| |
 |
| Paper: |
Machine Learning Bias and the Annotation of Large Databases of Astronomical Objects |
| Volume: |
538, ADASS XXXII |
| Page: |
162 |
| Authors: |
Hunter Goddard; Lior Shamir |
| DOI: |
10.26624/BFMV6514 |
| Abstract: |
One of the common approaches to annotating astronomical databases is
by applying machine learning (ML), and specifically artificial neural networks (ANNs).
But while ANNs can be invaluable for astronomy, they also have several downsides.
Here, we study the possible disadvantages of ANNs. Our results show that when using
ML, the annotations can have subtle but consistent biases. These biases are very difficult
to detect, can change in different parts of the sky, and are not intuitive for the users
of data products annotated by ANNs. Since these catalogs are, in many cases, very
large, these subtle biases can lead to statistically significant observations that are the
result of the neural network bias rather than a true reflection of the Universe. Based
on these observations, catalogs annotated by current ANNs should be used cautiously,
and statistical observations enabled by such catalogs should be analyzed in light of
possible biases in the machine learning systems. The results reinforce the need for
further research on explainable neural network architectures. |
|
|
|
|
|
