Programme by Session

For regression tasks, CNN typically provide only point estimates with no uncertainty, leading to overconfident predictions and limiting the scientific value of such methods.

We show that Bayesian CNN and generative label models can be combined to predict posteriors over all regression targets. This approach is highly general, with potential applications including fast radio burst detection and strong lensing mass estimation.

We apply this approach to Galaxy Zoo, predicting posteriors for the exact (as opposed to majority) responses of citizen scientists. We show that these posteriors are well-calibrated and hence are reliable for practical use in galaxy evolution research.

By predicting posteriors, we can identify which subjects would, if labelled, be most informative for our model (active learning). Using our Galaxy Zoo posteriors, we simulate iteratively requesting citizen scientist responses and retraining our model. We show that active learning significantly improves model performance given limited citizen scientist effort. This will allow researchers to classify morphologies in new surveys of any size on a timescale of weeks.