Leveraging Transfer Learning for Astronomical Image Analysis

Crescenzo Tortora; Demetra De Cicco; Gianluca Sasanelli; Giuseppe Angora; Giuseppe Longo; Lars Doorenbos; Massimo Brescia; Maurizio Paolillo; Olena Torbaniuk; Stefano Cavuoti

REVIEW 4 major objections 6 minor 22 references

Leveraging Transfer Learning for Astronomical Image Analysis

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A single ImageNet-pretrained feature extractor, used with no labeled astronomy training, can support AGN detection, galaxy property estimation, strong-lensing search, and anomaly detection because distances in its feature space track…

desk verdict A clearly written recap of the authors' own ULISSE work; no new results, and the core transfer claim remains under-defended, but it is an honest proceedings summary. read the letter →

arxiv 2411.18206 v1 pith:YLKCJTNS submitted 2024-11-27 astro-ph.IM

Stefano Cavuoti , Lars Doorenbos , Demetra De Cicco , Gianluca Sasanelli , Massimo Brescia , Giuseppe Longo , Maurizio Paolillo , Olena Torbaniuk

show 2 more authors

Giuseppe Angora Crescenzo Tortora

This is my paper · ORCID

classification astro-ph.IM

keywords transferlearningastronomicalimageanalysisfeatureextractionEfficientNetsimilaritysearchAGNdetectionanomalystronggravitationallensing

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 brings together four astronomical applications built on one idea: an ImageNet-pretrained EfficientNet, with its classification layer removed, converts any input image into a 1280-dimensional feature vector without any labeled astronomy training. Once every object is a point in this space, distances between points can be used to find similar objects, flag outliers, estimate physical properties, and detect rare events. The authors report that this label-free geometric representation identifies active galactic nuclei (AGN) candidates in three-band SDSS images, estimates galaxy stellar mass and star formation rate within about one dex for over 80% of targets, retrieves strong-lensing candidates with high accuracy in simulations, and flags bad epochs in time series. The point of the paper is that a single generic feature extractor can serve as a flexible foundation for many survey-analysis tasks, which matters as LSST-scale datasets make manually labeled training sets impractical.

What carries the argument

The central object is ULISSE, a feature-extraction and similarity-search framework built on EfficientNet, a convolutional neural network pretrained on ImageNet with the classifier head removed. The network converts each input image into a 1280-dimensional feature vector, and the paper treats those vectors as points in a geometric space where distance is the similarity measure. That single machinery does the load-bearing work across applications: query-by-example retrieval for AGN and lens candidates, neighbor averaging for galaxy properties, and comparison against a stacked reference image for time-series anomalies and outlier detection.

What would settle it

Run the same nearest-neighbor retrieval on a fixed galaxy sample with the ImageNet-pretrained EfficientNet features and again with the same architecture using random weights; if retrieval accuracy for AGN, stellar mass, or star formation rate is unchanged, then the ImageNet pretraining is not doing the work and the central transfer claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that the visual features a convolutional network learns on ordinary photographs transfer to astronomy well enough that similarity in the resulting feature space tracks astrophysical similarity, not just visual appearance. Concretely, the paper transforms the problem of analyzing a complex astronomical image into a geometric problem in a high-dimensional feature space: each image becomes a point among 1280 features, and the distance between points defines which objects are most similar or dissimilar. On this basis the authors report that AGN candidates can be retrieved without training for the task, that stellar mass and star formation rate can be estimated from images alone by averaging over retrieved neighbors, that strong-lensing candidates stand out in simulated data with 94% accuracy in the first hundred most frequent candidates, and that outlier epochs in light curves can be flagged against a stacked-image reference. The claim is that this works despite the small, noisy, and heterogeneous character of astronomical survey data.

Load-bearing premise

The load-bearing premise is that the visual regularities a network learns from ordinary photographs also order astronomical images by astrophysical similarity, so that points close in the feature space are physically alike.

