Publications

Wildfires are a disastrous phenomenon which cause damage to land, loss of property, air pollution, and even loss of human life. Due to the warmer and drier conditions created by climate change, more severe and uncontrollable wildfires are expected to occur in the coming years. This could lead to a global wildfire crisis and have dire consequences on our planet. Hence, it has become imperative to use technology to help prevent the spread of wildfires. One way to prevent the spread of wildfires before they become too large is to perform early detection i.e, detecting the smoke before the actual fire starts. In this paper, we present our Wildfire Detection and Alert System which use machine learning to detect wildfire smoke with a high degree of accuracy and can send immediate alerts to users. Our technology is currently being used in the USA to monitor data coming in from hundreds of cameras daily. We show that our system has a high true detection rate and a low false detection rate. Our performance evaluation study also shows that on an average our system detects wildfire smoke faster than an actual person.

We present the Earthquake Detective dataset - A crowdsourced set of labels on potentially triggered (PT) earthquakes and tremors. These events are those which may have been triggered by large magnitude and often distant earthquakes. We apply Machine Learning to classify these PT seismic events and explore the challenges faced in segregating such low amplitude signals. The data set and code are available online.

We present OtoWorld, an interactive environment in which agents must learn to listen in order to solve navigational tasks. The purpose of OtoWorld is to facilitate reinforcement learning research in computer audition, where agents must learn to listen to the world around them to navigate. OtoWorld is built on three open source libraries: OpenAI Gym for environment and agent interaction, PyRoomAcoustics for ray-tracing and acoustics simulation, and nussl for training deep computer audition models. OtoWorld is the audio analogue of GridWorld, a simple navigation game. OtoWorld can be easily extended to more complex environments and games. To solve one episode of OtoWorld, an agent must move towards each sounding source in the auditory scene and “turn it off”. The agent receives no other input than the current sound of the room. The sources are placed randomly within the room and can vary in number. The agent receives a reward for turning off a source. We present preliminary results on the ability of agents to win at OtoWorld. OtoWorld is open-source and available.

Training an autonomous agent in the real world is a cumbersome process. The hardware modules required are expensive and they need routine maintenance. The data collection process is time-consuming and it is difficult to collect data in different conditions and scenarios. Moreover, testing these agents in the real world requires many permissions and could be potentially hazardous. This paper introduces a virtual environment for training and testing of autonomous driving agents. The environment has features like customizable car parameters and sensors, different terrains, customizable data extraction parameters, and simulated pedestrian and vehicular traffic. The environment can connect to any learning agent via a communication interface. Therefore, the environment introduced in this paper expedites the training and testing process and the learned knowledge representations can be scaled to the real world.

It is commonly observed that deep networks trained for classification exhibit class-selective neurons in their early and intermediate layers. Intriguingly, recent studies have shown that these class-selective neurons can be ablated without deteriorating network function. But if class-selective neurons are not necessary, why do they exist? We attempt to answer this question in a series of experiments on ResNet-50s trained on ImageNet. We first show that class-selective neurons emerge during the first few epochs of training, before receding rapidly but not completely; this suggests that class-selective neurons found in trained networks are in fact vestigial remains of early training. With single-neuron ablation experiments, we then show that class-selective neurons are important for network function in this early phase of training. We also observe that the network is close to a linear regime in this early phase; we thus speculate that class-selective neurons appear early in training as quasi-linear shortcut solutions to the classification task. Finally, in causal experiments where we regularize against class selectivity at different points in training, we show that the presence of class-selective neurons early in training is critical to the successful training of the network; in contrast, class-selective neurons can be suppressed later in training with little effect on final accuracy. It remains to be understood by which mechanism the presence of class-selective neurons in the early phase of training contributes to the successful training of networks.

There have been many recent advancements in the field of artificial intelligence and machine learning. Nevertheless, the problem of learning from a few examples persists. The process of learning from just an example is easy for humans but not for a computer. Learning from a small number of samples is especially necessary in the case of facial recognition systems as the number of samples per person is limited. The aim is to explore, analyze and improve the different techniques which can be used for Face Recognition where the algorithm is fed with a few examples of faces i.e. the process of k shot learning for Face Recognition has been explored using the LFW and FEI datasets. The techniques of transfer learning have been used along with the famous Dlib library with some improvements using methods of deep learning.

