Publications

We present a data-driven method for reconstructing the galactic acceleration field from phase-space measurements of stellar streams. Our approach is based on a flexible and differentiable fit to the stream in phase-space, enabling a direct estimate of the acceleration vector along the stream. Reconstruction of the local acceleration field can be applied independently to each of several streams, allowing us to sample the acceleration field due to the underlying galactic potential across a range of scales. Our approach is methodologically different from previous works, since a model for the gravitational potential does not need to be adopted beforehand. Instead, our flexible neural-network-based model treats the stream as a collection of orbits with a locally similar mixture of energies, rather than assuming that the stream delineates a single stellar orbit. Accordingly, our approach allows for distinct regions of the stream to have different mean energies, as is the case for real stellar streams. Once the acceleration vector is sampled along the stream, standard analytic models for the galactic potential can then be rapidly constrained. We find our method recovers the correct parameters for a ground-truth triaxial logarithmic halo potential when applied to simulated stellar streams. Alternatively, we demonstrate that a flexible potential can be constrained with a neural network, though standard multipole expansions can also be constrained. Our approach is applicable to simple and complicated gravitational potentials alike, and enables potential reconstruction from a fully data-driven standpoint using measurements of slowly phase-mixing tidal debris.

In the coming decade, thousands of stellar streams will be observed in the halos of external galaxies. What fundamental discoveries will we make about dark matter from these streams? As a first attempt to look at these questions, we model Magellan/Megacam imaging of the Centaurus A's (Cen A) disrupting dwarf companion Dwarf 3 (Dw3) and its associated stellar stream, to find out what can be learned about the Cen A dark-matter halo. We develop a novel external galaxy stream-fitting technique and generate model stellar streams that reproduce the stream morphology visible in the imaging. We find that there are many viable stream models that fit the data well, with reasonable parameters, provided that Cen A has a halo mass larger than M200 >4.70×1012 M⊙. There is a second stream in Cen A's halo that is also reproduced within the context of this same dynamical model. However, stream morphology in the imaging alone does not uniquely determine the mass or mass distribution for the Cen A halo. In particular, the stream models with high likelihood show covariances between the inferred Cen A mass distribution, the inferred Dw3 progenitor mass, the Dw3 velocity, and the Dw3 line-of-sight position. We show that these degeneracies can be broken with radial-velocity measurements along the stream, and that a single radial velocity measurement puts a substantial lower limit on the halo mass. These results suggest that targeted radial-velocity measurements will be critical if we want to learn about dark matter from extragalactic stellar streams.

The brain must extract behaviorally relevant latent variables from the signals streamed by the sensory organs. Such latent variables are often encoded in the dynamics that generated the signal rather than in the specific realization of the waveform. Therefore, one problem faced by the brain is to segment time series based on underlying dynamics. We present two algorithms for performing this segmentation task that are biologically plausible, which we define as acting in a streaming setting and all learning rules being local. One algorithm is model based and can be derived from an optimization problem involving a mixture of autoregressive processes. This algorithm relies on feedback in the form of a prediction error and can also be used for forecasting future samples. In some brain regions, such as the retina, the feedback connections necessary to use the prediction error for learning are absent. For this case, we propose a second, model-free algorithm that uses a running estimate of the autocorrelation structure of the signal to perform the segmentation. We show that both algorithms do well when tasked with segmenting signals drawn from autoregressive models with piecewise-constant parameters. In particular, the segmentation accuracy is similar to that obtained from oracle-like methods in which the ground-truth parameters of the autoregressive models are known. We also test our methods on data sets generated by alternating snippets of voice recordings. We provide implementations of our algorithms at https://github.com/ttesileanu/bio-time-series.

Denoising is a fundamental challenge in scientific imaging. Deep convolutional neural networks (CNNs) provide the current state of the art in denoising photographic images. However, their potential has been inadequately explored for scientific imaging. Denoising CNNs are typically trained on clean images corrupted with artificial noise, but in scientific applications, noiseless ground-truth images are usually not available. To address this, we propose a simulation-based denoising (SBD) framework, in which CNNs are trained on simulated images. We test the framework on transmission electron microscopy (TEM) data, showing that it outperforms existing techniques on a simulated benchmark dataset, and on real data. We analyze the generalization capability of SBD, demonstrating that the trained networks are robust to variations of imaging parameters and of the underlying signal structure. Our results reveal that state-of-the-art architectures for denoising photographic images may not be well adapted to scientific-imaging data. For instance, substantially increasing their field-of-view dramatically improves their performance on TEM images acquired at low signal-to-noise ratios. We also demonstrate that standard performance metrics for photographs (such as peak signal-to-noise ratio) may not be scientifically meaningful, and propose several metrics to remedy this issue in the case of TEM images. In addition, we propose a technique, based on likelihood computations, to visualize the agreement between the structure of the denoised images and the observed data. Finally, we release a publicly available benchmark dataset containing 18,000 simulated TEM images.

We introduce a family of neural quantum states for the simulation of strongly interacting systems in the presence of spatial periodicity. Our variational state is parametrized in terms of a permutationally invariant part described by the Deep Sets neural-network architecture. The input coordinates to the Deep Sets are periodically transformed such that they are suitable to directly describe periodic bosonic systems. We show example applications to both one- and two-dimensional interacting quantum gases with Gaussian interactions, as well as to $4He$ confined in a one-dimensional geometry. For the one-dimensional systems we find very precise estimations of the ground-state energies and the radial distribution functions of the particles. In two dimensions we obtain good estimations of the ground-state energies, comparable to results obtained from more conventional methods.

