Multi-object tracking (MOT) systems often rely on accurate object detectors; however, accurate detectors are not available in every application domain. We present Robust Confidence Tracking (RCT), an offline MOT algorithm designed for settings where detection quality is poor. Whereas prior methods simply threshold and discard detection confidence information, RCT relies on the exact detection confidence values to increase track quality throughout the entire tracking pipeline. This innovation (along with some simple and well-studied heuristics) allows RCT to achieve robust performance with minimal identity switches, even when provided with completely unfiltered detections. To compare trackers in the presence of unreliable detections, we present a challenging real-world underwater fish tracking dataset, FISHTRAC. In an large-scale evaluation across FISHTRAC, UA-DETRAC, and MOTChallenge data, RCT outperforms a wide variety of trackers, including deep trackers and more classic approaches. We have publically released our FISHTRAC codebase and training dataset at https://github.com/tmandel/fish-detrac, which will facilitate comparing trackers on understudied problems.

Although artificial intelligence has revolutionized data analysis, significantly less work has focused on using AI to improve scientific data collection. Past work in AI for data collection has typically assumed the objective function is well-defined by humans before starting an experiment; however, this is a poor fit for scientific domains where new discoveries and insights are made as data is being collected. In this paper we present a new framework to allow AI systems to work together with humans (e.g. scientists) to collect data more effectively in simple scientific domains. We present a novel algorithm, TESA, which seeks to achieve good performance by learning from past human behavior how to direct data to places that are likely to become scientifically interesting in the future. We analyze the problem theoretically, defining a novel notion of regret in this setting and showing that TESA is zero regret. Next, we show that TESA outperforms other related algorithms in simulations using real data drawn from three diverse domains (economics, mental health, and cognitive psychology). Finally, we run experiments with human subjects across these scientific domains to compare our iterative human-in-the-loop process to a (more standard) workflow in which information is communicated to the AI a priori.

To prevent harmful AI behavior, people need to specify constraints that forbid undesirable actions. Unfortunately, this is a complex task, since writing rules that distinguish harmful from non-harmful actions tends to be quite difficult in real-world situations. Therefore, such decisions have historically been made by a small group of powerful AI companies and developers, with limited community input. In this paper, we study how to enable a crowd of non-AI experts to work together to communicate high-quality, reliable constraints to AI systems. We first focus on understanding how humans reason about temporal dynamics in the context of AI behavior, finding through experiments on a novel game-based testbed that participants tend to adopt a long-term notion of harm, even in uncertain situations that do not affect them directly. Building off of this insight, we explore task design for long-term constraint specification, developing new filtering approaches and new methods of promoting user reflection. Next, we develop a novel rule-based interface which allows people to craft rules in an accessible fashion without programming knowledge. We test our approaches on a real-world AI problem in the domain of education, and find that our new filtering mechanisms and interfaces significantly improve constraint quality and human efficiency. We also demonstrate how these systems can be applied to other real-world AI problems (e.g. in social networks).

Abstract: Data science is on the rise, being a high-demand skill for a number of employers, both in the tech field and in various scientific disciplines. Since it is a new field, only a handful of schools offer a data science program, most of those being large research institutions. This presents a unique opportunity for smaller schools to combine an exciting and highly relevant topic with the personalized education, small class sizes, and individual attention they are known for. On the flip side, building out data science at small schools also presents numerous challenges, including finding sufficient faculty time and expertise, attracting a sufficient number of students, forging interdepartmental connections, and convincing the administration to approve the new courses and program. In this poster, we present lessons learned from data science efforts at three small institutions, including one which recently successfully developed a data science certificate (containing four new data science courses).

In order for reinforcement learning systems to learn quickly in vast action spaces such as the space of all possible pieces of text or the space of all images, leveraging human intuition and creativity is key. However, a human-designed action space is likely to be initially imperfect and limited; furthermore, humans may improve at creating useful actions with practice or new information. Therefore, we propose a framework in which a human adds actions to a reinforcement learning system over time to boost performance. In this setting, however, it is key that we use human effort as efficiently as possible, and one significant danger is that humans waste effort adding actions at places (states) that aren't very important. Therefore, we propose Expected Local Improvement (ELI), an automated method which selects states at which to query humans for a new action. We evaluate ELI on a variety of simulated domains adapted from the literature, including domains with over a million actions and domains where the simulated experts change over time. We find ELI demonstrates excellent empirical performance, even in settings where the synthetic "experts" are quite poor.

In an increasingly computer-driven world, it is no surprise that undergraduate interest in computer science has grown enormously over the years. To meet this growing demand, it is vital to ensure there is a growing supply of talented and passionate CS educators. However, most computer science professors must first earn a PhD at a research university. Therefore, we seek to understand whether these institutions are meeting the growing need for passionate and highly-trained computer science educators. Specifically, we analyze 19 years of historical data gathered from the University of Washington Computer Science & Engineering department. Although focusing on graduates from a single school is a limitation, University of Washington is somewhat unique in that it publicly posts graduating PhD students' preferences online, and UW has had numerous notable graduates pursue careers in teaching, research, and industry. Our analysis of 309 graduating students reveals that the number of graduating students interested in teaching is alarmingly small, and has not grown over time to meet the increased need for quality educators. Our more detailed analysis of the change in interest over time suggests several potential reasons for this phenomenon. We hope that this poster opens dialogue surrounding possible interventions that might alleviate this issue.

