Invasions of species from one region to another (bioinvasions) is a worldwide danger to biodiversity and ecosystems. Marine bioinvasions are rapidly facilitated by maritime traffic via hull fouling and/or ballast-water discharge.

Despite numerous previous studies, a knowledge gap exists regarding the parameters that dictate which ship route presents higher invasion risks by introduction of fouling organisms. Factors such as changes in water salinity and temperatures, the type of vessel, the duration of the voyage, and turnover time in ports of call are likely to influence the survival of "passenger" organisms, and thus correlate with risk.

In our research, we harness machine learning to learn the parameters that affect the survival of invasive species during the voyage. Specifically, we conduct a series of controlled voyage simulations using an invasive ascidian which is exposed to changing environments that mimic real-time voyages. Machine-learning algorithms are then used to analyze the obtained experimental data and build models that can predict survival from route characteristics. To our knowledge, machine learning has never been used in such types of studies.

In this project we examine if and how echolocating bats can navigate over kilometers using only acoustic echoes. To study this, we generated a huge database of simulated echoes based on actual 3D maps of northern Israel, at a scale of dozens of kilometers.

This echoic information was examined, revealing the unique characteristics of different landscapes, that is - the statistics of the echo provide vast information about the landscape (e.g., whether it is a field, orchard, or lake and also sufficient information about unique landmarks that should allow echo-based navigation.

In the next part, we will construct a Navigation Neural-Network model that allows large-scale acoustic mapping. First, we used an Auto-Encoder network to represent the echoes in a low-dimensional space.

Next, we used Transformer LLMs to train a navigator that learns to navigate based on acoustic information. This network will rely on the previous ones in order to steer the agent’s navigation towards familiar locations. The network has already shown promising results. 

Is criminal behavior influenced by the quality of air?

In this project, we use detailed daily data from the Indian state of Karnataka to estimate the relationship between ambient air pollution and crime incidents.

Our results indicate that a one-standard-deviation increase in air pollution, measured as the daily concentration level of PM 2.5, is associated with an increase of 2.3% in the number of violent crimes (significant at the 1% level). We find no effect on property crimes.

To verify that the estimated relationship is in fact causal, we exploit atmospheric thermal inversions as a source of plausibly exogenous variation in the daily level of air pollution. These results suggest that investment in pollution abatement is advantageous not only through improvements in health outcomes but also via improving public safety.