Pith. sign in

REVIEW 4 major objections 4 minor 20 references

RandNet: deep learning with compressed measurements of images

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A recurrent sparse auto-encoder can be trained on random compressed measurements of images to perform dictionary learning and MNIST classification with minimal accuracy loss.

desk verdict A sensible integration of unrolled auto-encoders with compressed measurements, but the unsupervised recovery experiment starts from a near-oracle warm start and the MNIST evaluation has selection issues; still worth a real review. read the letter →

arxiv 1908.09258 v1 pith:SDVKVKU2 submitted 2019-08-25 cs.LG stat.ML

Thomas Chang , Bahareh Tolooshams , Demba Ba This is my paper · ORCID
classification cs.LGstat.ML
keywords randomprojectionsdictionarylearningsparsecodingcompressedmeasurementsconstrainedrecurrentauto-encoderFISTAMNISTclassificationcompressivesensing
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

The paper claims that a neural network can learn useful representations from random compressed measurements of data, rather than from the data themselves. It introduces RandNet, a modification of the constrained recurrent sparse auto-encoder in which the encoder operates on randomly projected images and the decoder reconstructs the projection. In an unsupervised setting, RandNet learns the true underlying dictionary from compressed simulations. In a supervised setting, the sparse code from the encoder serves as features for a classifier, reaching 1.56% MNIST test error with Gaussian random projections and 3.16% with row-sparse projections, at a 50% input reduction. If true, this means deep learning can proceed on data that are never fully acquired or stored, with memory and compute savings.

What carries the argument

The central object is the unrolled FISTA encoder: a recurrent network that repeats T iterations of a gradient step w_t + (1/L) A^T Phi^T (r - Phi A w_t) followed by a two-sided ReLU shrinkage nonlinearity eta_epsilon, with momentum coefficients s_t = (1 + $\sqrt$(1+4s_{t-1}^2))/2. The decoder applies Phi A, and A is shared (tied) between encoder and decoder, so backpropagation through the unrolled iterations updates only the dictionary A. The measurement matrix Phi is random and fixed; examples are divided into B blocks that share Phi_b, and only the random seeds are stored, yielding memory cost O($\beta$ N) and, for row-sparse Phi, matrix-operation cost O(gamma N) with $\beta$ = M/N and gamma = $\beta$ s.

What would settle it

Run the identical RandNet pipeline on a dataset that provably has no sparse dictionary (for instance, pixel-wise i.i.d. Gaussian noise images): if classification error under 50% Gaussian projections stays near 1.56%, the sparse-dictionary explanation is wrong. Alternatively, fix beta=0.5 on MNIST and sweep the row sparsity of Phi from 1 to 392; the RIP explanation predicts error increases monotonically as rows get sparser, so a flat or non-monotonic error curve would contradict it.

Watch

Extended reading notes

Core claim

RandNet demonstrates, for the first time, that a constrained recurrent sparse auto-encoder can be trained on compressed random measurements to perform dictionary learning and classification. The architecture solves the compressive dictionary-learning objective min_{x,A} sum_j 1/2 || r_j - Phi A x_j ||$_2^{2}$ + $\lambda$ ||x_j||_1 with unit-norm columns of A, using an encoder that unrolls T FISTA iterations and a tied linear decoder. On MNIST with 392 random measurements per 784-pixel image, the learned sparse codes classify digits at 1.56% error with Gaussian Phi and 3.16% with row-sparse Phi, compared with 3.72% and 5.20% for the CK-SVD baseline, and 1.08% and 1.05% for uncompressed DrSAE and SDL. The results are evidence that MNIST images admit a dictionary in which they are sparse and that this dictionary is identifiable from random projections, supporting a compressed-sensing explanation.

Load-bearing premise

The data must have a dictionary in which they are sparse, and the particular block-random projections used must preserve enough information to recover that dictionary and its sparse codes.

