Browsing by Subject "Generalization"
Now showing 1 - 7 of 7
- Results Per Page
- Sort Options
Item The contribution of fear conditioning to pathological anxiety: an investigation of conditioned fear generalization in OCD Traits and PTSD(2014-08) Kaczkurkin, AntoniaA review of the literature demonstrates a lack of research on fear-generalization processes in many anxiety disorders including obsessive-compulsive disorder (OCD) and posttraumatic stress disorder (PTSD). Chapter 2 represents the first study that attempted to investigate the generalization of conditioned fear in individuals with obsessive-compulsive traits using startle EMG and behavioral measures. The results of this study demonstrated that individuals with high levels of Threat Estimation as measured by the Obsessive Beliefs Questionnaire (OBQ-44) displayed overgeneralization of fear responses to a greater range of stimuli resembling the danger cue than those with low levels of Threat Estimation. In addition, despite etiological theories proposing that fear conditioning and overgeneralization of fear play prominent roles in the development and maintenance of PTSD, little research had been done on the neurobiological mechanisms that contribute to fear conditioning processes in PTSD patients and none have been specifically conducted on generalization. Chapter 3 investigated the neurobiological substrates associated with the overgeneralization of conditioned fear in PTSD patients using behavioral, skin conductance, and functional magnetic resonance imaging (fMRI) measures. This study provides evidence that PTSD patients demonstrate overgeneralization of conditioned fear in the dorsal medial prefrontal cortex, bilateral insula, left and right caudate, left inferior parietal lobule, and right superior frontal gyrus. This body of work provides novel evidence regarding the generalization of conditioned fear in OCD and PTSD.Item Empirical Analysis of Optimization and Generalization of Deep Neural Networks(2022-03) Li, XinyanDeep neural networks (DNNs) have gained increasing attention and popularity in the past decade. This is mainly due to their tremendous success in numerous commercial, scientific, and societal tasks. Despite the success of DNNs in practice, several aspects of the optimization dynamics and generalization are still not well understood. In practice, DNNs are usually heavily over-parameterized with far more parameters than training samples, making them easier to memorize all the training examples without learning. In fact, Zhang et al. have shown that DNNs indeed can fit training data perfectly. Training DNNs also requires first-order optimization methods such as gradient descent (GD) and stochastic gradient descent (SGD) to solve a highly non-convex optimization problem. The fact that such heavily over-parameterized DNNs trained by simple GD/SGD are still able to learn and generalize well deeply puzzles the deep learning community. In this thesis, we explore the optimization dynamics and generalization behavior of over-parameterized DNNs trained by SGD from two unique directions. First, we focus on studying the topology of the loss landscape of those DNNs through the analysis of the Hessian of the training loss (with respect to the parameters). We empirically study the second-moment matrix $M_t$ constructed by the outer product of the stochastic gradients (SGs), as well as the Hessian of the loss $H_f(\theta_t)$. With the help of existing tools such as the Lanczos method and the R-operator, we can compute the eigenvalues and the corresponding eigenvectors of both the Hessian matrix and the second-moment matrix efficiently. This allows us to reveal the relationship between the Hessian of the loss and the second moment of SGs. Besides, we discover the ``low-rank'' structure in both the eigenvalue-spectrum of the Hessian and in the stochastic gradients themselves. Such observations directly lead to the development of a new PAC-bayes generalization bound which considers the structure of the Hessian at minima obtained from SGD, as well as a novel noisy truncated stochastic gradient descent (NT-SGD) algorithm, aiming to tackle the communication bottleneck in the large-scale distributed setting. Next, we dive into the debate on whether it is sufficient to explain the success of DNNs in practice by their behavior in the infinite-width limit. On one hand, there has been a rich literature understanding of the infinite width limit of DNNs. Such analysis simplifies the learning dynamics of very wide neural networks by a linear model obtained from the first-order Taylor expansion around its initial parameters. As a result, the DNN training occurs in a ``lazy'' regime. On the other hand, both theoretical and empirical evidence has been presented, pointing out the limitations of lazy training. Those results suggest that training DNNs with gradient descent actually occurs in a ``rich'' regime which captures much richer inductive biases and the behavior of such models cannot be fully described by their infinite-width kernel equivalence. As an empirical complement of the recent work studying the transition from the lazy regime to the rich regime, we study generalization and optimization behaviors of commonly used DNNs, focusing on varying the width and to some extent the depth, and show what happens in typical DNNs used in practice in the rich regime. We also extensively study the smallest eigenvalues of the Neural Tangent Kernel, a crucial element that appeared in many recently theoretical analyses related to both the training and generalization of DNNs. Hopefully, our empirical study could provide fodder for new theoretical advances on understanding generalization and optimization in both rich and lazy regimes.Item Enhancing maintenance and generalization of incremental rehearsal through theory-based modifications(2013-05) Petersen-Brown, ShawnaThe attainment of basic early literacy skills at an early age is one way to ensure children become proficient readers as adults. Word recognition is an important basic early literacy skill that is related to reading fluency and overall reading competency. Incremental rehearsal (IR) is a flashcard technique that has produced strong outcomes for a variety of outcomes including word recognition. Modifications to IR that