skip to main content


Title: Interpretable Debiasing of Vectorized Language Representations with Iterative Orthogonalization
We propose a new mechanism to augment a word vector embedding representation that offers improved bias removal while retaining the key information—resulting in improved interpretability of the representation. Rather than removing the information associated with a concept that may induce bias, our proposed method identifies two concept subspaces and makes them orthogonal. The resulting representation has these two concepts uncorrelated. Moreover, because they are orthogonal, one can simply apply a rotation on the basis of the representation so that the resulting subspace corresponds with coordinates. This explicit encoding of concepts to coordinates works because they have been made fully orthogonal, which previous approaches do not achieve. Furthermore, we show that this can be extended to multiple subspaces. As a result, one can choose a subset of concepts to be represented transparently and explicitly, while the others are retained in the mixed but extremely expressive format of the representation.  more » « less
Award ID(s):
2115677
NSF-PAR ID:
10465704
Author(s) / Creator(s):
; ; ; ; ; ; ; ; ;
Date Published:
Journal Name:
International Conference on Learning Representations
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. null (Ed.)
    Perception of limb position and motion combines sensory information from spindles in muscles that span one joint (monoarticulars) and two joints (biarticulars). This anatomical organization should create interactions in estimating limb position. We developed two models, one with only monoarticulars and one with both monoarticulars and biarticulars, to explore how biarticulars influence estimates of arm position in hand ( x, y) and joint ( shoulder, elbow) coordinates. In hand coordinates, both models predicted larger medial-lateral than proximal-distal errors, although the model with both muscle groups predicted that biarticulars would reduce this bias. In contrast, the two models made significantly different predictions in joint coordinates. The model with only monoarticulars predicted that errors would be uniformly distributed because estimates of angles at each joint would be independent. In contrast, the model that included biarticulars predicted that errors would be coupled between the two joints, resulting in smaller errors for combinations of flexion or extension at both joints and larger errors for combinations of flexion at one joint and extension at the other joint. We also carried out two experiments to examine errors made by human subjects during an arm position matching task in which a robot passively moved one arm to different positions and the subjects moved their other arm to mirror-match each position. Errors in hand coordinates were similar to those predicted by both models. Critically, however, errors in joint coordinates were only similar to those predicted by the model with monoarticulars and biarticulars. These results highlight how biarticulars influence perceptual estimates of limb position by helping to minimize medial-lateral errors. NEW & NOTEWORTHY It is unclear how sensory information from muscle spindles located within muscles spanning multiple joints influences perception of body position and motion. We address this issue by comparing errors in estimating limb position made by human subjects with predicted errors made by two musculoskeletal models, one with only monoarticulars and one with both monoarticulars and biarticulars. We provide evidence that biarticulars produce coupling of errors between joints, which help to reduce errors. 
    more » « less
  2. Human spatial concepts, such as the concept ofplace, are not immediately translatable to the geometric foundations of spatial databases and information systems developed over the past 50 years. These systems typically rest on the concepts of objects and fields, both bound to coordinates, as two general paradigms of geographic representation. The match between notions of place occurring in everydaywherequestions and the data available to answer such questions is unclear and hinders progress in place‐based information systems. This is particularly true in novel application areas such as the Digital Humanities or speech‐based human–computer interaction, but also for location‐based services. Although this shortcoming has been observed before, we approach the challenges of relating places to information system representations with a fresh view, based on a set of core concepts of spatial information. These concepts have been proposed in information science with the intent of serving human–machine spatial question asking and answering. Clarifying the relationship of the notion of place to these concepts is a significant step toward geographically intelligent systems. The main result of the article is a demonstration that the notion of place fits existing concepts of spatial information, when these are adequately exploited and combined.

     
    more » « less
  3. This research investigates the design of structurally performant, lightweight architectural elements produced through concrete 3D printing (C3DP). Traditionally, concrete requires dense and sturdy formwork, whose production adds significantly to the total cost and results in massive and heavy parts after demolding. C3DP offers the unique opportunity to both eliminate the need for formwork and to create lighter parts by introducing internal voids and cavities. The advent of additive manufacturing in a broad range of scales, materials, industries, and applications, led to increased interest and intense research into different types of porous structures, their geometry, and structural performance under various boundary conditions. Precise control over the sparse distribution of material allows not only for parts with similar strength at reduced mass but even for modifications of mechanical properties, like turning brittle materials into elastic or shock-absorbent ones. While with powder-based additive manufacturing processes like metal 3D printing, truss-based lattices have become very popular for the light-weighting of parts or to provide tissue growth scaffolds for medical implants, their geometry – a sparse space frame resulting in numerous individual contour islands and accentuated overhangs – cannot as easily be produced by C3DP, which is based on a continuous material extrusion. Alternative types of micro-structures, so-called triply periodic minimal surfaces (TPMS), are better suited for this process as they are, as their name suggests, consisting of one continuous surface dividing space into two separate but interwoven subspaces. TPMS are therefore very popular for the efficient design of heat exchangers. We develop and present a continuous and integrated workflow, in which the architectural elements and their structural requirements are designed through transitioning back and forth between the force and the form diagram using 3D graphic statics [1]. The members and their topology from the abstract graph of the conceptual form diagram are seamlessly connected to the volumetric modeling (VM) framework, responsible for the definition of the part geometry [2]. VM represents form assigned distance functions (SDF) and can easily handle complex topologies and flawless Boolean operations of not only the outer shell geometry but also the internal micro-structural infill patterns (Fig. 1, a). In an iterative feedback loop, the infill can be further optimized to leave the material only along certain internal stress trajectories (force flows). This functional grading controlling the relative density is done based on the FE analysis results. The stress distribution is thereby defined as a three-dimensional field (Fig. 1, b). Its values can factor into the SDF equation and be used to modify the wavelength (periodicity) of the TPMS, the local thickness of the surface shell, the solid to void fraction by shifting the threshold iso-value or even the alignment and orientation of the unit cells (Fig. 1, c). They can be arranged in an orthogonal, polar- or even spherical coordinate system to optimally adapt to structural necessities. The TPMS pattern can also gradually transition from one type into another type along the gradient of a spatial function. 
