Galaxy Zoo: Probabilistic Morphology through Bayesian CNNs and Active Learning (1905.07424v2)

Abstract: We use Bayesian convolutional neural networks and a novel generative model of Galaxy Zoo volunteer responses to infer posteriors for the visual morphology of galaxies. Bayesian CNN can learn from galaxy images with uncertain labels and then, for previously unlabelled galaxies, predict the probability of each possible label. Our posteriors are well-calibrated (e.g. for predicting bars, we achieve coverage errors of 11.8% within a vote fraction deviation of 0.2) and hence are reliable for practical use. Further, using our posteriors, we apply the active learning strategy BALD to request volunteer responses for the subset of galaxies which, if labelled, would be most informative for training our network. We show that training our Bayesian CNNs using active learning requires up to 35-60% fewer labelled galaxies, depending on the morphological feature being classified. By combining human and machine intelligence, Galaxy Zoo will be able to classify surveys of any conceivable scale on a timescale of weeks, providing massive and detailed morphology catalogues to support research into galaxy evolution.

• The paper introduces Bayesian CNNs to probabilistically predict galaxy morphology while quantifying uncertainty in crowd-sourced labels from Galaxy Zoo.
• The Bayesian approach yields well-calibrated probabilistic predictions that achieve significantly lower RMS error compared to deterministic methods by leveraging label variability.
• Bayesian Active Learning by Disagreement (BALD) is employed to selectively request labels for the most informative galaxies, reducing the need for manual labelling by up to 60%.

Galaxy Classification with Bayesian CNNs and Active Learning

The paper, "Galaxy Zoo: Probabilistic Morphology through Bayesian CNNs and Active Learning," presents a sophisticated methodology to address galaxy morphology classification using a combination of Bayesian convolutional neural networks (CNNs) and active learning strategies. This work is pivotal for handling the burgeoning data from astronomical surveys, which traditional visual classification methods cannot efficiently process due to their sheer volume.

Bayesian CNNs for Galaxy Morphology

The research introduces Bayesian CNNs capable of predicting posteriors for galaxy morphology, which inherently account for the uncertainty in volunteer responses gathered through the Galaxy Zoo project. This approach tackles the challenge posed by uncertain and variable labelling, often encountered in crowd-sourced data collections. By employing a novel generative model and MC Dropout for uncertainty quantification, the Bayesian CNNs escape the limitations of deterministic CNNs that provide single-point estimates. The Bayesian framework mathematically integrates the uncertainty in model weights and the variability in labels, thereby offering a probabilistic interpretation that enhances both the training process and the reliability of morphological classifications.

Probabilistic Predictions and Calibration

The strength of this methodology lies in the ability to generate probabilistic predictions for categorising galaxies into morphological types, such as "Smooth or Featured" and "Bar", through well-calibrated posterior distributions of volunteer vote fractions. The Bayesian CNNs leverage the variability in these labels and achieve a root-mean-square error significantly below deterministic approaches, making them highly reliable for subsequent statistical analysis. The introduction of a calibration procedure ensures that these probabilities can be robustly interpreted, addressing common issues related to overconfident predictions in neural networks.

Active Learning with BALD

A notable contribution of the paper is using Bayesian Active Learning by Disagreement (BALD) to reduce the necessity for manually labelled data. By calculating the mutual information between model parameters and predictions, the system identifies the most informative galaxies for labelling. This process allows the learning algorithm to request additional labels selectively, thus enhancing learning efficiency and drastically reducing the number of necessary labelled examples by up to 60% for certain morphological features. The method optimally balances exploration and exploitation, targeting galaxies that contribute most toward improving model accuracy.

Implications and Future Directions

The integrated use of Bayesian CNNs and active learning offers a scalable solution to real-time galaxy classification, positioning this approach as a significant advancement for ongoing and future sky surveys like Euclid and LSST. The probabilistic nature of the predictions not only ensures robust scientific inferences but also highlights the potential for combining machine learning with citizen science, forming a symbiotic relationship that enhances both computational and human efforts in scientific exploration.

In future developments, fine-tuning dropout rates for even better calibration, employing deeper architectures for improved feature representation, and leveraging domain adaptation techniques to cross-apply these models to different surveys can further enhance performance. As AI methods continue to evolve, integrating human intuition and machine accuracy through active learning paradigms will be imperative for extracting meaningful insights from vast astronomical datasets, potentially redefining our approach to understanding cosmic phenomena.