22689 research outputs found
Sort by
Theory and Algorithms for Adversarially Robust Machine Learning via Geometric Properties of Data Distributions
Get PDFMachine learning is being increasingly used in many security sensitive tasks. However, the performance of these systems has been shown to fall drastically under specially crafted small changes to their inputs called adversarial examples. This has caused an arms race between defenders proposing adversarially robust systems, and attackers breaking these proposals. In practice, defenses rely on poorly understood training tricks, and attackers always seem to win this race. Even more, recent theoretical results suggest that the defenders' failure is inevitable, and that any classifier is susceptible to small adversarial examples.
In this thesis, we demonstrate that the reality is not so grim -- our main message is that the theory and practice of robustness can be improved by tightly incorporating the rich structure abundant in natural data distributions, \eg, natural images approximately lie on low-dimensional manifolds.
Our first contribution demonstrates that under a locally-linear structure on the data distribution, we can utilize game theory to analyse the attacker-defender arms race and obtain the optimal strategies, which happen to resemble well known attacks and defenses. Notably, the optimal defense generalizes randomized smoothing (RS), a popular method for defending classifiers against adversarial examples by smoothing their decision boundaries. However, RS's empirical performance relies on poorly understood training tricks.
Our second contribution identifies a key structural property of the data distribution, which we call \emph{interference distance}, and demonstrate that it characterizes the performance of RS. By informing the strength of smoothing by the interference distance of the underlying data distribution, we improve upon defenses agnostic to such structure.
Our third set of contributions generalize the above by identifying the property of localization, which measures the extent to which a data distribution localizes on small-volume regions of the input space. We prove that localization of a data distribution characterizes the robustness of the optimal defense. In particular, the existence of humans as a robust classifier suggests certain geometric properties of the support of the data distribution, which we utilize to construct practical classifiers that are equipped with mathematical guarantees of robustness against various families of attack algorithms.
Our contributions demonstrate the power of modelling the data distribution for adversarially robust classification, and pave the way for future research on building systems robust to realistic threat models by utilizing the rich geometry abundant in natural data distributions
A DATA-DRIVEN APPROACH FOR MICROSTRUCTURAL INVERSE DESIGN THROUGH MATERIALS INFORMATICS
No full textMachine learning (ML) has revolutionized computer vision, natural language processing, and other fields. However, it could also make important contributions to more traditional areas such metallurgy in which progress is often incremental. With the fast-moving development of technology, reducing the development time to bring new materials to market becomes an important topic. For instance, Apple has an Alloy Engineering Team dedicated to the application of computational and systematic methods to speed up the design of high-strength alloys for consumer products. Tesla and SpaceX have similarly invested in computational tools to facilitate development and production of innovative materials for components of electric vehicles and spacecraft.
Traditionally, material engineers would think about developing a material through a physics-based ``cause and effect'' (forward) process, which starts from a processing schedule that generates a specific material structure, which results in certain properties and performance. Product designers, in contrast, have in mind a set of design goals that require specific materials properties. This requires an inverse process, working backwards from the properties to the structure and ultimately to the necessary processing. For example, high strength is a desirable quality for the material making up the case of an electronic device. The strength of a metallic alloys is significantly affected by the distribution and size of particles (precipitates) in the matrix of parent material. For example, precipitates provide barriers to motion of crystal defects, and having more precipitates therefore increases strength. The distribution of these particles is, in turn, strongly correlated with processing variables such as the duration and temperature of heat treatments.
Integrated Computational Materials Engineering (ICME) is an emerging interdisciplinary approach that implements multi-scale modeling to accelerate the development of new materials. Although efficient forward simulation models exist to predict the structure (and hence properties) that will result from a given processing schedule, design with these models relies on iterative experimentation and simulation which yields incremental improvements in performance.
We use machine learning to create an integrated modeling framework to automate the inverse process and quickly find the processing parameters necessary to produce a material with a specified target structure. We demonstrate the utility of our approach through a case study in which we manipulate the resistance to spall failure of a commercial aluminum alloy. The choice of spall failure as a target property is motivated by the availability of high-throughput techniques for measuring spall strength. But our approach to the inverse problem of materials design is general, and could be applied to any other property or combination of properties of interest. If successful, material design though the inverse process described here can be dramatically accelerated over current approaches, significantly decreasing the time required to bring new products to market
NEURAL CONTROL OF THE TONGUE IN THE COMMON MARMOSET
No full textFlexible control of the tongue is critical across a wide range of human behaviors, including speech, mastication, respiration, and swallowing. Lingual movements exhibit dexterity, precision, and coordinated interactions with other oro-motor systems, making them an intriguing model for studying fine motor control. However, measuring tongue movements is challenging, and there is no established primate model for studying the neural basis of lingual control.
