Total: 14
Neural architecture search (NAS) has emerged as an algorithmic method of developing neural network architectures. Weight Agnostic Neural Networks (WANNs) are an evolutionary-based NAS approach. Fundamentally, WANNs find network structures that are relatively insensitive to shifts in weight values and are typically much smaller than an equivalent performance dense network. Here, we extend the WANN framework to search for spiking circuits and in doing so investigate whether these circuit motifs can also yield task performance that is weight agnostic. We analyze properties such as the complexity of the solution, as well as performance. Our results successfully show the performance of spiking WANNs on several exemplar tasks.
A significant portion of the veteran population suffers from PTSD, a mental illness that is often accompanied by social anxiety disorder. Student veterans are especially vulnerable as they struggle to adapt to a new, less structured college lifestyle. In order to assist psychologists and social workers in the treatment of social anxiety disorder we use machine learning to analyze transcribed interview text and apply topic modelling to highlight common stress factors for student veterans. The results detailed in this paper also have broader impacts in fields such as pedagogy and public health.
We introduce a novel technique that uses a multi-headed neural network to analyze symmetric games with a variable number of players, where the number of participants falls in a specified range. We hypothesize that the payoffs in a game with x players are similar or related to the same game with x + 1 players, given a large value of x. With this hypothesis, we generalize prior work to analyze games with a large, variable number of players.
Illegal wildlife poaching is driving the loss of biodiversity. To combat poaching, rangers patrol expansive protected areas for illegal poaching activity. However, rangers often cannot comprehensively search such large parks. Thus, the Protection Assistant for Wildlife Security (PAWS) uses machine learning to help identify the areas with highest poaching risk. As PAWS is deployed to parks around the world, we recognized that many parks have limited resources for data collection and therefore have scarce feature sets. To ensure under-resourced parks have access to meaningful poaching predictions, we introduce the use of publicly available remote sensing data to extract features for parks. By employing this data from Google Earth Engine, we also incorporate previously unavailable dynamic data such as climate and primary production patterns to enrich predictions with seasonal trends. We automate the entire data-to-deployment pipeline and find that, with only using publicly available data, we recuperate prediction performance comparable to predictions made using features manually computed by park specialists. We conclude that the inclusion of satellite imagery creates for a robust system through which parks of any resource level can benefit from poaching risks for years to come.
Rotating Savings and Credit Associations (Roscas) are a widely documented informal financial organization that is often used in low-income communities with limited funding sources. Roscas have been shown to serve as a tool for economic empowerment and a way of mitigating adverse shocks to income for vulnerable communities. In this paper, I describe my contributions to a research project in which we study the allocative efficiency of different Rosca structures and formats in the presence of risk-averse agents.
Safety properties of neural networks are critical to their application in safety-critical domains. Quantification of their robustness against uncertainties is an upcoming area of research. In this work, we propose an approach for providing probabilistic guarantees on the performance of a trained neural network. We present two novel metrics for probabilistic verification on training data distribution and test dataset. Given a trained neural network, we quantify the probability of the model to make errors on a random sample drawn from the training data distribution. Second, from the output logits of a sample test point, we measure its p-value on the learned logit distribution to quantify the confidence of the model at this test point. We compare our results with softmax based metric using the black-box adversarial attacks on a simple CNN architecture trained for MNIST digit classification.
Automated deception detection systems can enhance societal well-being by helping humans detect deceivers and support people in high-stakes situations across health, social work, and legal domains. Existing computational approaches for detecting deception have not leveraged dimensional representations of affect, specifically valence and arousal, expressed during communication. My research presents a novel analysis of the potential for including affect in machine learning models for detecting deception. My work informs and motivates the development of affect-aware machine learning approaches for modeling deception and other social behaviors during human interactions in-the-wild. This research, independently defined and conducted by me, is from work-in-progress towards my undergraduate thesis in the Department of Computer Science at the University of Southern California.
