kriegeskorte, N. (2017, February). Fixed versus mixed RSA: Explain- ing visual representations by fixed and mixed feature sets from shallow and deep computational models. Journal of Mathematical Psychology , 76, 184–197. Retrieved 2018- 04-03, from http://www.sciencedirect...

    arXiv 2019

  8. [11]

    IMAGENET1K_V1

    with "IMAGENET1K_V1" weights also trained on ImageNet-1k (Rus- sakovsky et al., 2014). This networks architecture is primarily based on 4 main layers which each contain two residual blocks. I included all layers before and after the 4 layers and outputs of the two residual blocks for each main layer. The third network was the Vision Transformer ViT -B-16 ...

    work page 2014

Show all 13 references
  1. [12]

    The displayed comparisons were chosen from the matrix of comparisons between all Alexnet lay- ers and all ResNet-18 layers, but both displayed comparisons remained within ResNet-18

    The images for this simulation were always the first n images from the unlabeled images from MS COCO as described above. The displayed comparisons were chosen from the matrix of comparisons between all Alexnet lay- ers and all ResNet-18 layers, but both displayed comparisons r...

    2021

  2. [13]

    For other applica- tions in machine learning these densities need to be approx- imated based on samples as well, which can severely bias these estimators (Murphy, 2022)

    Note, that these simple sampling approximations rely on us having analytic formulas for the densities. For other applica- tions in machine learning these densities need to be approx- imated based on samples as well, which can severely bias these estimators (Murphy, 2022). Comp...

    2022

  3. [2016]

    IMAGENET1K_V1

    with pytorch version 2.1.0 (Paszke et al., 2019). The first network was Alexnet (Krizhevsky et al., 2012; Krizhevsky, 2014), for which the "IMAGENET1K_V1" weights were used, which were trained on ImageNet-1k (Russakovsky et al., 2014). For all convolutional and fully connected...

    2019

  4. [2021]

    For the signal to noise ratio, a discrete prior with 10 different equally likely values placed logarithmically between exp(−5) and 1 was used

    as prepared for the Algo- nauts challenge 2023 (Gifford et al., 2023), more specifically on voxels of the first subject in the left hemisphere reacting to the “places” functional localiser. For the signal to noise ratio, a discrete prior with 10 different equally likely values...

    2023

  5. [2664]

    E., Larsen, B

    doi: 10.1162/0899766042321814 Harvey, S. E., Larsen, B. W., & Williams, A. H. (2023, Novem- ber). Duality of Bures and Shape Distances with Implications for Comparing Neural Representations. arXiv. Retrieved 2024-05-18, from http://arxiv.org/abs/2311.11436 (arXiv:2311.11436 [c...

    2023 arXiv

Pith tools