    more » « less
  4. This work proposes a new computational framework for learning a structured generative model for real-world datasets. In particular, we propose to learn a Closed-loop Transcriptionbetween a multi-class, multi-dimensional data distribution and a Linear discriminative representation (CTRL) in the feature space that consists of multiple independent multi-dimensional linear subspaces. In particular, we argue that the optimal encoding and decoding mappings sought can be formulated as a two-player minimax game between the encoder and decoderfor the learned representation. A natural utility function for this game is the so-called rate reduction, a simple information-theoretic measure for distances between mixtures of subspace-like Gaussians in the feature space. Our formulation draws inspiration from closed-loop error feedback from control systems and avoids expensive evaluating and minimizing of approximated distances between arbitrary distributions in either the data space or the feature space. To a large extent, this new formulation unifies the concepts and benefits of Auto-Encoding and GAN and naturally extends them to the settings of learning a both discriminative and generative representation for multi-class and multi-dimensional real-world data. Our extensive experiments on many benchmark imagery datasets demonstrate tremendous potential of this new closed-loop formulation: under fair comparison, visual quality of the learned decoder and classification performance of the encoder is competitive and arguably better than existing methods based on GAN, VAE, or a combination of both. Unlike existing generative models, the so-learned features of the multiple classes are structured instead of hidden: different classes are explicitly mapped onto corresponding independent principal subspaces in the feature space, and diverse visual attributes within each class are modeled by the independent principal components within each subspace. 
    more » « less
  5. BACKGROUND The past decade has witnessed considerable progress toward the creation of new quantum technologies. Substantial advances in present leading qubit technologies, which are based on superconductors, semiconductors, trapped ions, or neutral atoms, will undoubtedly be made in the years ahead. Beyond these present technologies, there exist blueprints for topological qubits, which leverage fundamentally different physics for improved qubit performance. These qubits exploit the fact that quasiparticles of topological quantum states allow quantum information to be encoded and processed in a nonlocal manner, providing inherent protection against decoherence and potentially overcoming a major challenge of the present generation of qubits. Although still far from being experimentally realized, the potential benefits of this approach are evident. The inherent protection against decoherence implies better scalability, promising a considerable reduction in the number of qubits needed for error correction. Transcending possible technological applications, the underlying physics is rife with exciting concepts and challenges, including topological superconductors, non-abelian anyons such as Majorana zero modes (MZMs), and non-abelian quantum statistics.­­ ADVANCES In a wide-ranging and ongoing effort, numerous potential material platforms are being explored that may realize the required topological quantum states. Non-abelian anyons were first predicted as quasiparticles of topological states known as fractional quantum Hall states, which are formed when electrons move in a plane subject to a strong perpendicular magnetic field. The prediction that hybrid materials that combine topological insulators and conventional superconductors can support localized MZMs, the simplest type of non-abelian anyon, brought entirely new material platforms into view. These include, among others, semiconductor-superconductor hybrids, magnetic adatoms on superconducting substrates, and Fe-based superconductors. One-dimensional systems are playing a particularly prominent role, with blueprints for quantum information applications being most developed for hybrid semiconductor-superconductor systems. There have been numerous attempts to observe non-abelian anyons in the laboratory. Several experimental efforts observed signatures that are consistent with some of the theoretical predictions for MZMs. A few extensively studied platforms were subjected to intense scrutiny and in-depth analyses of alternative interpretations, revealing a more complex reality than anticipated, with multiple possible interpretations of the data. Because advances in our understanding of a physical system often rely on discrepancies between experiment and theory, this has already led to an improved understanding of Majorana signatures; however, our ability to detect and manipulate non-abelian anyons such as MZMs remains in its infancy. Future work can build on improved materials in some of the existing platforms but may also exploit new materials such as van der Waals heterostructures, including twisted layers, which promise many new options for engineering topological phases of matter. OUTLOOK Experimentally establishing the existence of non-abelian anyons constitutes an outstandingly worthwhile goal, not only from the point of view of fundamental physics but also because of their potential applications. Future progress will be accelerated if claims of Majorana discoveries are based on experimental tests that go substantially beyond indicators such as zero-bias peaks that, at best, suggest consistency with a Majorana interpretation. It will be equally important that these discoveries build on an excellent understanding of the underlying material systems. Most likely, further material improvements of existing platforms and the exploration of new material platforms will both be important avenues for progress toward obtaining solid evidence for MZMs. Once that has been achieved, we can hope to explore—and harness—the fascinating physics of non-abelian anyons such as the topologically protected ground state manifold and non-abelian statistics. Proposed topological platforms. (Left) Proposed state of electrons in a high magnetic field (even-denominator fractional quantum Hall states) are predicted to host Majorana quasiparticles. (Right) Hybrid structures of superconductors and other materials have also been proposed to host such quasiparticles and can be tailored to create topological quantum bits based on Majoranas. 
    more » « less