(r6) InterOpNov2021KD < IVOA

IVOA Web>WebPreferences>InterOpNov2021>InterOpNov2021KD (revision 6) (raw view)~~EditAttach~~

---++ Knowledge Discovery 1

Time: [[https://www.timeanddate.com/worldclock/fixedtime.html?msg=KD&iso=20211103T2200][Wednesday Nov 03 22:00 UTC]]

| <p><span style="background-color: transparent;">Raffaele D'Abrusco</span></p> | <p> *Introduction* </p> | 5' | pdf |
| <p><span style="background-color: transparent;">Ashish Mahabal</span></p> | <p> *Data Sheets and Model Cards* </p> <p>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 &ldquo;is&rdquo; 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).</p> | 7' | [[%ATTACHURL%/Mahabal_IVOA_20211103.pdf][pdf]] |
| _Petr Skoda_ | <p> *SDSS redshift prediction based on Bayesian Deep Learning* </p> <p>Bayesian deep learning is a relatively new approach that starts to enter <span style="background-color: transparent;">the astronomy. Unlike majority of the current methods it does provide the </span><span style="background-color: transparent;">uncertainty of its predictions. So we can visually check the suspicious </span><span style="background-color: transparent;">cases with high uncertainty. We demonstrate this in the experiment with </span><span style="background-color: transparent;">spectroscopic redshift prediction from SDSS quasar catalogues .This </span><span style="background-color: transparent;">allowed us to find a number of quasars which are probably normal stars </span><span style="background-color: transparent;">with wrong estimate of redshift from the SDSS pipeline.</span></p> | 7' | [[%ATTACHURL%/skoda-bayesian-redshift.pdf][pdf]] |
| _Rafael Martinez Galarza_ | <p> *Harvesting outliers: data barriers to turn anomalies into discoveries* </p> <p>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.</p> | 7' | [[%ATTACHURL%/KD-IG_anomalies.pdf][pdf]] |

Moderator: [[Raffaele D'Abrusco][Raffaele D'Abrusco]], Notetaker: [[IVOA.TBD][TBD]], [[https://yopad.eu/p/IVOA_Nov3_KD][Etherpad link]]

Attachments

Topic attachments
I	Attachment	History	Action	Size	Date	Who
pdf	KD-IG_anomalies.pdf	r1	manage	34742.0 K	2021-11-03 - 13:21	RaffaeleDAbrusco
pdf	Mahabal_IVOA_20211103.pdf	r1	manage	59.2 K	2021-11-03 - 20:17	RaffaeleDAbrusco
pdf	skoda-bayesian-redshift.pdf	r1	manage	1812.3 K	2021-11-03 - 20:14	RaffaeleDAbrusco

Topic revision: r6 - 2021-11-03 - RaffaeleDAbrusco

IVOA

Log in or Register

IVOA.net
Wiki Home
WebChanges
WebTopicList
WebStatistics

Twiki Meta & Help
IVOA
Know
Main
Sandbox
TWiki

TWiki intro
TWiki tutorial
User registration
Notify me

Working Groups

Interest Groups

Committees

Stds&Procs

www.ivoa.net
Documents
Events
Members
XML Schema