|
META TOPICPARENT |
name="InterOpNov2021" |
|
|
< < | Knowledge Discovery 1 |
> > | Knowledge Discovery |
| |
|
< < | Time: Wednesday Nov 03 22:00 UTC
Raffaele D'Abrusco |
Introduction |
5' |
pdf |
Ashish Mahabal |
Data Sheets and Model Cards Astronomy datasets have been growing, and so are the attempts to use them wth a variety of machine learning techniques. While we would like to use all data, data fusion for diverse uneven or not-fully-matched datasets can be a challenge. Creating machine learning and artificial intelligence models for such datasets and follow-up validation can be challenging owing to lack of large labeled training datasets. To address this two related concepts that have emerged recently in data science are that of Data Sheets for data sets (Gebru et al. arXiv:1803.09010), and model cards for models (Mitchell et al., arXiv:1810.03993). This is just like each component in the electronics industry comes with a datasheet that describes operating characteristics, test results, recommended use etc., We recommend that for each astronomy dataset uniform and standardized datasheets that advertise similar meta-properties should be created, not just stating what “is” but also where each of the dataset could go, much like lego-blocks. This will enable data fusion, and also thwart mis-guided use of datasets. Similarly the models that we build will carry not just the usual provenance, but explicit characteristics displaying known biases and hence added caution when being used in certain ways. While this trend started in social fields where bias is explicit, it has been successfully applied in the Planetary Data System (PDS) setup for identifying key descriptors in an equally diverse dataspace (https://pds.nasa.gov/datastandards/documents/im/v1/index_1G00.html#10.31%C2%A0%C2%A0class_pds_observation_area). |
7' |
pdf |
Petr Skoda |
SDSS redshift prediction based on Bayesian Deep Learning Bayesian deep learning is a relatively new approach that starts to enter the astronomy. Unlike majority of the current methods it does provide the uncertainty of its predictions. So we can visually check the suspicious cases with high uncertainty. We demonstrate this in the experiment with spectroscopic redshift prediction from SDSS quasar catalogues .This allowed us to find a number of quasars which are probably normal stars with wrong estimate of redshift from the SDSS pipeline. |
7' |
pdf |
Rafael Martinez Galarza |
Harvesting outliers: data barriers to turn anomalies into discoveries Over the last few years astronomers have become increasingly effective at identifying anomalous objects in large astronomical datasets. So far, that has meant "finding objects in sparsely populated regions of a multidimensional feature space". This is done using a number of methods that includes ensemble methods such as random forests searches, and more recently generative models that identify anomalies as those objects are more difficult to reconstruct by the trained model. This has produced huge lists of anomalies in diverse datasets that include SDSS galaxy spectra, Kepler and TESS light curves, and X-ray catalogs. Yet, most of those anomalies are not followed up, because of a cultural difficulty for scientists to interpret multi-dimensional scatter plots that have no labels in their axes. We argue that such cultural barrier can be overcome with novel ways to combine domain knowledge expertise with data visualization, or even incorporating domain knowledge directly into the anomaly detection algorithms. We would like to discuss ways in which VO tools can help in the identification of anomalies that represent true astronomical discoveries, by harvesting the currently publicly available catalogs of anomalies. |
7' |
pdf |
Moderator: Raffaele D'Abrusco, Notetaker: TBD, Etherpad link |
> > | Time: Tuesday May 09 11:00 UTC
Speaker |
Title |
Time |
Abstract |
Material |
Raffaele D'Abrusco |
Greetings and Introduction |
10' |
|
pdf |
Sandor Kruk |
Exploring astronomy data archives at large scales using deep learning and crowdsourcing |
15'+5' |
The vast amount of data in astronomy archives presents an opportunity for new discoveries. Deep learning combined with crowdsourcing provides an efficient way to explore this data using. the intuition of the human brain and the processing power of machines. In the Hubble Asteroid Hunter project, we used a deep learning algorithm on Google Cloud, trained on volunteer classifications from the asteroidhunter.org Zooniverse project to search two decades of Hubble Space Telescope (HST) observations from the ESA HST archives for objects not targeted by the Hubble observations. The project, which was set up as a collaboration between Zooniverse, ESAC Science Data Center and engineers at Google, led to the discovery of 1700 asteroids (Kruk et al. 2022), including 1031 previously unknown ones (Garcia Martinet al., in prep.), 198 new strong gravitational lenses (Garvin et al. 2022), and quantified the impact of artificial satellites on HST observations (Kruk et al. 2023). In this talk, we will present the results of this project and highlight the benefits of scientifically exploiting the vast amounts of data available in astronomy data archives using novel techniques. |
pdf |
Raffaele/Yihan |
Introduction to mini-session on generative AI and language models |
5' |
|
pdf |
Yihan Tao |
Foundation models for Astronomy |
5' |
|
pdf |
Rafael Galarza-Martinez |
Intro to Transformers |
10' |
|
pdf |
Ioana Ciucă |
Galactic ChitChat: Using Large Language Models to Engage with Astronomy Literature |
5' |
We showcase the capacity of the OpenAI's large language model GPT-4 for meaningful engagement with Astronomy papers using in-context prompting. We employ a distillation technique to optimise efficiency, reducing the input paper size by 50% while preserving paragraph structure and semantic integrity. The in-context model excels at providing detailed answers contextualised by related research findings by examining its responses within a multi-document context (ten distilled documents). Our investigation highlights the potential of foundation models for the astronomical community. For example, they can help researchers gain insights from astronomical literature, such as validating new scientific hypotheses or proposing novel ideas. |
pdf |
|
|
> > |
Panel + audience |
Discussion - panel comprising speakers |
30' |
|
pdf |
Moderator: Raffaele D'Abrusco, Notetaker: TBD, Etherpad link |
|
META FILEATTACHMENT |
attachment="KD-IG_anomalies.pdf" attr="" comment="" date="1635945677" name="KD-IG_anomalies.pdf" path="KD-IG_anomalies.pdf" size="35575806" user="RaffaeleDAbrusco" version="1" |
META FILEATTACHMENT |
attachment="skoda-bayesian-redshift.pdf" attr="" comment="" date="1635970467" name="skoda-bayesian-redshift.pdf" path="skoda-bayesian-redshift.pdf" size="1855828" user="RaffaeleDAbrusco" version="1" |
META FILEATTACHMENT |
attachment="Mahabal_IVOA_20211103.pdf" attr="" comment="" date="1635970664" name="Mahabal_IVOA_20211103.pdf" path="Mahabal_IVOA_20211103.pdf" size="60610" user="RaffaeleDAbrusco" version="1" |
META FILEATTACHMENT |
attachment="slides_session_KDIG_IVOA_2021FallInterOp.pdf" attr="" comment="" date="1636049545" name="slides_session_KDIG_IVOA_2021FallInterOp.pdf" path="slides_session_KDIG_IVOA_2021FallInterOp.pdf" size="197990" user="RaffaeleDAbrusco" version="1" |
|