Editorial extensions

If this is right

  • Training on compressed data with Gaussian Phi cuts memory storage by a factor beta; with row-sparse Phi it cuts both memory-access and matrix-operation costs by gamma, allowing larger effective batch sizes on fixed hardware.
  • Unsupervised dictionary learning no longer requires access to full-resolution examples, so dictionaries can be learned from compressive sensors or from data that are only available as random projections.
  • The sparse codes obtained from compressed measurements are directly usable as classification features, indicating that reconstruction of the image is not a necessary intermediate step.
  • Because the measurement matrices are fixed and shared, the same framework applies to any block-structured dataset, with only random seeds stored rather than the full projection matrices.

Reading between the lines

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

  • If the result generalizes beyond MNIST, any dataset known to be sparsely representable in a fixed basis (wavelets, curvelets, or a learned dictionary) could be trained on directly from compressive-camera measurements, bypassing image reconstruction entirely.
  • The two-stage training (dictionary first, classifier second) leaves a small accuracy gap to jointly trained baselines; a testable extension is to train the classifier and dictionary jointly, which may recover that gap.
  • The paper's RIP-based explanation predicts that error rises as measurement rows become sparser; this is directly testable by sweeping row sparsity s at fixed beta, and a flat error curve would falsify that explanation.
  • The use of stored random seeds for shared projection matrices suggests a distributed or privacy-preserving variant in which each node only shares compressed projections and a seed, never raw data.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces RandNet, a recurrent sparse auto-encoder whose encoder implements T iterations of FISTA on the compressed-data dictionary-learning problem (Eq. 5), with a tied decoder ΦA; for supervised tasks, the resulting sparse code feeds a linear-softmax classifier. The authors provide a memory and computational efficiency analysis for Gaussian and row-sparse measurement matrices (Table 1), a simulation of dictionary recovery at β ∈ {0.1, 0.3, 0.5}, and MNIST classification experiments at β = 0.5 reporting 1.56% test error with Gaussian and 3.16% with row-sparse measurements.

Significance. If established, the ability to learn dictionaries and to classify from random projections with only a modest accuracy loss would be valuable in memory- and compute-limited settings. The architecture is clearly described, the extension of CRsAE to dense dictionaries and to compressed inputs is natural, and Table 1 gives a useful accounting of where the savings occur. The MNIST result at β = 0.5 is suggestive and, if reproduced with honest evaluation, would be a worthwhile contribution. However, the central unsupervised claim is not currently established because the recovery experiment starts from a near-oracle initialization, and the classification results are weakened by test-set tuning and selective reporting; these issues are fixable within the scope of the manuscript.

major comments (4)
  1. [Sec. 4.3.1, Fig. 2] The dictionary-recovery experiment initializes  by randomly perturbing A so that err(A, Â) ≈ 0.5, which places each estimated atom within roughly 30 degrees of the true atom. Fig. 2 therefore demonstrates that the unrolled FISTA auto-encoder can refine a near-oracle dictionary from compressed data; it does not demonstrate that the non-convex objective in Eq. 5 can be solved from a random initialization. Since the paper's central claim that RandNet performs dictionary learning from compressed data rests on this experiment, a random-start recovery experiment with the same simulation setup is needed, and the final err values for each β should be reported.
  2. [Sec. 4.3.2, Fig. 4] The regularization parameter λ is selected by sweeping on the test set: the text reports that the classification error reaches its minimum at λ = 2.2 for Gaussian and λ = 2 for row-sparse measurements, and these values are then used for the reported 1.56% and 3.16% test errors. This is test-set tuning, which makes the reported generalization performance optimistic. A validation split should be used for λ selection, and the test error should be reported once for the chosen λ.
  3. [Sec. 4.3.2, Table 2] The paper states that 'the accuracy for more efficient networks (lower β) was lower and hence not reported.' This omits exactly the data needed to support the memory/computation-versus-accuracy tradeoff that motivates RandNet. Please report the test errors for lower β values, or explicitly state that these experiments were not run, so that readers can judge whether the framework remains accurate at meaningful compression levels.
  4. [Sec. 4.3.2, Table 2] It is not clear from the text whether the DrSAE and SDL baselines are trained on compressed inputs or on the original images. If these baselines see the original data while RandNet sees only random projections, Table 2 does not establish RandNet's performance relative to alternatives under the same compressed-data constraint. Please clarify the input modality for each baseline and, if necessary, add compressed-input versions of the baselines.
minor comments (4)
  1. [Fig. 2] The figure shows learning curves but no final numeric err values and no indication of variability across random seeds; reporting the mean and standard deviation over several runs would make the comparison with CK-SVD more informative.
  2. [Sec. 4.2.1 and 4.2.2] The choices T = 400 for the simulation and T = 60 for MNIST are stated without sensitivity analysis; a brief discussion or ablation of the dependence on the number of FISTA iterations would help the reader assess the practical robustness of the architecture.
  3. [Sec. 2.2] The claim that this is the first time CRsAE is trained to learn a dense dictionary is strong and is not supported by a comparison with prior dense dictionary-learning auto-encoder methods; please soften the claim or add appropriate references.
  4. [Eq. (7)] The definition of the softmax output in the text is garbled ('û = eq / ∑ i e qi'); the vector notation for the cross-entropy loss should be written out cleanly so that the dimensions of C, d, and u are unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: RandNet's compressed-data dictionary learning is benchmarked against independent algorithms and a known ground-truth dictionary; cited prior work is not load-bearing in a way that reduces the claims.

