arXiv subjects

Michael Zhang

Publications and source records attributed to Michael Zhang.

At least 91 records · Page 5Linked to original sources

arXiv2022

Non-detection of He I in the atmosphere of GJ1214b with Keck/NIRSPEC, at a time of minimal telluric contamination

Observations of helium in exoplanet atmospheres may reveal the presence of large gaseous envelopes, and indicate ongoing atmospheric escape. Orell-Miquel et al. (2022) used CARMENES to measure a tentative detection of helium for the sub-Neptune GJ 1214b, with a peak excess absorption reaching over 2% in transit depth at 10830 Angstroms. However, several non-detections of helium had previously been reported for GJ 1214b. One explanation for the discrepancy was contamination of the planetary signal by overlapping telluric absorption- and emission lines. We used Keck/NIRSPEC to observe another transit of GJ 1214b at 10830 Angstroms, at a time of minimal contamination by telluric lines, and did not observe planetary helium absorption. Accounting for correlated noise in our measurement, we place an upper limit on the excess absorption size of 1.22% (95% confidence). We find that the discrepancy between the CARMENES and NIRSPEC observations is unlikely to be caused by using different instruments or stellar activity. It is currently unclear whether the difference is due to correlated noise in the observations, or variability in the planetary atmosphere.

astro-ph.EP↗

arXiv2022

Shoring Up the Foundations: Fusing Model Embeddings and Weak Supervision

Foundation models offer an exciting new paradigm for constructing models with out-of-the-box embeddings and a few labeled examples. However, it is not clear how to best apply foundation models without labeled data. A potential approach is to fuse foundation models with weak supervision frameworks, which use weak label sources -- pre-trained models, heuristics, crowd-workers -- to construct pseudolabels. The challenge is building a combination that best exploits the signal available in both foundation models and weak sources. We propose Liger, a combination that uses foundation model embeddings to improve two crucial elements of existing weak supervision techniques. First, we produce finer estimates of weak source quality by partitioning the embedding space and learning per-part source accuracies. Second, we improve source coverage by extending source votes in embedding space. Despite the black-box nature of foundation models, we prove results characterizing how our approach improves performance and show that lift scales with the smoothness of label distributions in embedding space. On six benchmark NLP and video tasks, Liger outperforms vanilla weak supervision by 14.1 points, weakly-supervised kNN and adapters by 11.8 points, and kNN and adapters supervised by traditional hand labels by 7.2 points.

stat.ML↗

arXiv2022

Perfectly Balanced: Improving Transfer and Robustness of Supervised Contrastive Learning

An ideal learned representation should display transferability and robustness. Supervised contrastive learning (SupCon) is a promising method for training accurate models, but produces representations that do not capture these properties due to class collapse -- when all points in a class map to the same representation. Recent work suggests that "spreading out" these representations improves them, but the precise mechanism is poorly understood. We argue that creating spread alone is insufficient for better representations, since spread is invariant to permutations within classes. Instead, both the correct degree of spread and a mechanism for breaking this invariance are necessary. We first prove that adding a weighted class-conditional InfoNCE loss to SupCon controls the degree of spread. Next, we study three mechanisms to break permutation invariance: using a constrained encoder, adding a class-conditional autoencoder, and using data augmentation. We show that the latter two encourage clustering of latent subclasses under more realistic conditions than the former. Using these insights, we show that adding a properly-weighted class-conditional InfoNCE loss and a class-conditional autoencoder to SupCon achieves 11.1 points of lift on coarse-to-fine transfer across 5 standard datasets and 4.7 points on worst-group robustness on 3 datasets, setting state-of-the-art on CelebA by 11.5 points.

