Table Stakes Discovery “Did something vary in this place in the sky?” (aka Real-Bogus) 0-1 Classification - Feature-based+tree - Learned Representation+neural Characterization “What might this event be?” Multi-class Classification - Learned Representation+neural - Semi-supervised neural Already being done with LSST precursors, contemplated in LSST Brokers Data Cleaning “What would my image look like without Cosmic rays?” “Was my image corrupted by asteroid streaking, Starlink?” Image Masking/Inpainting - UNet Zhang, Bloom 20 (1907.09500) Duev+19 (1904.05920) Brink+12 (1209.3775) Goldstein+15 (1504.02936) J. Bloom ML Club, Feb 2021 @profjsb Richards+12 (1204.4180) review: Jamal & Bloom (2003.08618) 

Forefront/Challenges ML-Assisted Exploration Robust Out-of-sample/Anomaly Detection - Find new types of events - Identify bad data (in real time) Need probability-calibrated classifications in a full Bayesian context, that makes explicit the priors. Publish full likelihoods. Calibrated Classification Novelty Discovery J. Bloom ML Club, Feb 2021 @profjsb - Critical for Resource Allocation Similarity searches at scale, recommendation engines - Accelerate exploration for LSST users ML Ops Model management, provenance tracking - Reproducible science Physical Inference Parameter fitting with Neural Density Estimation, Likelihood-free inference - Rapid and accurate Posteriors with amortized neural models Zhang+21 (2010.04156) 

“Ask Not What Your Country can do for You. Ask, what you can do for your Country.” ML can do for LSST. LSST can do for ML. J. Bloom ML Club, Feb 2021 @profjsb