Editorial extensions

If this is right

AGN candidates can be retrieved from a single three-band image without any AGN-specific training, with reported retrieval efficiency above 33% for the tested prototype morphologies and up to 60% for composite objects.
Galaxy stellar mass and star formation rate can be estimated directly from images by averaging over retrieved neighbors, with over 80% of target objects within 1 dex of the estimate.
Strong-lensing candidates in simulated data appear at the top of the retrieval list: 94% of the first 100 most frequent candidates are true positives, and 93% of the first 200.
Problematic epochs in time series can be flagged automatically by comparing each epoch's feature vector with the stacked 'typical' image, without manual inspection.
The same feature space enables large-scale anomaly detection by defining normality as low distance to the nearest neighbor, offering a route to finding rare or unknown object classes in upcoming large surveys.

Reading between the lines

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

An implication the authors leave implicit is that if feature-space distance tracks physical similarity, the same extractor can index entire survey databases, letting astronomers query any new phenomenon by example rather than retraining a model.
The reported drop in strong-lensing performance from simulated to real data points to a boundary of the transfer: it likely works best when the query distribution matches the data distribution, and the effect of noise and depth changes on retrieval accuracy is a direct test of that boundary.
The stacked-image 'typical behaviour' criterion for time-series epochs could be turned into a real-time alert filter for survey streams, flagging only epochs that deviate beyond a dynamic threshold.
A direct comparison against self-supervised or astronomy-specific pretrained encoders would isolate how much of the success comes from ImageNet pretraining itself, an extension the authors gesture at when they mention improving the feature extractor.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

This is a proceedings-style recap, and the authors say so themselves in the intro: 'we recap some of our transfer learning applications in astronomy.' There is no new derivation, no new data, and no new experiment. If you are looking for a novel contribution, this is not it. The only fresh-looking element is a qualitative figure of 'weird' SDSS images, which has no metrics.

What the paper does well is summarize a coherent line of work built around a fixed ImageNet-pretrained EfficientNet feature extractor used for retrieval across four tasks: AGN candidacy, galaxy SFR/stellar mass estimation, strong-lens ranking, and time-series epoch flagging. The method description is clear, and the authors are honest about which results come from which prior paper. For someone wanting an entry point to ULISSE, this is a useful map.

The soft spots are the ones you would expect in a recap. The load-bearing assertion—that distance in the 1280-dimensional feature space corresponds to astrophysical similarity—is not defended here beyond a citation to [10]. The stress-test concern is real: with g,r,i cutouts, neighbors may be retrieved mainly on apparent brightness, size, and color, and a k-NN baseline on simple photometry would be a fair comparison. But that is a critique of the underlying method, not of this recap, because no new evidence is presented. The one quantitative result with a simulated-to-real gap, strong lensing, is acknowledged as work in progress. The galaxy-property numbers (<1 dex for >80% of targets) are quoted without error bars or sample definitions; again, the details are in the cited papers.

Who is this for? A reader who wants a five-page overview of what ULISSE can do, or a proceedings audience. As a research preprint it should be desk rejected; there is nothing here for a referee to check. That said, the paper is not misleading—it labels itself as a recap and gives credit to the source papers.

My recommendation: do not send this to peer review; it is a summary, not a research contribution, and the evidence that would need checking lives in the cited papers. For a proceedings volume that invites recaps, it is acceptable, but it does not deserve referee time.

Referee Report

4 major / 6 minor

Summary. This manuscript is a short proceedings paper that describes ULISSE, a transfer-learning framework built on a frozen ImageNet-pretrained EfficientNet feature extractor. The authors' method maps multi-band (g,r,i) astronomical cutouts into a 1280-dimensional feature space and uses distances in that space for nearest-neighbour similarity search. Four applications are summarized: AGN candidate detection in SDSS, estimation of galaxy stellar mass and star formation rate, strong gravitational lensing candidate search, and anomaly/artifact identification in image and time-series data. The central claim is that no labeled training is needed: labels are used only for assessment, and the same geometric feature space serves multiple downstream tasks. The paper reports summary statistics from the authors' previous ULISSE papers (33-60% AGN candidate rates, 20% false positives, <1 dex distance for more than 80% of galaxies, 94%/93% lensing purity) and notes an ongoing simulated-to-real gap for lensing.