full rationale

The derivation chain is self-contained. RandNet minimizes the compressive dictionary-learning objective in Eq. 5 using an unrolled FISTA encoder with a tied decoder, and the forward and backward computations are explicitly given in Sec. 3 (Algorithms 1-2). The unsupervised claim is tested on simulated data with a ground-truth dictionary A and an external error metric (Eq. 8), and RandNet is compared against CK-SVD, an independently published algorithm (ref. [7]). The MNIST classification results are benchmarked against DrSAE and SDL; the reported error rates are empirical outcomes, not quantities forced by construction. The 'Why does RandNet work?' discussion invokes compressed sensing theory (ref. [14]) post hoc to interpret the empirical success, but the success itself is not derived from that theory. There are self-citations to CRsAE (refs. [2,3]), but RandNet's operation and evaluation do not depend on accepting those papers' conclusions: the architecture equations are stated in this paper, and the CRsAE baseline is independently tested. One methodological limitation, not a circular step, is that the simulation in Sec. 4.3.1 initializes the estimated dictionary by randomly perturbing the true A so that err(A, Ahat) ≈ 0.5; Fig. 2 therefore demonstrates refinement from a warm start rather than recovery from a random start. This weakens the strength of the 'RandNet performs dictionary learning' claim, but no prediction or derived result is equivalent to its input by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard FISTA theory, on prior compressive DL identifiability results, and on an empirical domain assumption about sparsity of MNIST. Three hyperparameters (lambda, L, T) are hand-chosen. No invented entities are introduced.

free parameters (3)
  • lambda (sparsity regularization) = 2.2 (MNIST Gaussian), 2 (MNIST sparse), via grid search; in simulation lambda = sigma*sqrt(2 log p) with sigma tuned…
    Controls the sparsity of the encoder output x_T; central to both reconstruction and classification accuracy. Selected based on test error (Fig. 4), which is a free choice affecting the headline results.
  • L (FISTA step-size bound) = 5 (CRsAE), 12 (Gaussian RandNet), 2 (sparse RandNet), 50 (MNIST)
    Chosen manually to be greater than the maximum eigenvalue of A^T Phi^T Phi A; an incorrect choice can break FISTA convergence.
  • T (number of FISTA iterations) = 400 (simulation), 60 (MNIST)
    Number of unrolled iterations in the encoder; larger T better approximates the sparse code but costs more compute. T=400 was described as 'crucial' in the simulation.
assumptions (4)
  • domain assumption Compressive dictionary learning is possible from random Gaussian projections when M<N (theory from [9] and [7])
    RandNet's unsupervised objective assumes the dictionary A can be identified from compressed measurements r=Phi y; the paper relies on prior compressive DL results rather than proving identifiability for its specific block-structured setting.
  • domain assumption MNIST images admit a sparse representation in some learned dictionary A, and this representation is useful for classification
    Acknowledged in Sec. 4.3.2: 'the close-to-state-of-the-art performance of RandNet on MNIST is evidence that there indeed exists a dictionary...'. If false, the classification result would not hold.
  • domain assumption The unrolled FISTA encoder with finite T iterations produces a sparse code x_T that is close enough to the fixed point for dictionary learning and classification
    The encoder (Algorithm 1) runs T iterations; convergence is only asymptotic. The choice T=60 for MNIST is an approximation that is not validated.
  • standard math The chosen L bounds the Lipschitz constant of the gradient, ensuring FISTA convergence (standard theory from [10])
    Algorithm 1 line 5 uses 1/L step; FISTA convergence requires L >= sigma_max(A^T Phi^T Phi A). The paper states L was set 'greater than the maximum eigenvalue' but provides no spectral verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RandNet: deep learning with compressed measurements of images." pith.science (2026). https://pith.science/paper/SDVKVKU2