Here, we present the common marmoset as the first nonhuman primate model for the study of goal-directed lingual movements across two neuroscientific domains: the neuroeconomics of lingual control and the neurophysiology of lingual control by the cerebellum. Marmosets are capable of remarkable dexterity of their tongue, a skill they use in the wild to feed through extraction of sap from small holes made in trees. In this work, we leverage this ability in a foraging task in which marmosets are trained to produce goal-directed licks for food reward, while we record electrophysiological activity from the cerebellum.
In Chapter 2, we examine the effects of economic variables such as reward and effort on foraging-based decision-making and movement vigor of both saccades and licks. We found that when the acquisition of reward became effortful, the pupils constricted, the decisions exhibited delayed gratification, and the movements displayed reduced vigor. This coordinated response suggests that decisions and actions are part of a single control policy that aims to maximize a variable relevant to fitness: the capture rate.
In Chapter 3, we shift our focus to the cerebellum, examining the mechanism by which the principal cells of the cerebellar cortex, P-cells, contribute to control of the tongue. We identified lingual regions in lobule VI of the cerebellar vermis and recorded from P-cells as subjects performed goal-directed licks. We found that when P-cell simple spike activity was suppressed following a complex spike, licks exhibited endpoint hypermetria. Further, we found that P-cell simple spike population firing rate responses exquisitely tracked lick deceleration, scaling with lick kinematics. These results suggest that P-cells contribute to the endpoint accuracy of a lick through enabling timely deceleration of movement
Towards Robust Natural Language Processing to Promote Health Equity
Get PDFNatural language processing (NLP) has rapidly become an integral component of contemporary healthcare infrastructure and is likely to become more deeply entrenched in the domain given the rise of large language models. As such, we find ourselves at a crossroads where it is perhaps more important than ever to reflect on the broader purpose of such systems. While much focus has been placed on the potential for NLP to improve general health outcomes (e.g., diagnostic accuracy, efficiency of care), there also exists an understated opportunity and responsibility to use this technology to promote health equity. How exactly do we accomplish this given the inherently complex and multifaceted nature of health disparities?
In this dissertation, I identify and make progress along two orthogonal dimensions for promoting health equity using NLP. As the first dimension, I consider the development of tools that leverage NLP to augment or supplant (potentially biased) decision making in health care. In this setting, there exists a need for "defensive" methods that ensure NLP systems behave robustly across populations (e.g., patient populations, time periods). We ask and answer two questions – 1) how can we identify and measure sampling induced artifacts that arise in health-oriented training data, and 2) how can we counteract the effects of these artifacts or of more sweeping distribution shifts to promote model robustness? As the second dimension, I consider the development of tools that leverage NLP to detect or measure implicit bias in health care. Here, there is a need for "proactive" methods that translate hypotheses regarding implicit bias and discrimination into machine learnable tasks. We ask and answer: to what extent do these translations align with the broader pool of NLP research concerning implicit bias
The role of men’s gender egalitarian attitudes in men’s and couples’ reproductive desires and contraception decision-making in Nigeria
Get PDFBackground: Globally, there is substantial evidence that measurements tools for men's gender-egalitarian attitudes are limited despite strong evidence of their impact on reproductive health outcomes. In sub-Saharan African countries such as Nigeria, non-egalitarian gender norms could be contributing to persistently low modern contraceptive use and high fertility levels. This dissertation aimed at developing a measurement for men’s gender egalitarian attitudes and investigating the relationships between men’s gender attitudes, their fertility desires, couples’ fertility desires and contraception decision-making in Nigeria, a country marked by significant socio-cultural diversity.
Methods: This dissertation used data from the 2018 Nigeria Demographic Health Survey (NDHS). The Men’s Gender Egalitarian Attitudes Scale (MGEAS) was developed based on 10 items using data from 8,057 married men which also allowed to investigate the relationship between gender attitudes, urbanicity, education and region of residence. To explore the relationships between couples’ fertility desires and contraception decision-making dynamics, samples of 5,388 couples interviewed for gender-based violence and 3,110 couples eligible for contraception were used respectively. Descriptive analyses were used to explore sample characteristics and MGEAS score distributions. Weighted bivariate and multivariable logistic regression models were employed to investigate these relationships.
Results: The MGEAS revealed three distinct dimensions namely “Opposition to violence”, “Upholding equal decision-making power” and “Affirmation of women’s contraception autonomy”. Further analysis revealed that men with higher gender-egalitarian attitudes were likely to report wanting smaller families and were more likely to report concordant fertility desires with their wives. Men’s gender egalitarian views contributed to greater joint contraception decision-making, while less egalitarian views increased female or male driven decisions. Education levels of both partners, religion, wealth, and regional differences were also significant predictors of joint decision-making. Sensitivity analyses confirmed the robustness of these associations.