stat.ML↗

arXiv2022

Contrastive Adapters for Foundation Model Group Robustness

While large pretrained foundation models (FMs) have shown remarkable zero-shot classification robustness to dataset-level distribution shifts, their robustness to subpopulation or group shifts is relatively underexplored. We study this problem, and find that FMs such as CLIP may not be robust to various group shifts. Across 9 robustness benchmarks, zero-shot classification with their embeddings results in gaps of up to 80.7 percentage points (pp) between average and worst-group accuracy. Unfortunately, existing methods to improve robustness require retraining, which can be prohibitively expensive on large foundation models. We also find that efficient ways to improve model inference (e.g., via adapters, lightweight networks with FM embeddings as inputs) do not consistently improve and can sometimes hurt group robustness compared to zero-shot (e.g., increasing the accuracy gap by 50.1 pp on CelebA). We thus develop an adapter training strategy to effectively and efficiently improve FM group robustness. Our motivating observation is that while poor robustness results from groups in the same class being embedded far apart in the foundation model "embedding space," standard adapter training may not bring these points closer together. We thus propose contrastive adapting, which trains adapters with contrastive learning to bring sample embeddings close to both their ground-truth class embeddings and other sample embeddings in the same class. Across the 9 benchmarks, our approach consistently improves group robustness, raising worst-group accuracy by 8.5 to 56.0 pp over zero-shot. Our approach is also efficient, doing so without any FM finetuning and only a fixed set of frozen FM embeddings. On benchmarks such as Waterbirds and CelebA, this leads to worst-group accuracy comparable to state-of-the-art methods that retrain entire models, while only training $\leq$1% of the model parameters.

cs.LG↗

arXiv2022

Triangle and Four Cycle Counting with Predictions in Graph Streams

We propose data-driven one-pass streaming algorithms for estimating the number of triangles and four cycles, two fundamental problems in graph analytics that are widely studied in the graph data stream literature. Recently, (Hsu 2018) and (Jiang 2020) applied machine learning techniques in other data stream problems, using a trained oracle that can predict certain properties of the stream elements to improve on prior "classical" algorithms that did not use oracles. In this paper, we explore the power of a "heavy edge" oracle in multiple graph edge streaming models. In the adjacency list model, we present a one-pass triangle counting algorithm improving upon the previous space upper bounds without such an oracle. In the arbitrary order model, we present algorithms for both triangle and four cycle estimation with fewer passes and the same space complexity as in previous algorithms, and we show several of these bounds are optimal. We analyze our algorithms under several noise models, showing that the algorithms perform well even when the oracle errs. Our methodology expands upon prior work on "classical" streaming algorithms, as previous multi-pass and random order streaming algorithms can be seen as special cases of our algorithms, where the first pass or random order was used to implement the heavy edge oracle. Lastly, our experiments demonstrate advantages of the proposed method compared to state-of-the-art streaming algorithms.

cs.DS↗

arXiv2021

Detection of Ongoing Mass Loss from HD 63433c, a Young Mini Neptune

We detect Lyman $α$ absorption from the escaping atmosphere of HD 63433c, a $R=2.67 R_\oplus$, $P=20.5$ d mini Neptune orbiting a young (440 Myr) solar analogue in the Ursa Major Moving Group. Using HST/STIS, we measure a transit depth of $11.1 \pm 1.5$% in the blue wing and $8 \pm 3$% in the red. This signal is unlikely to be due to stellar variability, but should be confirmed by an upcoming second visit with HST. We do not detect Lyman $α$ absorption from the inner planet, a smaller $R=2.15 R_\oplus$ mini Neptune on a 7.1 d orbit. We use Keck/NIRSPEC to place an upper limit of 0.5% on helium absorption for both planets. We measure the host star's X-ray spectrum and FUV flux with XMM-Newton, and model the outflow from both planets using a 3D hydrodynamic code. This model provides a reasonable match to the light curve in the blue wing of the Lyman $α$ line and the helium non-detection for planet c, although it does not explain the tentative red wing absorption or reproduce the excess absorption spectrum in detail. Its predictions of strong Lyman $α$ and helium absorption from b are ruled out by the observations. This model predicts a much shorter mass loss timescale for planet b, suggesting that b and c are fundamentally different: while the latter still retains its hydrogen/helium envelope, the former has likely lost its primordial atmosphere.