Intraplate earthquakes have longer recurrence times and less confined locations than plate boundary earthquakes. Consequently, they can wreak havoc by catching populations that are relatively unprepared. Interplate areas also suffer from sparse instrumentation and historical records, compared to regions along tectonic boundaries. The Chicago metropolitan area, which has witnessed a handful of felt earthquakes in the past century, is a data-sparse area of strong concern due to its high population and building density. To understand seismic hazards in this urban inraplate area, we aim to detect small local earthquakes, including potential aftershocks of the 2013 M3.2 earthquake that occurred near a suburban quarry, using data from a single broadband seismometer. Due to the noisy, urban and industrial setting of the seismometer and the lack of recorded earthquakes to be used as templates, traditional methods like the STA/LTA ratio method and waveform cross correlation are ill-suited for our aim. To build a model that is better adapted to our noisy urban and industrial environment, we trained a tree-based classifier to detect and distinguish local earthquakes, man-made blasts, and other sources of noise from single-station data. Features for our classifier include conventional metrics such as STA/LTA ratios and power spectral density (PSD) values along with features based on higher-order statistics (e.g. skewness and kurtosis) and frequency bands pertinent to urban environments. The sparsity of earthquakes in the area posed the primary challenge for our aim, in both training and testing our model. To combat this sparsity, we created synthetic samples of earthquakes by augmenting seismograms from the STanford EArthquake Dataset (STEAD) with synthetic, stationary noise, created from averaged Chicago noise PSDs. We will present the results of our model’s application on continuous, single-station data in the Chicago area and compare our detections with those using the STA/LTA method and the deep learning model EQTransformer. We will also discuss the model’s interpretability and generalizability by presenting the results from our feature analysis and the application of our classifier on urban earthquake data outside of Chicago.

To improve seismic event detection and classification in our increasingly urban and often sparsely-instrumented environment, it is important to expand our single-station methods to detect small seismic events in highly-fluctuating, high-amplitude background noise. In this study, we aim to identify effective waveform features which can detect and discriminate small local earthquakes, explosions such as quarry blasts, recurring industrial activity, and other sources of environmental and anthropogenic noise in urban seismic data. Some explored features include measures of power spectral density (PSD) misfit and modified STA/LTA ratios using skewness and kurtosis in the frequency domain. To assess the ability of our features in detecting transient events in urban seismic data, we apply a simple unsupervised learning model (K-means clustering) to continuous feature data from a single broadband station in the Chicago area. We systematically investigate how the findings of our clustering model change with additional features and processing steps. For instance, we explore how filtering our waveforms at characteristic frequency bands of environmental and anthropogenic noise can improve our model performance. We will present a few notable clusters of seismic events in the Chicago area and discuss their characterizing features and possible sources. To assess the efficacy of our features in different urban environments, we also apply our clustering model to continuous data from a single station in Singapore and present our preliminary findings. We conclude by discussing additional features and methods that will be explored in the future to improve our model performance and analysis.

Study of dynamically triggered earthquakes and tremors allows a more comprehensive understanding of how faults evolve and the temporal conditions under which transient stresses may activate them. In this citizen science project, we asked volunteer scientists to identify earthquakes and tremors in Alaska and the Western US that could have been dynamically triggered. The project is called Earthquake Detective and is hosted on zooniverse.org. After a brief training module, users view graphs and listen to relevant sections of seismograms, accelerated to audible frequencies, and classify signals in four ways: earthquake, tremor, noise, and null. Over 6000 volunteer scientists have made 130,000 classifications after listening to and viewing relevant, filtered sections of seismic recordings of Mw>7.5 earthquakes spanning the years 2012-2018. We find 12 potentially triggered earthquakes in Alaska and 4 earthquakes in the Western US, while classification of tremors was quite rare, and we identify one in southcentral Alaska. Engaging citizen scientists in this type of experiment has many advantages: 1) we exploit the ability of the human ear to implicitly perform a spectrogram (time-frequency analysis) capable of discerning a wide range of different signals; 2) multiple human ears listening to the same signal provides redundancy of observations and therefore statistical analysis can be performed on the classifications; 3) a blind volunteer is positioned to perform and contribute at least as well as a seeing volunteer; 4) the project enhances informal learning because the online platform, Zooniverse, is available to persons of all identities, ages, cultural backgrounds, and education level; 5) the Comments section of our project on Zooniverse can be mined for keywords, FAQ, and tracking user and expert dialogue for learning patterns and communication styles to both refine and inform improvements to our outreach strategies. We present our results from Earthquake Detective to date and share plans to extend its use to classify small, local seismic events on faults of interest in and around the San Francisco Bay Area by recruiting local high school and undergraduate students to be our citizen scientists thereby providing them the experience of scientific research.