37200 research outputs found
Sort by
Digging Deeper with Deep RAM Networks
While Deep Neural Networks (DNNs) have driven major breakthroughs in artificial intelligence, their internal complexity often makes their behavior hard to explain, resulting in the well-known “black box” dilemma. This thesis addresses the challenge of interpretability in DNNs and deep reinforcement learning (DRL) through two main contributions.
In Part I, we revisit and extend the use of Deep RAM Networks (DRNs) within the Arcade Learning Environment (ALE), showing that, with modern architectures and careful hyperparameter tuning, RAM-based agents can achieve performance competitive with established pixel-based baselines on Atari 2600 games, while offering additional advantages for research and analysis. We also train and evaluate a hybrid agent, that integrates both RAM and pixel observations, demonstrating that in most games it outperforms agents relying on either modality alone. By leveraging the compact and Markovian nature of RAM observations, DRNs not only act as competitive agents, but also enable new forms of analysis, making them particularly well-suited for interpretability studies.
In Part II, we dig deeper into agent behavior and DNN internals by introducing two general analysis and interpretability techniques. Trajectory Tracking provides a model-agnostic framework for examining long-term behavioral patterns, applicable to any reinforcement learning agent by querying basic trajectory attributes or attributes augmented by the user, within and across episodes. Neural Pathway Decomposition (NP-Decomp) offers a systematic approach for decomposing compact fully connected DNNs into their constituent neural pathways, tracing from input features and biases to final outputs. This method yields exact, context-specific attributions of how the collective of neural pathways influence an agent\u27s decisions. While computationally intensive, this method reveals structural insights not likely to be obtained through other analysis methods.
Applying these methods to DRN agents trained on games such as Breakout, we uncover insights invisible to score-based evaluation alone, such as the agent\u27s lapses in learned behaviors, the behavioral consequences of epsilon-greedy exploration, and the influence of particular RAM addresses on action selection. Trajectory Tracking and NP-Decomp together enable a step from assessing merely what an agent achieves score-wise, towards understanding how and why it behaves as it does.
By combining compact RAM-based architectures with deep analytical tools, this thesis lays the foundation for more transparent and interpretable reinforcement learning systems. Trajectory Tracking enables us to survey and explore the landscape of agent behavior over time, while NP-Decomp allows for the extraction of analytical core samples, cross-sections through the network\u27s internal structure that reveal not just what decisions an agent makes, but how those decisions emerge from the collective influence of many neural pathways
The Center Cannot Hold: Structures of Political Choice
In times of political polarization, we are well-advised to heed Spinoza’s injunction, “Not to lament, not to curse, but to understand.” Understanding can be facilitated by ideas about the archetypal structures of political choice. The primary question is whether available or desirable political positions are dyadic or triadic, and if dyadic, whether they are good versus bad or complementary. Various modes of polarization are illustrated by locating two or more positions on a line, horseshoe, or circle. It behooves us to grasp the structure of the political situation we face to avoid making intellectual and/or moral mistakes
The Geography of Rock Glaciers in the American West
Rock glaciers are relatively common in conterminous United States west of the 100° meridian, yet their distribution and number are not well known. Using aerial and satellite imagery, we inventoried the population of rock glaciers, yielding 2,257 rock glaciers totaling 234.28 km2 with an area of uncertainty of 7 percent. Of these, 1,561 (155.44 km2) are considered intact (moving) and 696 (76.84 km2) relict. Based on published estimates of ice content, total water-equivalent volume is roughly 2.52 km3. Rock glacier elevations increase from the northwest (Washington) to the southeast (Colorado), following a spatial trend of decreasing winter precipitation and increasing summer air temperature. Compared to alpine glaciers, rock glaciers are found in warmer, drier environments where winter precipitation \u3c 1,200 mm, and relict rock glaciers inhabit warmer and drier conditions than intact rock glaciers. Locally, alpine glaciers exist at higher, colder, and wetter elevations than rock glaciers, and intact rock glaciers are found at higher elevations that relict ones. We also noted 7,850 “features of interest” that appeared to be nonflowing ice-rich landforms, many of which are similar in form to protalus ramparts and found in a climate common to relict rock glaciers
Preparing Teachers to Seek Out Student Voice: Grounding the Work in Critical Consciousness
Seeking out student voice is a powerful and beneficial practice that requires teachers to be thoughtful and intentional. While students may feel validated when teachers respond to their input, soliciting student voice can also be harmful if not done carefully. It is necessary to use a critically conscious lens when soliciting student voice and responding to their input and feedback to best support students with historically marginalized identities. Teachers must be prepared to seek out multiple perspectives, deeply listen, and take critical action in response to the information that the students provide. By doing so, teachers can create empowering learning environments where students feel seen and heard
Real-World Evidence for Herbal Medicine Benefit in 9728 Type 2 Diabetes Patients-Peridonotitis Risk and Ambulatory Care Utilization.