are based on theory may enhance maintenance and generalization, outcomes that are frequently overlooked in flashcard intervention research. A number of theoretical frameworks were utilized in this research to enhance maintenance and generalization. The depth of processing framework was used, which suggests that semantic processing of information leads to better maintenance. The theory of common elements and Stokes and Baer's (1977) generalization framework were utilized as frameworks for enhancing generalization. In addition to intervention design, individual factors may also impact intervention efficacy. For example, working memory capacity may influence the amount of verbal information an individual can process. Decoding skill may also lead to enhanced outcomes in word recognition interventions. The current study examined the impact of theory-based modifications to IR on maintenance and generalization. A within-subjects design was utilized in which all participants were taught seven unknown words in each of three IR variants - IR, IR with vocabulary (IR-V, which leveraged the depth of processing framework), and IR with context (IR-C, which leveraged Stokes and Baer's [1977] generalization framework and the theory of common elements). Auditory working memory and decoding skills were also measured as individual factors and potential moderators of intervention effects. Primary dependent variables were maintenance and generalization at 1 and 2 weeks after teaching. Results of the study indicated that maintenance and generalization were high across conditions, suggesting possible ceiling effects. However, 1-week maintenance was significantly greater in IR-V than IR and in IR-C than IR. Additionally, 2-week generalization was significantly greater in IR-V than IR. Effect sizes were small across outcomes in favor of IR-V and IR-C. IR was the most efficient intervention variant in words maintained per minute. A moderating effect of auditory working memory capacity was not observed. However, participants with low decoding skill maintained significantly more words through IR-V than IR at 1 week, while this difference was not observed in participants at higher decoding skill levels. Future research may investigate theory-based modifications applied to different populations (such as students with low decoding skills) or different information types. In addition, future research may investigate theory-based modifications to more efficient modifications of IR (perhaps with fewer opportunities to respond) to enhance the efficiency of these approaches.Item Generalizations in Practice: Investigating Generality and Specificity in Developmental Biology(2023) Yoshida, YoshinariAlthough there is a consensus that pursuits of general knowledge are crucial in almost all fields of science, the majority of philosophical analyses of generalizations have focused narrowly on universal generalizations or laws of nature and what role generalizations play in scientific explanations. This narrow focus has limited the scope of philosophical discussions about scientific generalizations. This dissertation proposes and exemplifies a broader inquiry into scientific generalizations that is motivated by the question: how do scientists pursue, formulate, reason about, utilize, and communicate generalizations? In other words, how are generalizations practiced in science? To address this broad set of questions, I focus on a particular field—developmental biology—and examine investigative and representational practices surrounding generalizations. Like many other fields, developmental biology seeks both widely shared regularities and the details of causal processes peculiar to specific systems. My analyses show how this dual interest in generality and specific details is interconnected and mutually contribute to each other. This dissertation is organized as follows. Chapter 1 provides a brief overview of how philosophers have discussed generalizations. I point out that the interests in laws and explanation have dominated the past discussions. In contrast, my approach focuses on investigative and representational practices of generalization, which have received very little philosophical attention. Chapter 2 analyzes two approaches to generalizations in developmental biology: mechanisms and principles. These are distinguished based on the relevance of abstraction. I show that the two approaches are associated with different investigative practices. This analysis illustrates what forms of non-universal generalizations developmental biologists seek and formulate, which serves as a basis for discussions in the following chapters. Chapter 3 explores generalizations from the perspective of modeling desiderata. I offer a characterization of what I call multiple-models juxtaposition (MMJ), a strategy for managing a trade-off between generality and detail in scientific models. MMJ displays models of distinct processes together and fulfills different desiderata both in the individual models and by a comparison of those models. I also clarify the distinction between MMJ and multiple-models idealization (MMI), which also uses multiple models to manage trade-offs among desiderata. Chapter 4 focuses on the use of model systems. Biologists often study particular biological systems as models of a phenomenon of interest, even if they know that the phenomenon is produced by diverse mechanisms and hence none of those systems alone can sufficiently represent it. I argue that even if generalizability of results from a single model system is significantly limited, generalizations concerning specific aspects of mechanisms often hold across certain ranges of biological systems. This enables multiple model systems to jointly represent such a phenomenon. Chapter 5 considers the question “how and why do scientists generalize?” by challenging three influential assumptions: (1) generalizations are expressed linguistically; (2) scientists generalize by formulating a single representation with wide applicability; and (3) generalizations are valuable because they enable scientific explanations. My analysis of a concrete example illustrates roles that visual representations play in generalizations. It also shows that formulating a single, unified representation is not the only way to