Significance. If the central claim is correct, this would be a valuable and economical tool for survey-era astronomy because it promises useful astronomical inference from a frozen extractor without task-specific labeled training, which is especially relevant for LSST-scale data. The paper is commendably explicit about the role of labels (assessment only) and about the simulated-to-real gap in lensing, and the underlying architecture is easy to reproduce. The significance, however, is conditional: the evidence offered is a set of summary statistics quoted from the authors' own papers, with no baseline comparisons, no error bars, and no ablation isolating the contribution of ImageNet-pretrained features from low-level photometric properties. The paper therefore currently functions better as a research programme statement than as a demonstration that the method works as claimed.

major comments (4)

[Section 2.1 and Section 3.2] The load-bearing assertion that distances in the 1280-dimensional EfficientNet feature space rank galaxies by astrophysical similarity is not established, because no control is reported for the photometric proxies that the input g,r,i cutouts directly encode (total flux, angular size, colour, PSF, background subtraction). The galaxy-property results in Section 3.2 are equally consistent with a k-nearest-neighbour regressor on simple photometry; the reported '<1 dex for more than 80%' does not by itself demonstrate that transfer learning contributes anything beyond low-level photometric proximity. Please add an ablation using photometric features (magnitude, size, colours) as an explicit baseline, or demonstrate that feature-space neighbours remain astrophysically similar after controlling for these quantities.
[Section 3.1] The AGN detection evidence is quoted as 'on average higher than 33%' and '20% false positives', but the sample sizes, definitions of 'true' non-AGN, and the retrieval protocol (number of neighbours, threshold) are not given here, and no comparison is made to standard AGN colour-colour or emission-line selection. As written, the claim that the method is 'really competitive' cannot be evaluated from the numbers reported. Please give the complete contingency table or refer the reader to specific tables in [10] and restate the corresponding metrics.
[Section 3.3] The only application with a quantitative simulated result (94% purity in the first 100 candidates and 93% in the first 200) is also the one where the authors state that 'the transition to real observational data has presented additional challenges'. This tension needs to be addressed explicitly in the conclusions: currently the abstract and Section 3.6 generalize across applications as 'successful implementations', which overstates what the paper itself reports. Please either report quantitative results on real lensing candidates or explicitly limit the claim to simulated data.
[Section 3.4 and Section 3.5] The anomaly-detection applications are described without any quantitative evaluation: 'great potential', 'very few good epochs lost', and 'a lot of problematic images' are not measurable claims. Since the full studies are available ([20] and Fig. 4), please include the relevant evaluation metrics (e.g., precision/recall for flagged epochs, number of images inspected) or state clearly that these are proof-of-concept results with no quantitative validation.

minor comments (6)

[Section 2.1, footnote] 'classificator' should be 'classifier'.
[Figure 3 caption] 'Top forthy' should be 'Top forty'.
[Section 3.3] 'sosia' is an unusual non-standard term; use 'counterparts' or 'similar images'.
[References] References [11] and [15] lack author names and are formatted inconsistently; update to standard journal format.
[Figure captions] Figure captions say 'See [10] for further details' for reproduced figures; if figures are taken from other papers, permission and attribution should be stated.
[Section 3.2] The text says 'the averaged one for the set of the retrieved neighbours' but Fig. 2 also refers to a 'weighted mean'; define both averaging procedures and explain why they differ.

