A human brain is a network of neurons; our society is a network of people connected via social relationships. Do they share anything else in common? What kinds of analogies can we make? This talk will examine complex contagion dynamics and optimal modularity phenomenon that occur in both systems. In particular, we will discuss optimal modularity as a potential organizing principle that shapes the structure of neural networks.

Cascading failures in power systems represent a significant risk to electric transmission systems and yet quantifying this risk remains challenging. One such challenge is how to cope with the combinatorial search space of potential triggering events (small sets of power system elements failing simultaneously) which may result in a cascade. Here, we present a new, group-testing inspired algorithm for efficiently sampling cascade-inducing events and compare it to the current state-of-the-art, demonstrating our new approach is easier to optimize and, under certain circumstances, can reduce the running time for finding events which trigger cascades.

Enabling machines to respond appropriately to natural language commands could greatly expand the number of people to whom they could be of service. Recently, advances in neural network-trained word embeddings have empowered non-embodied text-processing algorithms, and suggest they could be of similar utility for embodied machines. Here we introduce a method that does so by training robots to act similarly to semantically-similar word2vec encoded commands. We show that this enables them to act appropriately, after training, to previously-unheard commands. Finally, we show that inducing such an alignment between motoric and linguistic similarities can be facilitated or hindered by the mechanical structure of the robot. This points to future, large scale methods that find and exploit relationships between action, language, and robot structure.

Artificial Neural Networks, a core element of many modern machine learning applications, are unable to adapt to changing environmental conditions. Once trained on a specific set of training data, which may not cover all of the situations that the network may experience when deployed in the "real world", the brain of traditional neural networks, the weights, are fixed and do not change in response to new experiences. We utilize a variety of memory mechanisms, such as recurrent connections, plasticity, and external memory modules, in an effort to develop models that are better able to adapt to changing conditions and learn from new experiences, even after the initial training phase has been completed.

Chagas disease is a neglected tropical disease in Central and South America which is particularly prevalent in poor and rural communities. The main vector of the disease in Guatemala, Triatoma Dimidiata, maintains a notable domestic population in southern Guatemala, despite ongoing insecticide spraying programs. This talk will present a new network model for understanding the factors which contribute to domestic infestation. We demonstrate that the model, which takes into account the socio-economic variables of individual houses as well as the spatial relationship between houses, can predict patterns of infestation better than taking into account either spatial or social-economic variables alone.

In power systems, accurate device modeling is the key factor to grid reliability, availability, and resiliency. Many critical tasks such as planning or even realtime operation decisions rely on accurate modeling. In the US, all generators of 10 MVA or larger are required to validate their models every five years. This project aims to use advanced machine learning such as convolutional neural network (CNN) to estimate the model parameters from power system response data. The data recorded after system disturbances proved to have valuable information to verify power system devices. Existing methods use finite differential equations methods and statistical optimization methods to estimate the model parameters which suffer from being ill-posed and might result in more than one solution. So far, we achieved promising results on simple systems and the next steps are to work on more complicated scenarios to prove that Deep Leaning based methods can be useful.

Hydrological storm events are main cause of the transport of river water quality constituents such as turbidity, suspended sediments, and nutrients, and thus analyzing those events helps to characterize the dynamics and flux of such constituents. The conventional approach to the storm event analysis has been drawing event concentration-discharge (C-Q) variable plots and identifying two-dimensional hysteresis loop patterns from the plots. While effective and informative to some extent, this approach has shortcoming in capturing the temporality of variables, as their values are projected onto the C-Q plane and consequently a lot of details in temporal variations of the C-Q relationship are obscured. In this study, we address this issue by using a temporality-aware approach, namely, multivariate time series clustering. This clustering is applied to sequences of measurements from multiple sensors, which in this study are river discharge and suspended sediment concentration. This research utilizes turbiditybased monitoring of suspended sediment from Mad River watershed, located in the Lake Champlain Basin in the northeastern United States. The resulting computational clusters were found to have correlation to hysteresis loops and watershed sites and have higher explanatory power than categorization through conventional C-Q plots in terms of metrics characterizing astorm event.

