SpecLess: A Multi Survey Python Package for Machine Learning Classification of Galaxy Morphology and Activity from Photometry

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: We're looking at the paper: "SpecLess: A Multi Survey Python Package for Machine Learning Classification of Galaxy Morphology and Activity from Photometry".

Jocelyn: Here's the gist of the paper. The gist The SpecLess package implements two photometric classification tasks for galaxy morphology and activity using machine learning classifiers trained on survey-specific photometry, demonstrating that catalogue dependence can affect classification performance even when object identities are controlled.

Vera: Model 1: Morphology Classification Model 1 predicts high-confidence Galaxy Zoo 1 elliptical and spiral classes using broad band photometry without explicit structural measurements The native morphology classifiers recover the highconfidence Galaxy Zoo 1 elliptical and spiral reference classes with macro F1 of 0.9332 on 35,554 SDSS+AllWISE test galaxies and 0.9237 on 3,269 KiDS+VIKING test galaxies The integrated magnitudes and colours therefore carry substantial information about the selected visual morphology labels, even though spiral arms and other structural properties are not supplied to the classifiers.

Jocelyn: Controlled Paired Survey Experiment (Morphology) The controlled paired survey experiment evaluates whether photometric information remains predictive under a change in how the same galaxies are measured XGBoost achieves similar within-survey macro F1 in SDSS (0.9049) and KiDS (0.9065), but its transferred predictions are less accurate in both directions.

Vera: The signed source-relative differences FA→B − FA→A are −0.0137 ± 0.0067 for SDSS→KiDS and −0.0370 ± 0.0104 for KiDS→SDSS.

Title and authors: Tom: So, this paper is introducing SpecLess, which is this Python package that lets you classify galaxy morphology and activity just using photometry from different surveys, without needing spectroscopy at inference time

Vera: Right. So what's the core idea here? It's not just training one model; it’s building a system where you can use these pretrained models across totally different catalogues

Jocelyn: And the big takeaway is that we have to be careful when we move our galaxy samples around because how you measure them changes what the AI learns

Vera: Exactly. They show that even when you control for the object itself, if the catalogue changes, your classification performance shifts in a predictable way

Jocelyn: They did this by running paired experiments where they compare SDSS data to KiDS data on morphology and activity classes

Subrahmanyan: And what they found is that for morphology, XGBoost gets similar accuracy in both surveys, but the predictions transferred between them are less accurate in both directions

Vera: That means there’s a quantifiable performance deficit when you try to apply a model trained on one survey to another one directly

Jocelyn: And for activity classification, they found that the recovery of star-forming and starburst classes is stronger, but there's still this confusion between them because the photometric features overlap too much

Subrahmanyan: It points toward a physical dependence here, as they showed that the STARBURST classification error is strongly linked to H-alpha equivalent width, which tells us what’s driving that specific label assignment

Vera: They also tested how sensitive these models are to changes in input features by looking at interaction terms, and those numbers show that colour-only refits don't always fix the transfer loss

Jocelyn: So the authors are essentially giving us a reusable tool for classification, but they’re also clearly flagging that catalogue dependence is a real issue we have to account for when using these models across different survey volumes

Subrahmanyan: It sets clear observational limits on what we can infer from photometry alone, and it shows us exactly where the catalogue representation becomes a crucial part of our validation process

Vera: That’s the main point. SpecLess gives us the models, but it also forces us to be explicit about which survey we are using and how that affects the results

Jocelyn: It’s about making sure we understand exactly what those numbers mean for real-world science and not just looking at a single accuracy score

Subrahmanyan: This means if you want to use these tools for structure formation, you can't just assume the classification works everywhere; you have to account for that asymmetry in transfer deficits between catalogues

Vera: This means if we want to use these tools for cosmology, you have to account for that asymmetry in transfer deficits between catalogues because it dictates how much the model might drift depending on where you are observing

The paper's summary: Tom: So, we've looked at how SpecLess works across different surveys for morphology and activity classification >

Vera: Right. Now, the paper isn't just stopping there; they suggest ways to actually improve this package itself >

Jocelyn: What kind of improvements are they talking about? Are we getting better models or better testing methods?

Vera: They're focusing on making it a more reusable interface and providing explicit diagnostics for when you need catalogue-specific evaluation > They're showing that colour-only refits don't always reduce those transfer losses, so you have to use more complex input constructions and nonadditive interactions

Jocelyn: That means we can’t just simplify the input features down to just colours and expect it to work universally across all survey types

Subrahmanyan: It suggests that the learned response of the AI is highly sensitive to how those input features are combined, which is important for connecting observations to physical structure

Vera: They also pointed out how feature ablations reveal complementary contributions from both optical and mid-infrared photometry when we look at activity classes > This suggests we can’t rely on just one band to tell the whole story for things like starburst activity

Jocelyn: So it’s a reminder that we can’t rely on just one band to tell the whole story for things like starburst activity

Subrahmanyan: This pushes us toward needing more diverse input data, which is something we see in other papers discussing dust or PAH properties in simulations

Vera: Ultimately, SpecLess is positioned as a documented model interface, but it’s also an experimental guide showing us the observational limits of what we can do with photometric classification alone