Circularity Check

0 steps flagged · score 1.0 of 10

No structural circularity: ImageNet-pretrained features are frozen and evaluated against external spectral labels; the evidence is self-cited but the cited experiments are externally falsifiable.

full rationale

The derivation chain is not circular in the structural sense. The feature extractor (EfficientNet, Section 2.1(1)) is pretrained on ImageNet and frozen; Section 2.2(1) states that 'there is no need for any labeled dataset in order to train a complex model; labels can be used just to assess the results', so no astronomical label enters the construction of the feature space. The downstream tasks are nearest-neighbour retrievals in this fixed space, and the reported evaluations compare against external measurements: AGN candidates against SDSS/BPT classifications (Section 3.1), and SFR/stellar-mass estimates against the MPA-JHU galSpec catalogue (Section 3.2), whose labels are not used to fit any parameter of the encoder. No equation in the paper reduces to its own input, and no fitted quantity is renamed as a prediction. The evidentiary base is, however, entirely composed of the authors' own prior works ([10], [16], [20]) plus one 'still in preparation' study (Section 3.3). These self-citations are real evidence rather than circular support because the cited experiments are externally falsifiable against independent spectroscopic and photometric catalogues, so they do not raise the circularity score under the stated rules. Section 3.3 itself acknowledges the simulated-to-real validation gap for lens finding, and the absence of ablations controlling for total flux, PSF, and angular scale is a correctness and confound concern about whether the ImageNet features or low-level photometric proxies drive the similarities; this is an empirical vulnerability, not a by-construction reduction. Overall, the paper is a self-cited proceedings summary with one minor non-load-bearing self-citation pattern, hence a score of 1.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities and no fitted parameter values. Its central claim rests on the transferability of ImageNet features, on the semantic validity of distances in feature space, and on user-chosen neighborhood and threshold settings that are not specified here. The numerical evidence is inherited from the authors' prior publications rather than produced in this preprint.

free parameters (2)

Number of retrieved neighbors or similarity threshold
The galaxy-property estimator in Section 3.2 averages properties over retrieved neighbors, but the paper does not state the neighborhood size or the distance threshold, and the reported accuracy depends on these choices.
Dynamic anomaly threshold
The time-series anomaly detector in Section 3.4 uses a dynamical threshold derived from a stacked reference image, but its definition and numerical setting are not specified.

assumptions (3)

domain assumption Convolutional features learned on ImageNet transfer to astronomical images without retraining.
Stated in Section 2.1(1) and Section 2.2(1); the whole framework depends on this transfer.
domain assumption Distance in the 1280-dimensional feature space is a meaningful measure of astrophysical similarity.
Section 2.1(3) defines similarity via distance between feature points; no validation of this metric against physical labels is given in this paper.
domain assumption A stacked image is a valid proxy for typical behavior of a time series, so deviations from it in feature space flag anomalies.
Section 3.4 defines the anomaly criterion as distance from the stacked image; this is assumed rather than tested here.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Leveraging Transfer Learning for Astronomical Image Analysis." pith.science (2026). https://pith.science/paper/YLKCJTNS

@misc{pith2026241118206,
  author       = {Pith},
  title        = {Pith review of: Leveraging Transfer Learning for Astronomical Image Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YLKCJTNS}},
  note         = {Machine review of arXiv:2411.18206}
}

read the original abstract

The exponential growth of astronomical data from large-scale surveys has created both opportunities and challenges for the astrophysics community. This paper explores the possibilities offered by transfer learning techniques in addressing these challenges across various domains of astronomical research. We present a set of recent applications of transfer learning methods for astronomical tasks based on the usage of a pre-trained convolutional neural networks. The examples shortly discussed include the detection of candidate active galactic nuclei (AGN), the possibility of deriving physical parameters for galaxies directly from images, the identification of artifacts in time series images, and the detection of strong lensing candidates and outliers. We demonstrate how transfer learning enables efficient analysis of complex astronomical phenomena, particularly in scenarios where labeled data is scarce. This kind of method will be very helpful for upcoming large-scale surveys like the Rubin Legacy Survey of Space and Time (LSST). By showcasing successful implementations and discussing methodological approaches, we highlight the versatility and effectiveness of such techniques.

Figures

Figures reproduced from arXiv: 2411.18206 by the authors.

