Fall 2017¶

12 / 1 Prof. Michael Stewart¶

Algorithms ruin everything

I’ll be presenting 2 papers that touch on people’s perceptions of algorithms and their impact on society:

Michael A. DeVito, Darren Gergle, and Jeremy Birnholtz. 2017. "Algorithms ruin everything": #RIPTwitter, Folk Theories, and Resistance to Algorithmic Change in Social Media. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems (CHI '17). ACM, New York, NY, USA, 3163-3174. DOI: https://doi.org/10.1145/3025453.3025659

Tufekci, Zeynep. "Algorithmic Harms beyond Facebook and Google: Emergent Challenges of Computational Agency," Colorado Technology Law Journal vol. 13, no. 2 (2015): p. 203-218.

11 / 10 Diana Godja (Advisor: Prof. Nathan Sprague)¶

Artificial Echolocation Using Deep Neural Networks

Our study explores the effectiveness of bat-inspired echolocation as a practical sensory modality for extracting high-resolution depth information. Traditional ultrasonic depth sensors provide a single scalar depth estimate based on the time delay between an emitted pulse and the detection of an echo. Bats use a similar mechanism to extract depth information. However, bats and other echolocating animals are able to create a sensory percept that is much richer than a single distance measurement. We demonstrate that a deep neural network can be trained to accurately reconstruct two-dimensional depth fields by analyzing the echoes from a single 10 millisecond frequency-modulated chirp.

10 / 27 Andrew Jones (Advisor: Prof. Nathan Sprague)¶

Preventing Blackout Catastrophe When Learning Sequentially with Elastic Weight Consolidation

Artificial neural networks have become the dominant machine learning approach across a wide range of domains. Significant strides have been made on single-task learning. The problem of sequentially learning multiple tasks has proved to be more challenging. Naive approaches suffer from “catastrophic forgetting”: networks lose accuracy on previously learned tasks when trained on new tasks. The recently introduced Elastic Weight Consolidation (EWC) algorithm is a novel and promising approach that calculates the importance of individual weights to previously learned tasks. A penalty term is introduced that preserves weights in proportion to their importance. However, preserving weights via EWC eventually results in total network failure due to the limited capacity of a fixed-size network, a phenomenon referred to as “Blackout Catastrophe”. Our proposed algorithm addresses this problem by employing EWC until the network capacity has been reached and then increasing the size of the network to accommodate additional tasks. We further investigate the potential of Fisher Information, which is used by EWC to evaluate the importance of each weight for each task, to probabilistically predict when the network needs to expand to prevent blackout catastrophe - an insight which could theoretically allow the network to learn sequentially in perpetuity without the need for human supervision.

10 / 13 Patricia D. Soriano and Garrett Folks (Advisor: Prof. Mike Lam)¶

ANALYSIS OF PARALLEL IMPLEMENTATIONS OF CENTRALITY ALGORITHMS

Speaker: Patricia D. Soriano

This talk explores parallel implementations of three network analysis algorithms for detecting node centrality: betweenness centrality (BC), eigenvalue centrality (EC), and degree and line importance (DIL). All solutions were written in the C programming language using OpenMP library for parallelization. We evaluated these implementations for accuracy and parallel scaling performance using five example networks. We found that the algorithms accurately reflect different notions of centrality. While DIL performs better in general because it is asymptotically faster than the other two algorithms, BC demonstrates better parallel strong scaling.

TRAVELING SALESMAN: A HEURISTIC SCALING ANALYSIS

Speaker: Garrett Folks

In this talk, we analyze two heuristics that approximate the Traveling Salesman Problem: K-Opt search and ant colony optimization. Our goal was to explore how these heuristics perform when run in parallel on multiple CPU cores as well as using GPU computing. We found that the K-Opt search heuristic showed impressive performance scaling results, especially when executed on a GPU. We also parallelized portions of the ant colony optimization and found good scaling. We conjecture that the ant colony optimization could be greatly improved with the use of GPU computing.

10 / 6 Prof. Mike Lam¶

Scheduling, Reproducibility, and Resilience

Dr. Lam and JMU student Garrett Folks spent Summer 2017 doing research at Lawrence Livermore National Laboratory (LLNL) in California. LLNL is a Department of Energy lab that supports a variety of computational research in the broad area of high-performance and scientific computing. This survey talk will give an overview of three ongoing research collaborations between LLNL and academic institutions with a general theme of improving the performance and reliability of numeric code.

9 / 22 Kevin Münch¶

Development of voice user interfaces and their impact to the userexperience of mobile applications

Voice user interfaces help users in hands-free situations and whenever graphical interfaces cannot be used. Developing voice user interfaces is a complex process. Besides the technical challenges of understanding and synthesizing human speech, the design of the dialogs is the key part to create a good interface. An easy dialog design includes the right wording and sentence structure. In addition, voice user interfaces can be combined with graphical interfaces to ensure a better usability. This presentation introduces the process of voice user interface development and the design of understandable dialogs. It displays this process through a practical example. This example explains the development of an Android application to experience interactive fiction. The application implements a voice user interface, created during the research for the thesis which this presentation is based on.

9 / 15 Prof. Jingwei Yang¶

A situation-centric, knowledge-driven requirements elicitation approach

Human factors have been increasingly recognized as one of the major driving forces of requirement changes. We believe that the requirements elicitation (RE) process should largely embrace human-centered perspectives, and my work focuses on changing human intentions and desires over time. To support software evolution due to requirement changes, Situ framework has been proposed to model and detect human intentions by inferring their desires through monitoring environmental contexts and human behavioral contexts prior to or after system deployment. Earlier work on Situ reported that the technique is able to infer users’ desires with a certain degree of accuracy using the Conditional Random Fields method. However, new intention identification and new requirements elicitation still primarily depends on manual analysis.

In this talk, I will discuss our attempt to find a computable way to identify users’ new intentions with limited help from human oracle. I will first discuss the feasibility of implementing the concept of Data-Information-Knowledge-Wisdom (DIKW) to bridge the gap between requirements and data pertaining to user behaviors and environmental contexts, and will then introduce our proposed situation-centric, knowledge-driven requirements elicitation approach using the Multi-strategy, Task-adaptive Learning (MTL) method and the Strategic Rationale (SR) model. Our case study shows that the proposed approach is able to identify users’ new intentions, and is especially effective to capture alternatives of low-level tasks. I will also demonstrate how these newly identified intentions can be fused to the existing domain knowledge network using the SR model, and harvest high-level wisdom, in terms of new requirements and design insights.