Conclusion: The MGEAS scores can predict men’s fertility desires, couples’ fertility desires and contraception decision-making behaviors. Higher gender-egalitarian attitudes among men are associated with smaller family size desires and greater joint contraception decision-making. These results demonstrate the need to implement measurements for men’s gender attitudes and their pivotal role in promoting gender equality while enhancing family planning programs and policies in Nigeria
Thermodynamic consequences of the coupling between structural reorganization and the ionization of buried residues in proteins
Get PDFIonizable amino acids buried in hydrophobic environments in proteins are rare but play essential functional roles in energy transduction processes.1 Buried ionizable groups have unusual physical properties, and the impact of dehydration of charged moieties in hydrophobic environments is poorly understood.2–5 Previously in this laboratory, we demonstrated that the pKa values of Asp, Glu, and Lys residues buried in the interior of staphylococcal nuclease (SNase),6,7 can be shifted by as many as 5 pH units relative to their normal pKa values in water, always in the direction that favors the neutral state. Most of the buried Lys, Glu, and Asp residues are neutral at physiological pH range. The correlation of the anomalous pKa values of these buried ionizable residues with the properties of their microenvironments revealed by crystal structures suggests that the pKa values cannot be determined from the electrostatic and chemical properties of their microenvironments in the folded structure. Previous NMR spectroscopy studies in this laboratory have shown that the ionization of these buried groups is coupled to local, sub-global, or global reorganization of the protein backbone. The ionization of the buried group appears to involve the transfer of the neutral group buried inside the hydrophobic protein interior to an environment where the charged form of the group is in contact with water.8–12
The data from this laboratory and others collectively suggest that the most important determinant of the pKa values of buried ionizable groups in proteins is the propensity of the protein backbone to reorganize in response to changes in pH. In this dissertation, I test the hypothesis that the pKa values of ionizable residues are determined by the thermodynamic stability of the protein because the ionization of buried groups is coupled to structural reorganization of the protein backbone. This hypothesis is based on the fact that the propensity for the protein to reorganize to alternative conformations (i.e. local, sub-global, or global unfolding) that is governed by the thermodynamic stability of the protein (∆Gconf) in the folded and denatured states.13,1
"We Did Our Part": The Roles and Influences of Women at the U.S. Capitol on January 6, 2021
Get PDFTo the detriment of national security, the women who contributed to the violent extremism of January 6, 2021 at the United States Capitol have largely been ignored, dismissed, misrepresented or sensationalized. Recent scholarship has begun to complicate our understanding of how women participate in violent extremism, yet large-scale empirical studies that explore women’s specific and potentially overlapping or contradicting contributions to violent extremism remain rare. Through a quantitative content analysis of 113 cases of women who allegedly participated at the Capitol and a qualitative analysis of select cases, this study reveals that the women participants of January 6th actively contributed to and influenced the violent extremism at the Capitol through participation in a variety of ways beyond nonviolent, supporting roles, indicating a contrast to traditional gendered assumptions. By illustrating the overlapping of roles along with nuances revealed in each case, this study additionally demonstrates that women who play more traditional, nonviolent supporting roles may also participate in violent extremism in a number of other influential ways, including as violent participants. Future researchers should aim to further unpack the nuances of women’s modes of participation and influence through interviews and more extensive case studies, as well as through larger quantitative analyses when more data becomes available. To ignore women’s roles and influences on January 6th inhibits researchers and policymakers alike from drawing a complete picture of how the violence at the Capitol occurred. If we are to prevent the next January 6th, greater attention toward this gap is essential
Compositional Neuro-Symbolic Reasoning over Natural Language
Get PDFMuch of classical artificial intelligence in the tradition of McCarthy (1959) pursued symbolic theorem proving over carefully hand-crafted ontologies of facts and rules. Due to challenges in scalability and expressivity, such reasoning has fallen out of favor in light of data-driven large language models (LLMs), which capture vast amounts of knowledge and reasoning power by learning patterns over natural language (NL) text corpora. Yet, aspects of the symbolic theorem proving paradigm – namely, mechanistic systematicity, groundedness, and logical underpinnings – are lost in most applications of LLMs, leading to common issues such as hallucinations. This thesis revisits theorem proving in the context of recent advances in LLM reasoning and knowledge retrieval, aiming to leverage the strengths of both the neural and symbolic traditions while addressing their respective pitfalls. I develop a neuro-symbolic entailment engine that reasons over natural language, building on the skeleton of a symbolic prover but leveraging neural models to reason flexibly and dynamically. The entailment engine is designed to systematically determine whether the natural language (NL) hypothesis is compositionally entailed by pieces of evidence provided in large and eadily available text corpora. Through its symbolic search, the engine constructs complex and interpretable entailment trees explaining how complex inferences follow from recursive multi-hop granular inferences, leveraging recent advances in information retrieval and modular reasoning with language models.