@misc{pith2026190809258,
  author       = {Pith},
  title        = {Pith review of: RandNet: deep learning with compressed measurements of images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SDVKVKU2}},
  note         = {Machine review of arXiv:1908.09258}
}
read the original abstract

Principal component analysis, dictionary learning, and auto-encoders are all unsupervised methods for learning representations from a large amount of training data. In all these methods, the higher the dimensions of the input data, the longer it takes to learn. We introduce a class of neural networks, termed RandNet, for learning representations using compressed random measurements of data of interest, such as images. RandNet extends the convolutional recurrent sparse auto-encoder architecture to dense networks and, more importantly, to the case when the input data are compressed random measurements of the original data. Compressing the input data makes it possible to fit a larger number of batches in memory during training. Moreover, in the case of sparse measurements,training is more efficient computationally. We demonstrate that, in unsupervised settings, RandNet performs dictionary learning using compressed data. In supervised settings, we show that RandNet can classify MNIST images with minimal loss in accuracy, despite being trained with random projections of the images that result in a 50% reduction in size. Overall, our results provide a general principled framework for training neural networks using compressed data.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 19 canonical work pages

  1. [1]

    In signal processing, dictionary learning (DL) [1] is the de facto method for learning adaptive data rep- resentations

    INTRODUCTION Representation learning has become an important problem in recent years both in the signal processing and machine learn- ing communities. In signal processing, dictionary learning (DL) [1] is the de facto method for learning adaptive data rep- resentations. In machine learning, deep learning is the method of choice to learn representations th...

    work page

  2. [2]

    DICTIONARY LEARNING Let xj ∈ Rp be a sparse vector and yj ∈ RN be the vec- tor obtained as the sum of the sparse linear combination of columns of a dictionary A∈ RN×p and additive noise yj = Axj + vj,j = 1,··· ,J, (1) 978-1-7281-0824-7/19/$31.00 c©2019 IEEE arXiv:1908.09258v1 [cs.LG] 25 Aug 2019 r ΦT 1 LAT + xt xT A Φ ˆr rg(·) lossLA C 1 d u + ˆusoftmax l...

    work page Pith review arXiv 2019

  3. [3]

    First, we in- troduce the architecture in unsupervised and supervised set- tings

    RANDNET We introduce a class of NNs, which we call RandNet, for learning representations from compressed data. First, we in- troduce the architecture in unsupervised and supervised set- tings. In unsupervised settings, RandNet is a variant of the CRsAE architecture (Sec. 2.2) to solve Eq. 5. In supervised cases, RandNet uses the output of the encoder as f...

    work page

  4. [4]

    G 0.1” stands for RandNet with Gaussian Φ andβ = 0.1. “S 0.5

    EXPERIMENTS We train RandNet on a simulated dataset and on MNIST. In the simulated case, we train a) a CRsAE architecture to learn the dictionary underlying data simulated according to Eq. 1, and b) an unsupervised RandNet architecture to learn the same dictionary from random projections of the data when Φ is Gaussian and also when it is row sparse. As a ...

    work page

  5. [5]

    Specifically, we introduced RandNet, a class of networks that, in the unsupervised setting, per- forms dictionary learning from random projections of the original data

    CONCLUSION We proposed a general framework to train NNs from com- pressed measurements. Specifically, we introduced RandNet, a class of networks that, in the unsupervised setting, per- forms dictionary learning from random projections of the original data. In the supervised setting, we highlighted the ability of RandNet in the classification of MNIST when t...

    work page

  6. [6]