Recent evidence manifests that individuals with type 2 diabetes (T2D) are increasingly affected by oral disorders. Although medicinal herbs have shown promise in managing T2D, their benefit in managing periodontitis risk and subsequent healthcare utilization remains uncertain
Human-Drought Systems
Advancing our understanding of the processes linking social and hydrological systems is an essential step toward managing drought risk and increasing drought resilience. Over the past decade, numerous studies have advanced our understanding of human influences on drought propagation, drought responses and their immediate and long-term consequences, and decisions and conditions leading to drought resilience. These advances have been achieved through: (1) the development of datasets that document drought progression, hazards, risk, vulnerability, and adaptation capacity; (2) creation of new modeling methods and application of models to build and test theory; and (3) empirical analyses from in depth case studies to large comparative analyses. While these categories are not mutually exclusive, they provide a way to track scientific progress, build a more fundamental understanding of human-drought systems to advance our capacity to generalize findings, and overcome systematic silos to develop more integrated drought management strategies that can span traditional sectoral and jurisdictional boundaries. The advances to date have enabled researchers to identify common phenomena describing human-drought interactions across space and time. Research on the influence of human actions on drought propagation sets the foundation for further work that can build generalized knowledge about the interrelationships within the human-drought system, across related systems, and in connection with other hazards in the context of many types of human action
Future Vision and Action Recommendations
Let us imagine a future world where water management is more fully integrated and supported by the science of sociohydrology and related metadisciplines. This is a vision for the future where the social, cultural, environmental, and economic values of water are explicitly considered through the lens of the science–water policy interface. The ultimate value of research will be measured by how much it supports improvements in water resources and hazards management and, in particular, completes and enriches the science–water policy interface
Teaching Research as an Interdependent Practice: Bringing Alternate Models into the Scholarly Conversation
The customs and practices of research in academia are often mischaracterized as individualistic pursuits, ignoring the reality that research requires a great deal of collaboration. As teaching librarians, our research instruction often reinforces individualism as a scholarly virtue, even as we simultaneously engage with themes of connection between researchers and their ideas. We propose an approach to teaching research and information literacy grounded in the concept, intention, and practice of interdependence rather than individualism. In doing so, we will call in models of interdependence from critical disability studies, relational-cultural theory, Indigenous Ways of Knowing, and the Cite Black Women movement. These models both describe and exemplify interconnected and inclusive approaches to research, which we can apply to our practice of teaching in academic libraries
The Library Instructor as Learner: A Survey of Reflective Teaching Practices in US Academic Libraries
Reflective practice provides library workers with a critical opportunity to examine professional experiences, question assumptions and approaches, explore new perspectives, and develop innovative solutions to existing problems. When applied to instruction as reflective teaching, this practice better situates library instructors to meet the evolving needs of twenty-first-century library users. This research study explores how library workers engage with reflective teaching in academic libraries across the United States. A survey was distributed to academic library communities, and a total of 153 responses were collected. While 92% of respondents reported participating in reflective teaching practice in a variety of instructional contexts, results indicated that respondents have utilized a broad range of methods for reflection and experienced numerous benefits and barriers. These diverse experiences suggest that library instructors see value in reflective work and would benefit from greater training and opportunities to participate in this practice
The King\u27s Wolf
The King\u27s Wolf is a queer medieval fantasy retelling of the 12th century Breton lai Bisclavret by Marie de France. My version of the story seeks to expand upon and make explicit the queerness that can be found in the original text. As such, The King\u27s Wolf is, at its core, a romance between a werewolf and his king.
The novel is still in the early stages of development. Included here are the first three chapters, told from the perspectives of two of the central characters, as well as a craft statement detailing how the story came to be and some of the challenges I have faced thus far in its development