Historical Deadlines and Objectives

ICLR Climate Workshop February 4th Most likely ICML Climate Workshop 4/1

Timeline:

• Workshops announced ICML: 3/7
• Submission deadline ICML: ?
• Conference: July 12-18th in Austria

Paper goals:

• Have dataset ready and fully prepared (this includes soil depth, lat/long, elevation, and other catchment attributes from CAMELs)
• Have simple LSTM, DA-RNN, and Transformer baseline
• Finalize evaluation metrics
• Explore if one neural network given lat and longitude can predict for all streams/rivers (i.e. one network for them all v.s. separate network for each)
• Explore (if there is time) different configurations of transformer (i.e 1DConv vs LinearLayer etc)

Transfer learning and meta-data integration on time series paper (May 23 for Neurips)

• Ideally this would be complete for ICML deadline of 2/1 Neurips main conference track. Goal for this paper would be publication in main track of major ML conference not only workshop track.
• Evaluate the effects of pre-training on solar/wind data and other "semi-related" time series data
• Evaluate if neural networks trained on specific basins in one part of the country can effectively predict basins in other regions of the country given only limited time series data. Explore exactly where this threshold is (i.e. if pre-trained on n rivers is training on 1 year of data, 1 week, 1 day, enough for target river???).
• Visualize dimension of attention activations over different rivers and how it changes on different tasks
• Investigate if pre-taining on flow data could improve neural network performance on other time series data (i.e wind, solar, even medical). Transfer learning has proven highly effective in for DL on images (even completely unrelated) would be good to study on time series.
• Create a effective multimodal architecture for integrating static attributes with dynamic information.

Flash flood damage forecasting/Dataset (i.e. task 2)

• Build dataset for forecasting damage done by flash floods.
• Construct out of aerial imagery + previously built time series dataset + flash flood (SNAP) data
• Study multimodal deep learning models that incorporate static aerial images with dynamic time series data
• If possible also predict maximum flood depth (though don't know of source for this data)