In this study, we combined a machine learning pipeline and human supervision to identify and label swallow and martin roost locations on data captured from 2000 to 2020 by 12 Weather Surveillance Radars in the Great Lakes region of the US. We employed radar theory to extract the number of birds in each roost detected by our technique. With these data, we set out to investigate whether roosts formed consistently in the same geographic area over two decades and whether consistency was also predictive of roost size. We used a clustering algorithm to group individual roost locations into 104 high‐density regions and extracted the number of years when each of these regions was used by birds to roost. In addition, we calculated the overall population size and analyzed the daily roost size distributions. Our results support the hypothesis that more persistent roosts are also gathering more birds, but we found that on average, most individuals congregate in roosts of smaller size. Given the concentrations and consistency of roosting of swallows and martins in specific areas throughout the Great Lakes, future changes in these patterns should be monitored because they may have important ecosystem and conservation implications.

Organisms have been shifting their timing of life history events (phenology) in response to changes in the emergence of resources induced by climate change. Yet understanding these patterns at large scales and across long time series is often challenging. Here we used the US weather surveillance radar network to collect data on the timing of communal swallow and martin roosts and evaluate the scale of phenological shifts and its potential association with temperature. The discrete morning departures of these aggregated aerial insectivores from ground-based roosting locations are detected by radars around sunrise. For the first time, we applied a machine learning algorithm to automatically detect and track these largescale behaviors. We used 21 years of data from 12 weather surveillance radar stations in the Great Lakes region to quantify the phenology in roosting behavior of aerial insectivores at three spatial levels: local roost cluster, radar station, and across the Great Lakes region. We show that their peak roosting activity timing has advanced by 2.26 days per decade at the regional scale. Similar signals of advancement were found at the station scale, but not at the local roost cluster scale. Air temperature trends in the Great Lakes region during the active roosting period were predictive of later stages of roosting phenology trends (75% and 90% passage dates). Our study represents one of the longest-term broad-scale phenology examinations of avian aerial insectivore species responding to environmental change and provides a stepping stone for examining potential phenological mismatches across trophic levels at broad spatial scales.

The current state-of-the-art task-oriented semantic parsing models use BERT or RoBERTa as pretrained encoders; these models have huge memory footprints. This poses a challenge to their deployment for voice assistants such as Amazon Alexa and Google Assistant on edge devices with limited memory budgets. We propose to learn compositional code embeddings to greatly reduce the sizes of BERT-base and RoBERTa-base. We also apply the technique to DistilBERT, ALBERT-base, and ALBERT-large, three already compressed BERT variants which attain similar state-of-the-art performances on semantic parsing with much smaller model sizes. We observe 95.15% ∼ 98.46% embedding compression rates and 20.47% ∼ 34.22% encoder compression rates, while preserving >97.5% semantic parsing performances. We provide the recipe for training and analyze the trade-off between code embedding sizes and downstream performances.

While large-scale pretrained language models have significantly improved writing assistance functionalities such as autocomplete, more complex and controllable writing assistants have yet to be explored. We leverage advances in language modeling to build an interactive writing assistant that generates and rephrases text according to fine-grained author specifications. Users provide input to our Intent-Guided Assistant (IGA) in the form of text interspersed with tags that correspond to specific rhetorical directives (e.g., adding description or contrast, or rephrasing a particular sentence). We fine-tune a language model on a dataset heuristically-labeled with author intent, which allows IGA to fill in these tags with generated text that users can subsequently edit to their liking. A series of automatic and crowdsourced evaluations confirm the quality of IGA's generated outputs, while a smallscale user study demonstrates author preference for IGA over baseline methods in a creative writing task. We release our dataset, code, and demo to spur further research into AI-assisted writing. * * Most of the work done during an internship at Adobe. 1 As these systems are not open-sourced, it is unclear how exactly they are implemented.