Jocelyn: It gives us the reusable models while making sure we're honest about where those models might fail when we change our survey selection function

Subrahmanyan: This means for structure formation, if you want to use these tools, you have to account for that asymmetry in transfer deficits between surveys because it dictates how much the model might drift depending on where you are observing

The paper's improvements: Tom: So we've seen how SpecLess uses machine learning to classify galaxy morphology and activity using just photometry from different surveys >

Vera: And we’ve discussed the catalogue dependence, seeing how performance dips when you transfer predictions between SDSS and KiDS data sets

Jocelyn: And Subrahmanyan shows that if you want to use these photometric tools for structure formation, we need to be very explicit about the catalogue dependence because it dictates how much the model might drift depending on where you are observing

Subrahmanyan: It shows that if you want to use these photometric tools for cosmology, you can't just assume the classification works everywhere; you have to account for that asymmetry in transfer deficits between catalogues

Vera: Exactly. SpecLess gives us a reusable interface and an experimental way to test those limits across morphology and activity classifications

Jocelyn: It’s a tool that lets us use pretrained models, but it’s also an explicit demonstration of when catalogue-specific evaluation is necessary for the science we're doing

Subrahmanyan: That means if you want to use these tools for cosmology, you can't just assume the classification works everywhere; you have to account for that asymmetry in transfer deficits between catalogues

Vera: We’ve established that this package has defined observational limits based on the data we feed it, which is really important when we are trying to constrain physical properties of galaxies

Jocelyn: It’s not just about getting a number; it’s about understanding the observational limits of what we can infer from photometry alone

Subrahmanyan: This paper "SpecLess: A Multi Survey Python Package for Machine Learning Classification of Galaxy Morphology and Activity from Photometry" provides that explicit demonstration of where our assumptions need to be checked

Conclusion: Vera: So we've seen how SpecLess uses machine learning to classify galaxy morphology and activity using just photometry from different surveys >

Jocelyn: And we’ve discussed the catalogue dependence, seeing how performance dips when you transfer predictions between SDSS and KiDS data sets >

Subrahmanyan: It shows that if you want to use these photometric tools for structure formation, we need to be very explicit about the catalogue dependence because it dictates how much the model might drift depending on where you are observing >

Vera: Exactly. SpecLess gives us a reusable interface and an experimental way to test those limits across morphology and activity classifications >

Jocelyn: It’s a tool that lets us use pretrained models, but it’s also an explicit demonstration of when catalogue-specific evaluation is necessary for the science we're doing >

Subrahmanyan: That means if you want to use these tools for cosmology, you can't just assume the classification works everywhere; you have to account for that asymmetry in transfer deficits between catalogues >

Vera: We’ve established that this package has defined observational limits based on the data we feed it, which is really important when we are trying to constrain physical properties of galaxies >

Jocelyn: It’s not just about getting a number; it’s about understanding the observational limits of what you can infer from photometry alone >

Subrahmanyan: This paper, "SpecLess: A Multi Survey Python Package for Machine Learning Classification of Galaxy Morphology and Activity from Photometry," provides that explicit demonstration of where our assumptions need to be checked >

Vera: It’s a tool with defined observational limits based on the data we feed it, which is really important when we are trying to constrain physical properties of galaxies >

Jocelyn: So the main thing is that this paper shows us how catalogue representation affects classification, and it gives us a reusable interface for these models >

Subrahmanyan: We have to keep in mind that the performance numbers are tied to specific test populations, and those results are not direct benchmarks for all experiments considered here, which is a crucial caveat >

Vera: And they did address that by showing the controlled paired survey experiment where XGBoost achieved similar macro F1 scores in both SDSS and KiDS but with less accurate transferred predictions >

Jocelyn: That shows how important it is to test on matched galaxies with identical labels and data partitions >

Subrahmanyan: If you want to use this for cosmology, you have to account for that asymmetry in transfer deficits—the larger target baseline from SDSS to DESI—because it dictates how much the model might drift depending on where you are observing >

Vera: So SpecLess is essentially a tool that lets us use these machine learning classifiers while being very honest about where they might fall short when we move data around >

Jocelyn: Right. We've seen how much more complex the input needs to be than just simple colors to get reliable results when you want that level of performance across different surveys >

Subrahmanyan: The paper provides a reusable interface to documented pretrained models and an explicit experimental demonstration of when catalogue-specific evaluation is necessary >

Vera: It defines the demonstrated capabilities and the observational limits of photometric galaxy classification in those reference samples considered here, which is a really important thing for anyone using these tools >

Jocelyn: So, we’ve seen how survey differences mess with performance even when we control for object identity >

Subrahmanyan: For structure formation, this means that if you want to use these photometric tools to study large-scale structure, you need to be very explicit about the catalogue dependence and the selection function of your input data >

Vera: It’s a demonstration of how observational constraints dictate what we can actually learn from photometry without needing expensive spectroscopy for every single object >

Jocelyn: And it points toward the need for more robust ways to handle those differences between catalogues >

Subrahmanyan: We've seen how much more complex the input needs to be than just simple colors to get reliable results when you want that level of performance across different surveys >