In this talk, I’ll introduce the automated design and manufacture of robots built entirely from pluripotent, Xenopus stem cells: Xenobots

Robotics-human interaction is often challenging and dangerous in Space. Creating a personalized robotics control language that is gesture-based and user-specific allow aspects of accessibility in robotics control systems not seen in aerospace today. In this talk, I will discuss a gesture-based language called MISL that was proposed for the NASA L'Space workshop in Summer 2019. The talk will explore the advantages and disadvantages of such a system, as it relates to controlling robots in space in the proximity of humans.

Mathematics is a language. Many researchers are using neural networks to generate, translate, and interpret language. My goal is to use a neural network to generate a mathematical function that explains a given dataset. In any language, the placement of a single word depends greatly on the surrounding words. This fact is not reflected in the way that many versions of genetic programming – an algorithm that performs symbolic regression (function generation) – evolve functions. In my work, I train a neural network to label the nodes of a tree with a given shape. The tree will represents a mathematical function. The network iteratively labels each node in the tree by taking the existing node labels (if any) and the location of the node to be labeled as input when generating the next label. In this way, my method considers the surrounding labels (or “words”) as each new label is generated. To determine if this is a viable method to perform symbolic regression, I will compare it with genetic programming.

Data privacy is a growing concern for every individual, businesses, government and organization. The state of the art in balancing useful data analysis with user privacy is a family of techniques which enforce a quantitative notion of data anonymization called differential privacy. In this project we will leverage Duet—a domain specific programming language for writing differentially private programs to perform arbitrary data analysis tasks. Our aim is to construct a proof of correctness for both the design and the implementation of Duet's analysis for data privacy. Our approach is to use the Agda proof assistant to embed (1) the operational semantics of Duet programs, (2) the specification for Duet typing judgments, and (3) the algorithm for type-based static analysis. We will connect the analysis (3) to the meaning of programs (1) in two phases: by establishing the soundness of type judgments (2) w.r.t. the semantics of programs (1), and then by demonstrating the soundness of the analysis algorithm (3) w.r.t. the specification for typing (2).

I tend to make the same set of comments to the student collaborators. Although some of these comments can be annoying, they highlight major differences between a student’s and a professor’s perspective. In this talk, I will go through some of the most common comments that I routinely make and explain where they come from. I hope this talk can share some insights into the ‘professor’s perspective’.

This presentation presents a novel approach for a transportation signage (asset) management system employing radio-frequency identification (RFID) technology. Taking advantage of the remote sensing feature of RFID, the system operations can be mobile and remote, consequently improving the efficiency significantly. Additionally, user-friendly control software (C#/.net) and a database (MSSQL/AZURE SQL) have been developed as part of the system. In the current prototype, the former comprises an in-vehicle RFID reader software (called the “UVM Reader Assistant”) and a handheld RFID reader/writer software. The in-vehicle reader can read from signs and can read tags at various heights and angles. The two types of readers are connected through separate software that display relevant tag and asset data on the screen and connect them to a remote shared database, which is synchronized with their own local copies whenever connectivity is established. The developed system can serve different scenarios of RFID tag-based signage management efficiently and is resilient to intermittent and unexpected interruptions to database connectivity common in rural as well as urban environments. This research not only focused on developing a system, but also focused on investigating the effects of antenna placement, tag orientation, and tag type on received signal strengths measured by an RFID reader.

Our research leverages novel deep learning techniques to construct an automated system which detects, classifies, and determines the GPS location of road signs using images captured from a road vehicle. Our baseline system uses the state-of-the-art detector RetinaNet, which we have heavily modified to perform GPS localization in addition to object classification. To convert signs detected by RetinaNet into a geospatial road sign mapping, we build a tracker which pairs detections belonging to the same sign into tracklets, such that each tracklet represents a single sign. Our tracker uses a neural network trained to compute a distance metric which represents the similarity between pairs detections. Using the Hungarian algorithm, we merge detections with large levels of similarity, to the condense detections into tracklets. Our results show RetinaNet achieves impressive results when classifying and estimating the GPS location of road signs, particularly considering no depth information is provided to the network. The distance metric network and Hungarian Algorithm show promising initial results for sign tracking. Lastly, our work introduces a large-scale dataset to serve as one of the very few benchmarks for US. traffic signs recognition (TSR), and the first to feature object-related GPS information.