generalize; scientists often generalize by configuring multiple representations. Finally, I argue that generalizations serve to facilitate cross-fertilization among studies of different target systems, which complements the explanation-centered view.Item A move towards studying both pavlovian & instrumental contributions to conditioning abnormalities in the anxiety disorders(2013-07) van Meurs, Brian GregoryFear-conditioning experiments in clinical anxiety have focused almost exclusively on passive- emotional, Pavlovian conditioning, rather than active-behavioral, instrumental conditioning. Paradigms capable of eliciting both Pavlovian and instrumental conditioning are thus needed to experimentally study the maladaptive behavioral consequences of Pavlovian abnormalities. One such abnormality is overgeneralization of conditioned fear, a core feature of anxiety pathology. Such generalization can be assessed by studying generalization gradients and until now has only been examined using Pavlovian conditioning. The current study validates a novel paradigm which applies a validated Pavlovian generalization experiment in the context of a `virtual farmer' computer game in which the participant is a farmer whose task it is to successfully plant and harvest crops. While playing the game, shapes are superimposed on the screen with one such shape, paired with shock, serving as the conditioned danger cue (CS+). Generalization stimuli (GS), parametrically vary in similarity to the CS+, but are never paired with shock. While playing the game, participants are given the opportunity to avoid shock (instrumental conditioning) at the cost of poorer performance. Fear-potentiated startle (FPS), skin conductance responses (SCR) and online risk ratings were obtained and each displayed the expected Pavlovian generalization gradient. Instrumental avoidance responses also form a generalization gradient and are strongly associated with Pavlovian indices of generalization (FPS and risk ratings but not SCR). Additionally, FPS at acquisition was a significant predictor of subsequent avoidance behavior. This novel experimental tool will be useful in describing and testing individual differences associated with clinical anxiety.Item Neurocircuitry of Generalization of Avoidance Behavior following Pavlovian Conditioning in Adults with High and Low Trait Anxiety(2016-08) van Meurs, BrianOne of the cardinal features of many anxiety disorders is maladaptive avoidance. While behavioral avoidance is important for survival and adaptive when danger is present, in the absence of a threat it is maladaptive. Signaled avoidance depends on Pavlovian learning that a neutral conditioned stimulus signals an ensuing aversive unconditioned stimulus. Maladaptive signaled avoidance could therefore result from abnormalities in Pavlovian conditioning. Overgeneralization of conditioned fear is one such abnormality that has been demonstrated in several anxiety disorders. To assess the relationship between anxiety and generalization of signaled avoidance behavior, 22 participants, with a range of trait anxiety scores split into two group of high and low anxiety, completed a generalization gradient, approach-avoidance fMRI task following Pavlovian discrimination conditioning. Results indicated the expected curvilinear generalization gradient in avoidance responses and ratings of risk, with group differences in avoidance responses. There were several functional regions of interest which also demonstrated the expected curvilinear gradient as well as group differences in percent BOLD signal change across the gradient. This was true for both Pavlovian trials, as well as during the decision making stage of the Instrumental trials. There were also several regions in which activations were significantly related to avoidance behavior. These results indicate that individuals with higher levels of trait anxiety are at increased risk of ‘maladaptive’ avoidance of safe stimuli that resemble danger-cues. Moreover brain areas such as the anterior insula and subgenual anterior cingulate cortex, and primary visual cortex, which are involved in the Pavlovian generalization of fear, are also involved in the overgeneralization of the avoidance response. Additional unexpected findings highlight the role of the cerebellum, somatosensory cortex, and gender in production and maintenance of an avoidance response.Item Teaching for transfer of an evidence-based reading strategy: an experimental field trial.(2012-04) Pinto, Viveca VictoriaThe purpose of this study was to examine the effects of explicitly teaching for transfer of PALS, and to examine whether transfer training helped participants maintain the strategy taught. Sixty-two participants from two third-grade classrooms and one fourth-grade classroom participated in the study. A pretest-posttest-maintenance control group design was used, in which participants within classroom were matched on their one-min oral reading scores and assigned randomly to receive PALS plus transfer training (n = 31) or PALS practice only (n = 31). Proximal and distal measures were used to assess transfer. Proximal measures were defined as measures closely aligned to the intervention and included main idea identification of narrative and informational text. Distal measures assessed general reading competence and were not closely aligned to the strategy taught. A repeated measures ANOVA was conducted with time (pretest, posttest, and maintenance) as the within-subject factor, and treatment (PALS vs. PALS plus transfer training) as the between-subjects factor. On the main idea identification of narrative text, there was a significant main effect of time, but the main effect of condition and the interaction of time by condition was non-significant. On the main idea identification of informational text, there was a significant main effect of time, and the interaction of time by condition approached significance, with participants receiving PALS plus transfer training correctly identifying and producing more main idea statement (ES = .04). On the distal measures, there were significant main effects of time, but no significant effects of condition, or interactions of time by condition. Implications for educational practice and future research directions are discussed.