Hi everyone! I’m Vivek and this semester I’m working with Dr. Xiaowei Jia to study how temporal convolutional neural networks can be used to predict water conditions.

What are temporal convolutional neural networks, you might ask? Great question! A temporal convolutional neural network (TCNN) is a type of machine learning (ML) architecture which can take extremely large datasets, such as decades-long time series of water temperatures, and use them to predict what specific values in the same series might look like at some point in the future. A simplified diagram of what a TCNN looks like is shown below:

Essentially, data comes in at the input layer, is transformed as it passes through the hidden layer or layers, and ends up as a specific prediction in the output layer. By measuring how far away this prediction is from a known or expected value — the prediction error or “loss” — we can adjust the transformations that the hidden layers perform on an initial input time so that the prediction is slightly more accurate the next time around. By repeating this process many times, it becomes possible to “train” a TCNN to make accurate temperature predictions even when an actual temperature value is not known in advance.

What’s so special about water temperatures in particular? An interactive app created by the U.S. Geological Survey explains the importance of monitoring water temperatures within the Delaware River Basin (DRB), which provides drinking water to more than 15 million people:

Many species – including economically important game species like brook trout and endangered species like the Dwarf wedgemussel – thrive or spawn in specific temperature ranges. These species are threatened by rising stream temperatures due to urbanization, climate change, and human alteration of streamflow patterns. . . . in the upper portion of the basin, some of this warming can be mitigated by releasing cold water from New York City reservoirs. Because the cold reservoir water is a limited resource, we can help protect cold water habitat in the DRB by improving our ability to predict stream temperature, and therefore make an educated guess at when, where, and how much cold water to release.

In other words, having the ability to accurately predict changes in water temperatures will help reservoir staff effectively manage their scarce cold water resources, thus helping to maintain the integrity of aquatic ecosystems and preserve their critical roles as habitats and water supplies for both human and animal species. Needless to say, tasks like this are only going to become more and more important as the planet warms as a result of greenhouse gas emissions from human societies.

As such, I hope to use my CURF to help develop technologies that can be deployed to help communities mitigate and adapt to the effects of global warming. However, as a computer science & psychology double major who hopes to become an academic cognitive scientist one day, I am also hoping to gain more experience with modern deep learning techniques as well as their limitations. Accordingly, my CURF gives me the perfect opportunity to combine my interests in artificial intelligence and machine learning with contributing to the ineffably important fight against anthropogenic climate change.