In many real-world reinforcement learning problems, we have access to an existing dataset and would like to use it to evaluate various learning approaches. Typically, one would prefer not to deploy a fixed policy, but rather an algorithm that learns to improve its behavior as it gains more experience. Therefore, we seek to evaluate how a proposed algorithm learns in our environment, meaning we need to evaluate how an algorithm would have gathered experience if it were run online. In this work, we develop three new evaluation approaches which guarantee that, given some history, algorithms are fed samples from the distribution that they would have encountered if they were run online. Additionally, we are the first to propose an approach that is provably unbiased given finite data, eliminating bias due to the length of the evaluation. Finally, we compare the sample-efficiency of these approaches on multiple datasets, including one from a real-world deployment of an educational game.

The current state-of-the-art method for generating educational content, such as math word problems and hints, is manual authoring by domain experts. Unfortunately, this is costly, time consuming, and produces content that lacks diversity. Attempts to automatically address the time and diversity issues through natural language generation still do not produce content that is sufficiently creative and varied. Crowdsourcing is a viable alternative - there has been a great deal of research on leveraging human creativity to solve complex problems, such as user interface design. However, these systems typically decompose complex tasks into subtasks. Writing a single word problem or hint is a small enough problem that it is unclear how to further break it down, but also far more complex than typical microtasks like image labeling. Therefore, it is not obvious how to apply these worker improvement methods or which ones are most effective (if at all). We build upon successful task design factors in prior work and run a series of iterative studies, incrementally adding different worker-support elements. Our results show that successive task designs improved accuracy and creativity.

Current algorithms for the standard multi-armed bandit problem have good empirical performance and optimal regret bounds. However, real-world problems often differ from the standard formulation in several ways. First, feedback may be delayed instead of arriving immediately. Second, the real world often contains structure which suggests heuristics, which we wish to incorporate while retaining the best-known theoretical guarantees. Third, we may wish to make use of an arbitrary prior dataset without negatively impacting performance. Fourth, we may wish to efficiently evaluate algorithms using a previously collected dataset. Surprisingly, these seemingly-disparate problems can be addressed using algorithms inspired by a recently-developed queueing technique. We present the Stochastic Delayed Bandits (SDB) algorithm as a solution to these four problems, which takes black-box bandit algorithms (including heuristic approaches) as input while achieving good theoretical guarantees. We present empirical results from both synthetic simulations and real-world data drawn from an educational game. Our results show that SDB outperforms state-of-the-art approaches to handling delay, heuristics, prior data, and evaluation.

We present a general automatic experimentation and hypothesis generation framework that utilizes a large set of users to explore the effects of different parts of an intervention parameter space on any objective function. We also incorporate importance sampling, allowing us to run these automatic experiments even if we cannot give out the exact intervention distributions that we want. To show the utility of this framework, we present an implementation in the domain of fractions and numberlines, using an online educational game as the source of players. Our system is able to automatically explore the parameter space and generate hypotheses about what types of numberlines lead to maximal short-term transfer; testing on a separate dataset shows the most promising hypotheses are valid. We briefly discuss our results in the context of the wider educational literature, showing that one of our results is not explained by current research on multiple fraction representations, thus proving our ability to generate potentially interesting hypotheses to test.

The challenge of mobile text entry is exacerbated as mobile devices are used in a number of situations and with a number of hand postures. We introduce ContextType, an adaptive text entry system that leverages information about a user's hand posture (using two thumbs, the left thumb, the right thumb, or the index finger) to improve mobile touch screen text entry. ContextType switches between various keyboard models based on hand posture inference while typing. ContextType combines the user's posture-specific touch pattern information with a language model to classify the user's touch events as pressed keys. To create our models, we collected usage patterns from 16 participants in each of the four postures. In a subsequent study with the same 16 participants comparing ContextType to a control condition, ContextType reduced total text entry error rate by 20.6%.

Bit errors are common in wireless networks, and techniques for overcoming them traditionally consist of expensive retransmission (e.g. Automatic Repeat reQuest (ARQ)) or expensive Forward Error Correction (FEC), both of which are undesirable in resource-constrained wireless networks such as wireless sensor networks (WSNs). In this paper, we present TVA (Transmit-Verify-Acknowledge), a protocol that can correct errors without adding additional redundancy to data packets. Instead, TVA corrects errors using the redundancy inherent in Cyclic Redundancy Checks (CRCs). The ubiquity of CRCs has the advantage of allowing TVA to be both backwards-compatible and backwards-efficient with link-layer protocols such as IEEE 802.15.4. We present a novel method of CRC error correction, which is compact and computationally efficient, and is designed to correct the most common error patterns observed in WSNs. We demonstrate that TVA provides reliability effectively equivalent to that of ARQ. We perform trace-driven simulations using data from sensor network deployments in different environments and analyze TVA's performance at different message lengths. To demonstrate the practicality of TVA, we implement it in TinyOS, and perform experiments on MicaZ motes to evaluate TVA in the presence of 802.11 interference. We find that TVA improves over ARQ and FEC-based protocols, using 31% less redundant communication and 30% less additional time to recover errored packets compared to ARQ.