Detecting Outliers in SDSS using Convolutional Neural Network

Kaushal Kumar Sharma, Ajit Kembhavi, Aniruddha Kembhavi, T. S. Sivarani, Sheelu AbrahamBulletin 2019 · DOI

Abstract

We propose an automated algorithm based on Convolutional Neural Network (CNN) for the detection of peculiar objects in large databases using their spectral observations. A convolutional neural network is a class of deep-learning algorithms which allows the detection of significant features/patterns in sequential data like images, audio, time-series etc. by applying convolutional neurons (kernels) along the sequence. For detecting unusual spectra, we use eight-layer deep convolutional network with autoencoder architecture on ~ 60,000 spectra collected from the Sloan Digital Sky Survey. The training of the network is done in an unsupervised manner. We show that the trained network is able to retrieve the spectra of rare objects from a large collection of spectra. Such algorithms can easily be rescaled to other surveys and therefore can serve as a potential component of the data reduction pipelines for automatically detecting spectra with unusual features and recovering defective spectra.

Introduction

Modern astronomical surveys, such as the Sloan Digital Sky Survey (SDSS) and LAMOST, have amassed millions of optical spectra. While automated pipelines effectively classify common objects like stars, galaxies, and quasars, they often struggle with rare, "peculiar" objects that deviate from standard templates. These anomalies may include rare stellar types or observations affected by technical artifacts like cosmic-ray hits or calibration errors. Manually identifying these outliers in datasets of millions is impractical. This research addresses the need for an unsupervised, automated method to detect anomalies without requiring pre-labeled datasets of rare objects.

Methodology

The core of the proposed pipeline is a Deep Convolutional Autoencoder (DCNN). An autoencoder is designed to compress input data into a lower-dimensional "latent space" (the encoder) and subsequently reconstruct the original input (the decoder).

Model Architecture

The model utilizes an 8-layer deep architecture consisting of four encoding layers and four decoding layers. Unlike standard CNNs used for 2D images, this model employs 1D convolutional kernels to process sequential spectral data (flux vs. wavelength).

Architecture of the 8-layer deep convolutional autoencoder used for spectral reconstruction.

The "outlierness" of a spectrum is quantified by its Reconstruction Error, specifically Mean Squared Error (MSE) and Root Mean Squared Error (RMSE). The underlying logic is that a model trained on "normal" spectra will struggle to accurately reconstruct "peculiar" spectra it has never encountered.

Data Preparation

The study utilized data from SDSS Data Release 13 (DR13), specifically 61,627 stellar spectra with a signal-to-noise ratio (SNR) > 20. Spectra were normalized at 5550 Å and resampled to a 5 Å wavelength step within the 4000–7200 Å range, resulting in 640 data points per spectrum. The dataset was split into 55,000 spectra for training/validation and 6,627 for testing.

Results and Discussion

The model demonstrated high fidelity in reconstructing standard stellar classes (O, B, A, F, G, K, M), achieving an average RMSE of 0.0222.

Comparison of original and reconstructed spectra for three different stellar classes, showing high reconstruction fidelity and low residuals.

Discovery of Rare Objects

By examining spectra with the highest reconstruction errors, the researchers successfully identified scientifically significant outliers, including Cataclysmic Variables (characterized by intense emission lines) and Carbon Stars. This confirms that the model's reconstruction failure serves as an effective filter for extraordinary objects.

Efficiency and Robustness

The training process is highly efficient, completing in just 314 seconds on a single GPU node, which suggests high scalability for larger future surveys. Furthermore, the model proved robust against moderate noise; while lower SNR spectra naturally exhibited slightly higher errors, they did not produce significant false positives. The method also proved effective at identifying data quality issues, such as improper cosmic-ray corrections.

Conclusion and Future Work

This unsupervised approach allows for the discovery of novel object types without predefined templates. While the current workflow involved a manual cleaning step to ensure the training set represented "normal" behavior, future iterations could explore truly end-to-end unsupervised learning.

Potential future directions include utilizing the "latent space" for semi-supervised classification or implementing "denoising autoencoders" to improve the signal-to-noise ratio of observed spectra, thereby enhancing subsequent chemical and physical analyses.

Summary updated

Abstract

We propose an automated algorithm based on Convolutional Neural Network (CNN) for the detection of peculiar objects in large databases using their spectral observations. A convolutional neural network is a class of deep-learning algorithms which allows the detection of significant features/patterns in sequential data like images, audio, time-series etc. by applying convolutional neurons (kernels) along the sequence. For detecting unusual spectra, we use eight-layer deep convolutional network with autoencoder architecture on ~ 60,000 spectra collected from the Sloan Digital Sky Survey. The training of the network is done in an unsupervised manner. We show that the trained network is able to retrieve the spectra of rare objects from a large collection of spectra. Such algorithms can easily be rescaled to other surveys and therefore can serve as a potential component of the data reduction pipelines for automatically detecting spectra with unusual features and recovering defective spectra.
BibTeX
@article{sharma2019detecting,
  author = {Kaushal Kumar Sharma and Ajit Kembhavi and Aniruddha Kembhavi and T. S. Sivarani and Sheelu Abraham},
  title = {{Detecting Outliers in SDSS using Convolutional Neural Network}},
  journal = {Bulletin},
  year = {2019},
  doi = {10.25518/0037-9565.8811},
}
Figure