**Figure 1.** Some examples of the application of ULISSE to the detection of AGNs. See [10], from which the image is extracted, for further details [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Left panel: The distributions of distances between the mean and the weighted mean M∗ (weighted for the heterogeneous distribution of galaxies in the SFR-M∗ parameter space in our primary sample) for the retrieved neighbors and the target object, i.e. Mtarg ∗ − ⟨Mneig ∗ ⟩ (by blue color) and Mtarg ∗ − Mweighted ∗ (by green color). Right panel: The same as on the left panel, but for SFRtarg − ⟨SFRneig⟩ and SFRtarg − S… view at source ↗

**Figure 3.** Top forthy images, in terms of frequency, proposed by [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: An example of the top 81 most weird images in a dataset extracted from SDSS, as it can be easily seen most of them contain some obvious problems [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages

[10]

Doorenbos, O

L. Doorenbos, O. Torbaniuk, S. Cavuoti et al. , ULISSE: A tool for one-shot sky November 28, 2024 1:36 ws-procs961x669 WSPC Proceedings - 9.61in x 6.69in cavuoti page 9 9 exploration and its application for detection of active galactic nuclei, A&A 666, p. A171 (October 2022)

work page 2024
[20]

Cavuoti, D

S. Cavuoti, D. De Cicco, L. Doorenbos et al., Identification of problematic epochs in astronomical time series through transfer learning, A&A 687, p. A246 (July 2024)

work page 2024
[1]

D. G. York, J. Adelman, J. E. Anderson, Jr. et al. , The Sloan Digital Sky Survey: Technical Summary, AJ 120, 1579 (September 2000)

work page 2000
[2]

J. T. A. de Jong, G. A. Verdoes Kleijn, D. R. Boxhoorn et al. , The first and second data releases of the Kilo-Degree Survey, A&A 582, p. A62 (October 2015)

work page 2015
[3]

Aihara, Y

H. Aihara, Y. AlSayyad, M. Ando et al. , Second data release of the Hyper Suprime- Cam Subaru Strategic Program, PASJ 71, p. 114 (December 2019)

work page 2019
[4]

Ivezi´ c, S

ˇZ. Ivezi´ c, S. M. Kahn, J. A. Tyson et al. , LSST: From Science Drivers to Reference Design and Anticipated Data Products, ApJ 873, p. 111 (March 2019)

work page 2019
[5]

Scaramella, J

R. Scaramella, J. Amiaux, Y. Mellier et al. , Euclid preparation-i. the euclid wide survey, Astronomy & Astrophysics 662, p. A112 (2022)

work page 2022
[6]

´Alvarez-M´ arquez, L

J. ´Alvarez-M´ arquez, L. Colina, R. Marques-Chaveset al. , Investigating the physical properties of galaxies in the Epoch of Reionization with MIRI/JWST spectroscopy, A&A 629, p. A9 (September 2019)

work page 2019

Show all 22 references

[7]

Baron, Machine Learning in Astronomy: a practical overview, arXiv e-prints , p

D. Baron, Machine Learning in Astronomy: a practical overview, arXiv e-prints , p. arXiv:1904.07248 (April 2019)

1904 arXiv
[8]

Longo, E

G. Longo, E. Mer´ enyi and P. Tiˇ no, Foreword to the Focus Issue on Machine Intelli- gence in Astronomy and Astrophysics, PASP 131, p. 100101 (November 2019)

2019
[9]

C. J. Fluke and C. Jacobs, Surveying the reach and maturity of machine learning and artificial intelligence in astronomy,WIREs Data Mining and Knowledge Discovery 10, p. e1349 (January 2020)

2020
[11]

PMLR, Efficientnet: Rethinking model scaling for convolutional neural networks 2019

2019
[12]

Schmidhuber, Deep learning in neural networks: An overview, Neural Networks 61, 85 (2015)

J. Schmidhuber, Deep learning in neural networks: An overview, Neural Networks 61, 85 (2015)

2015
[13]