While “symbol pushing” engines are semi-guaranteed to arrive at proofs if they exist, the neural variant that I explore in this thesis does not have such a guarantee. NL is infinitely expressive, and by creating an NL proof search, we introduce numerous challenges regarding the kinds of inferences and search branches that are most viable to reason over for a given problem. Moreover, by using neural LLMs to generate inferences, we face challenges in getting them to reason robustly. This thesis explores ways to tackle this large and intractable search problem: using a noisy recent tool for reasoning (LMs) for mechanistic backward chaining proof search
Optimizing community-based strategies for tuberculosis screening
Get PDFIn 2022, approximately 10.6 million people fell ill with tuberculosis (TB) worldwide but only 7.5 million were diagnosed and referred for care. To help close the case detection gap, the World Health Organization (WHO) has identified several high-risk groups who should be systematically screened for TB, including 1) household contacts of people diagnosed with TB, and 2) people living in high TB prevalence areas. Targeting TB screening among these two high-risk populations requires community-based implementation strategies that are both feasible and effective. I sought to generate policy-relevant evidence for optimizing community-based strategies for TB screening in high TB burden settings.
In Aim 1, we conducted a non-inferiority trial (TB Aftermath) in India to assess whether phone screening is non-inferior to home-based screening after TB treatment. We randomized TB survivors to receive phone or home-based symptom screening at 6- and 12-months post-treatment. Overall, we found that phone was non-inferior to home-based screening for detecting TB among all household members. However, for detecting recurrence, home-based screening was more effective.
In Aim 2, we used TB Aftermath data to develop a parsimonious model for predicting recurrence that can help target post-TB screening among high-risk households. The best-performing model included unhealthy alcohol use, peak expiratory flow, body mass index, monthly income, and more than one TB episode. Our five-item tool, measurable at treatment completion, showed moderate predictive accuracy for recurrent TB and performed better for men compared to women.
In Aim 3, we assessed the accuracy of c-reactive protein (CRP) for detecting TB in Uganda. While CRP did not meet WHO TB screening benchmarks in the community setting, it demonstrated high specificity, and sensitivity was high among individuals with high bacillary burden who are likely to be most infectious. In an ambulatory care setting, estimated sensitivity and specificity were each within four percentage points of WHO benchmarks with no meaningful difference in performance by human immunodeficiency virus (HIV) status.
Together, these results inform the optimization of community-based strategies for TB screening. At scale, such strategies could result in substantial reductions in time to TB diagnosis and treatment, thereby improving patient outcomes and reducing transmission
Fichte’s Theory of Property and Political Freedom
No full textIn this work, I analyze Fichte’s theory of property with a focus on how Fichtean agents solve the coordination problem that arises from their expression of free efficacy. A coordination problem exists when agents share a joint end, but the optimal choice for one depends on the others’ choice. It is resolved only when all parties’ expectations of each other’s decisions converge. The problem of establishing a relation of right in Fichte’s account can be understood as a coordination problem. Agents share an interest in establishing norms that constrain others’ behaviors in their shared sphere of efficacy, allowing them to make and execute rational plans without interference. However, they also prefer different allocations of such constraints. The need to reach a stable coordination solution through bargaining has implications both for the allocation of external freedom and the adoption of institutions that can facilitate peaceful negotiation.
The main idea advanced in my reading of Fichte is that Fichte’s philosophy of right is oriented towards “negative freedom”–freedom from interference. I therefore challenge both the long-standing interpretive tradition that portrays him as a totalitarian thinker who advocates the total subjugation of the individual to the collective reason embodied by a coercive state, and the recent attempts in Anglophone scholarship to align Fichte with liberalism by presenting him as endorsing distributive justice. I argue that from 1793 to 1800, Fichte has consistently identified freedom from interference as the raison d'être for a system of right. Against the totalitarian reading, I argue that Fichte envisions the government’s power being checked by coordinated resistance from politically engaged citizens and has offered an account of the institutions that can facilitate this. Against the liberal reading, I argue that Fichte does not consider distributive justice as an independent value, but rather as pro-tanto constraints that contribute to the stability of a coordination scheme aimed at protecting noninterference. I thus present Fichte as offering a libertarian-left perspective on property rights and political freedom, a position on the political spectrum that perhaps deserves to be explored more than it has been