My research presents a method for efficient exploration of an outdoor, unknown area, which aims to achieve precise coverage of regions of interest within that area. While this method for autonomous exploration was designed for autonomous controllers in unmanned aerial vehicles (UAVs), the concepts apply to any vehicle which uses autonomous navigation. We consider an environment with areas of interest of various sizes littered throughout, and a reinforcement learning agent which is tasked with discovering and mapping these areas in an efficient manner.
Given recent advances in deep music source separation, a feature representation method is proposed that combines source separation with a state-of-the-art representation learning technique that is suitably repurposed for computer audition (i.e. machine listening). A depthwise separable convolutional neural network is trained on a challenging electronic dance music (EDM) data set and its performance is compared to convolutional neural networks operating on both source separated and standard spectrograms. It is shown that source separation improves classification performance in a limited-data setting compared to the standard single spectrogram approach.
The project developed a device to assist blind and visually impaired users in complying with social distance rules. The main challenge was how to identify people from other objects and notify the user accordingly. Equally ambiguous was how to best notify the user that a person was violating the social distance rules. The hardware used is a Raspberry Pi running an image classification algorithm and hooked to a depth camera. The use of prediction labels and accuracy combined with the distance calculated by the depth camera made it possible to detect when a person was getting closer than 2 meters (6 ft) to the user. The device was tested both in daylight and at night, and with different lighting conditions. The device responded with a accuracy close to 1.9 meters. which is very acceptable. Testing showed that the camera is capable of identifying people coming from a 30 degrees angle from either side at around 1.9 meters of distance, giving a good range of object detection.
Current applications of Agent-based Modeling (ABM) in natural phenomena like wildfire land suppression and hurricane forecasting are in monitoring emergent behavior patterns among large groups of people (Hilljegerdes 2018). However, current evacuation times and plans for natural disaster management leave underserved communities vulnerable to substantial financial and welfare losses, especially when false positives during current predictions continue to influence evacuation decisions. A Machine Learning ABM (ML-ABM) model of hurricane trajectories introduces a pioneering opportunity to capture complex physical processes associated with hurricanes while minimizing computational costs and errors, thereby providing more accurate real-time prediction of hurricanes for improved disaster management. This Hurricane Track Prediction ML-ABM model aims to quickly model and predict hurricane tracks in only a few minutes while retaining some of the complex physical process interactions of real storms through feature engineering and deep learning. This work focuses on the implementation of an RNN with bidirectional time-distributed Long-Short Term Memory cells, accounting for positive and negative time direction in time series forecasting. The observations and predictions were represented as a multi-agent system in NetLogo for further emergent pattern analysis in an expanded research by Arthur Drake et. al (2020). The model also evaluates benchmark comparisons against the National Hurricane Center’s 5-Year Average Forecast Errors and the BCD5 Error Model, a combined intensity and track prediction error model that utilizes best track input and models decay over land.
Our research focuses on developing matching policies that match drivers and riders for ride-pooling services. We aim to develop policies that balance efficiency and various forms of fairness. We did this through two methods: new matching algorithms that include a fairness term in the objective function, and income redistribution methods based on the Shapley value of a driver. I tested these methods on New York City Taxicab data to evaluate their performance and found that they succeed in reducing certain forms of fairness.
We propose a MOTIF-driven contrastive framework to pretrain a graph neural network in a self-supervised manner so that it can automatically mine motifs from large graph datasets. Our framework achieves state-of-the-art results on various graph-level downstream tasks with few labels, like molecular property prediction.
The prevalence of impostor syndrome in computer science students from underrepresented backgrounds contributes to low retention rates. Bison Hacks the Yard is an augmented reality game that aims to reduce impostor syndrome in underrepresented students by presenting a novel way to strengthen their knowledge of fundamental data structures and providing specialized videos of Historically Black College or University alumni, sharing their struggles with impostor syndrome.