W. S. McCulloch and W. Pitts, A logical calculus of the ideas immanent in nervous activity, Bulletin of Mathematical Biophysics 5, 115 (1943)

1943
[14]

Rosenblatt, The perceptron: A probabilistic model for information storage and organization in the brain, Psychological Review , 65 (1958)

F. Rosenblatt, The perceptron: A probabilistic model for information storage and organization in the brain, Psychological Review , 65 (1958)

1958
[15]

Ieee, Imagenet: A large-scale hierarchical image database 2009

2009
[16]

Torbaniuk, L

O. Torbaniuk, L. Doorenbos, M. Paolillo et al. , Ulisse: Determination of star- formation rate and stellar mass based on the one-shot galaxy imaging technique, A&A (2024 submitted)

2024
[17]

Brinchmann, S

J. Brinchmann, S. Charlot, S. White et al. , The physical properties of star-forming galaxies in the low-redshift universe, Monthly Notices of the Royal Astronomical So- ciety 351, 1151 (2004)

2004
[18]

J. D. Scargle, Astronomical time series analysis, in Astronomical Time Series , eds. D. Maoz, A. Sternberg and E. M. Leibowitz (Springer Netherlands, Dordrecht, 1997)

1997
[19]

Aigrain and D

S. Aigrain and D. Foreman-Mackey, Gaussian Process Regression for Astronomical Time Series, ARA&A 61, 329 (August 2023)

2023
[21]

Capaccioli and P

M. Capaccioli and P. Schipani, The VLT Survey Telescope Opens to the Sky: History of a Commissioning, The Messenger 146, 2 (December 2011)

2011
[22]

De Cicco, F

D. De Cicco, F. Bauer, M. Paolillo et al., A random forest-based selection of optically variable agn in the vst-cosmos field, Astronomy and Astrophysics 645 (2021)

2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [10]

Doorenbos, O

L. Doorenbos, O. Torbaniuk, S. Cavuoti et al. , ULISSE: A tool for one-shot sky November 28, 2024 1:36 ws-procs961x669 WSPC Proceedings - 9.61in x 6.69in cavuoti page 9 9 exploration and its application for detection of active galactic nuclei, A&A 666, p. A171 (October 2022)

work page 2024

[2] [20]

Cavuoti, D

S. Cavuoti, D. De Cicco, L. Doorenbos et al., Identification of problematic epochs in astronomical time series through transfer learning, A&A 687, p. A246 (July 2024)

work page 2024

[3] [1]

D. G. York, J. Adelman, J. E. Anderson, Jr. et al. , The Sloan Digital Sky Survey: Technical Summary, AJ 120, 1579 (September 2000)

work page 2000

[4] [2]

J. T. A. de Jong, G. A. Verdoes Kleijn, D. R. Boxhoorn et al. , The first and second data releases of the Kilo-Degree Survey, A&A 582, p. A62 (October 2015)

work page 2015

[5] [3]

Aihara, Y

H. Aihara, Y. AlSayyad, M. Ando et al. , Second data release of the Hyper Suprime- Cam Subaru Strategic Program, PASJ 71, p. 114 (December 2019)

work page 2019

[6] [4]

Ivezi´ c, S

ˇZ. Ivezi´ c, S. M. Kahn, J. A. Tyson et al. , LSST: From Science Drivers to Reference Design and Anticipated Data Products, ApJ 873, p. 111 (March 2019)

work page 2019

[7] [5]

Scaramella, J

R. Scaramella, J. Amiaux, Y. Mellier et al. , Euclid preparation-i. the euclid wide survey, Astronomy & Astrophysics 662, p. A112 (2022)

work page 2022

[8] [6]

´Alvarez-M´ arquez, L

J. ´Alvarez-M´ arquez, L. Colina, R. Marques-Chaveset al. , Investigating the physical properties of galaxies in the Epoch of Reionization with MIRI/JWST spectroscopy, A&A 629, p. A9 (September 2019)

work page 2019