A conserved phase of gametogenesis is the development of oocytes and sperm within cell clusters (germline cysts) that arise through serial divisions of a founder cell. The resulting cell lineage trees (CLTs) exhibit diverse topologies across animals and can give rise to numerous emergent behaviors. Despite their centrality, sex-specific differences underlying the evolution and patterning of these cell trees are unknown. In Drosophila melanogaster, oocytes develop within a highly invariant and maximally branched 16-cell tree whose topology is constrained by the fusome – a branched membranous organelle critical for proper mitosis in females; the same division pattern and topology are widely thought to occur during spermatogenesis. Using highly-resolved three-dimensional reconstructions based on a supervised learning algorithm, we show that cell divisions in male cysts can deviate from the maximally branched pattern, leading to greater topological variability. Furthermore, in contrast to females, fusome fragmentation is common, suggesting germ cell mitoses can occur in its absence. These findings thus add to the repertoire of CLT formation strategies, highlighting the diversity of mechanisms employed during gametogenesis in the animal kingdom.

Quaternions are important for a wide variety of rotation-related problems in computer graphics, machine vision, and robotics. We study the nontrivial geometry of the relationship between quaternions and rotation matrices by exploiting the adjugate matrix of the characteristic equation of a related eigenvalue problem to obtain the manifold of the space of a quaternion eigenvector. We argue that quaternions parameterized by their corresponding rotation matrices cannot be expressed, for example, in machine learning tasks, as single-valued functions: the quaternion solution must instead be treated as a manifold, with different algebraic solutions for each of several single-valued sectors represented by the adjugate matrix. We conclude with novel constructions exploiting the quaternion adjugate variables to revisit several classic pose estimation applications: 2D point-cloud matching, 2D point-cloud-to-projection matching, 3D point-cloud matching, 3D orthographic point-cloud-to-projection matching, and 3D perspective point-cloud-to-projection matching. We find an exact solution to the 3D orthographic least squares pose extraction problem, and apply it successfully also to the perspective pose extraction problem with results that improve on existing methods.

Quaternions are important for a wide variety of rotation-related problems in computer graphics, machine vision, and robotics. We study the nontrivial geometry of the relationship between quaternions and rotation matrices by exploiting the adjugate matrix of the characteristic equation of a related eigenvalue problem to obtain the manifold of the space of a quaternion eigenvector. We argue that quaternions parameterized by their corresponding rotation matrices cannot be expressed, for example, in machine learning tasks, as single-valued functions: the quaternion solution must instead be treated as a manifold, with different algebraic solutions for each of several single-valued sectors represented by the adjugate matrix. We conclude with novel constructions exploiting the quaternion adjugate variables to revisit several classic pose estimation applications: 2D point-cloud matching, 2D point-cloud-to-projection matching, 3D point-cloud matching, 3D orthographic point-cloud-to-projection matching, and 3D perspective point-cloud-to-projection matching. We find an exact solution to the 3D orthographic least squares pose extraction problem, and apply it successfully also to the perspective pose extraction problem with results that improve on existing methods.

We propose Pathfinder, a variational method for approximately sampling from differentiable log densities. Starting from a random initialization, Pathfinder locates normal approximations to the target density along a quasi-Newton optimization path, with local covariance estimated using the inverse Hessian estimates produced by the optimizer. Pathfinder returns draws from the approximation with the lowest estimated Kullback-Leibler (KL) divergence to the true posterior. We evaluate Pathfinder on a wide range of posterior distributions, demonstrating that its approximate draws are better than those from automatic differentiation variational inference (ADVI) and comparable to those produced by short chains of dynamic Hamiltonian Monte Carlo (HMC), as measured by 1-Wasserstein distance. Compared to ADVI and short dynamic HMC runs, Pathfinder requires one to two orders of magnitude fewer log density and gradient evaluations, with greater reductions for more challenging posteriors. Importance resampling over multiple runs of Pathfinder improves the diversity of approximate draws, reducing 1-Wasserstein distance further and providing a measure of robustness to optimization failures on plateaus, saddle points, or in minor modes. The Monte Carlo KL divergence estimates are embarrassingly parallelizable in the core Pathfinder algorithm, as are multiple runs in the resampling version, further increasing Pathfinder's speed advantage with multiple cores.

We employ machine learning to decode the composition of unknown gas mixtures from the output of an array of four electrochemical sensors. The sensors use metal oxide electrodes paired with a ceramic electrolyte, yttria-stabilized zirconia (YSZ), to produce voltage responses to the presence of gases in complex mixtures. The voltages from the sensor array serve as inputs to a machine learning pipeline which first carries out multi-class classification of mixtures into types based on which gases are present at non-zero concentrations, and subsequently predicts gas concentrations given the mixture type. Thus, our model is able to take a single reading from the sensor array in response to gas mixtures involving NO, NO2, C3H8, and NH3, and output a highly accurate prediction of which gases are present in the mixture, along with the concentrations of each constituent gas. Our computational framework can be easily expanded to include additional gases and additional mixture types, allowing it to be used in numerous automotive, industrial and environmental monitoring settings.