    ACKNOWLEDGMENTS This work is partially supported by the Quantitative Biology Initiative at Harvard University

    work page

  7. [7]

    Learning sparsely used overcomplete dic- tionaries via alternating minimization,

    A Agarwal, A Anandkumar, P Jain, P Netrapalli, and R Tandon, “Learning sparsely used overcomplete dic- tionaries via alternating minimization,” SIAM Journal on Optimization, vol. 26, pp. 2775–2799, 2016

    work page 2016

  8. [8]

    Scalable convo- lutional dictionary learning with constrained recurrent sparse auto-encoders,

    B Tolooshams, S Day, and D Ba, “Scalable convo- lutional dictionary learning with constrained recurrent sparse auto-encoders,” in Proc. of 2018 IEEE 28th In- ternational Workshop on Machine Learning for Signal Processing, Sept. 2018, pp. 1–6

    work page 2018

Show all 20 references
  1. [9]

    Deep residual auto- encoders for expectation maximization-based dictionary learning,

    B Tolooshams, S Day, and D Ba, “Deep residual auto- encoders for expectation maximization-based dictionary learning,” 2019, arXiv:1904.08827

    2019 arXiv

  2. [10]

    Convolutional neu- ral networks analyzed via convolutional sparse coding,

    V Papyan, Y Romano, and M Elad, “Convolutional neu- ral networks analyzed via convolutional sparse coding,” Journal of Machine Learning Research , vol. 18, pp. 1– 52, 2017

    2017

  3. [11]

    Deeply-sparse signal representations (DS 2P),

    D Ba, “Deeply-sparse signal representations (DS 2P),” 2018, arXiv:1807.01958

    2018 arXiv

  4. [12]

    Memory and computation efficient pca via very sparse random pro- jections,

    F Pourkamali Anaraki and Sh Hughes, “Memory and computation efficient pca via very sparse random pro- jections,” in Proc. of the 31st International Conference on Machine Learning , Bejing, China, 22–24 Jun 2014, vol. 32, pp. 1341–1349

    2014

  5. [13]

    Efficient dictionary learning via very sparse random projections,

    F Pourkamali-Anaraki, S Becker, and Sh M Hughes, “Efficient dictionary learning via very sparse random projections,” in Proc. of 2015 International Conference on Sampling Theory and Applications , 2015, pp. 478– 482

    2015

  6. [14]

    Gradient-based learning applied to document recogni- tion,

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recogni- tion,” in Proc. of the IEEE , vol. 86, no. 11, pp. 2278– 2324, Nov 1998

    1998

  7. [15]

    Compressive k-svd,

    F. Pourkamali Anaraki and S. M. Hughes, “Compressive k-svd,” in Proc. of 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, May 2013, pp. 5469–5473

    2013

  8. [16]

    A fast iterative shrinkage- thresholding algorithm for linear inverse problems,

    A Beck and M Teboulle, “A fast iterative shrinkage- thresholding algorithm for linear inverse problems,” SIAM journal on imaging sciences , vol. 2, no. 1, pp. 183–202, 2009

    2009

  9. [17]

    Discriminative recurrent sparse auto-encoders,

    J T Rolfe and Y LeCun, “Discriminative recurrent sparse auto-encoders,” in Proc. of International Con- ference on Learning Representations, 2013, pp. 1–15

    2013

  10. [18]

    Atomic decomposition by basis pursuit,

    S S Chen, D L Donoho, and M A Saunders, “Atomic decomposition by basis pursuit,” SIAM Review, vol. 43, pp. 129–159, 1998

    1998

  11. [19]

    Supervised dictionary learning,

    J Mairal, J Ponce, G Sapiro, A Zisserman, and F R Bach, “Supervised dictionary learning,” in Proc. of Advances in Neural Information Processing Systems 21, D. Koller, D. Schuurmans, Y . Bengio, and L. Bottou, Eds., pp. 1033–1040. Curran Associates, Inc., 2009

    2009

  12. [20]

    The restricted isometry property and its implications for compressed sensing,

    E J Candes, “The restricted isometry property and its implications for compressed sensing,” Comptes rendus mathematique, vol. 346, no. 9-10, pp. 589–592, 2008

    2008

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.