astro-ph.EP↗

arXiv2021

Escaping Helium from TOI 560.01, a Young Mini Neptune

We report helium absorption from the escaping atmosphere of TOI 560.01 (HD 73583b), a $R=2.8 R_\oplus$, $P=6.4$ d mini Neptune orbiting a young ($\sim$600 Myr) K dwarf. Using Keck/NIRSPEC, we detect a signal with an average depth of $0.68 \pm 0.08$% in the line core. The absorption signal repeats during a partial transit obtained a month later, but is marginally stronger and bluer, perhaps reflecting changes in the stellar wind environment. Ingress occurs on time, and egress occurs within 12 minutes of the white light egress, although absorption rises more gradually than it declines. This suggests that the outflow is slightly asymmetric and confined to regions close to the planet. The absorption signal also exhibits a slight 4 km/s redshift rather than the expected blueshift; this might be explained if the planet has a modest orbital eccentricity, although the radial velocity data disfavors such an explanation. We use XMM-Newton observations to reconstruct the high energy stellar spectrum and model the planet's outflow with 1D and 3D hydrodynamic simulations. We find that our models generally overpredict the measured magnitude of the absorption during transit, the size of the blueshift, or both. Increasing the metallicity to 100$\times$ solar suppresses the signal, but the dependence of the predicted signal strength on metallicity is non-monotonic. Decreasing the assumed stellar EUV flux by a factor of 3 likewise suppresses the signal substantially.

astro-ph.EP↗

arXiv2021

Deciphering hierarchical organization of topologically associated domains through change-point testing

Background: The nucleus of eukaryotic cells spatially packages chromosomes into a hierarchical and distinct segregation that plays critical roles in maintaining transcription regulation. High-throughput methods of chromosome conformation capture, such as Hi-C, have revealed topologically associating domains (TADs) that are defined by biased chromatin interactions within them. Results: Here, we introduce a novel method, HiCKey, to decipher hierarchical TAD structures in Hi-C data and compare them across samples. We first derive a generalized likelihood-ratio (GLR) test for detecting change-points in an interaction matrix that follows a negative binomial distribution or general mixture distribution. We then employ several optimal search strategies to decipher hierarchical TADs with p-values calculated by the GLR test. Large-scale validations of simulation data show that HiCKey has good precision in recalling known TADs and is robust against random collision noise of chromatin interactions. By applying HiCKey to Hi-C data of seven human cell lines, we identified multiple layers of TAD organization among them, but the vast majority had no more than four layers. In particular, we found that TAD boundaries are significantly enriched in active chromosomal regions compared to repressed regions, indicating finer hierarchical architectures in active regions for precise gene transcription regulation. Conclusions: HiCKey is optimized for processing large matrices constructed from high-resolution Hi-C experiments. The method and theoretical result of the GLR test provide a general framework for significance testing of similar experimental chromatin interaction data that may not fully follow negative binomial distributions but rather more general mixture distributions.

stat.AP↗

arXiv2021

Personalized Federated Learning with First Order Model Optimization

While federated learning traditionally aims to train a single global model across decentralized local datasets, one model may not always be ideal for all participating clients. Here we propose an alternative, where each client only federates with other relevant clients to obtain a stronger model per client-specific objectives. To achieve this personalization, rather than computing a single model average with constant weights for the entire federation as in traditional FL, we efficiently calculate optimal weighted model combinations for each client, based on figuring out how much a client can benefit from another's model. We do not assume knowledge of any underlying data distributions or client similarities, and allow each client to optimize for arbitrary target distributions of interest, enabling greater flexibility for personalization. We evaluate and characterize our method on a variety of federated settings, datasets, and degrees of local data heterogeneity. Our method outperforms existing alternatives, while also enabling new features for personalized FL such as transfer outside of local data distributions.

cs.LG↗

arXiv2021

No escaping helium from 55 Cnc e

We search for escaping helium from the hot super Earth 55 Cnc e by taking high-resolution spectra of the 1083 nm line during two transits using Keck/NIRSPEC. We detect no helium absorption down to a 90\% upper limit of 250 ppm in excess absorption or 0.27 milli-Angstrom in equivalent width. This corresponds to a mass loss rate of less than $\sim10^9$ g/s assuming a Parker wind model with a plausible exosphere temperature of 5000-6000 K, although the precise constraint is heavily dependent on model assumptions. We consider both hydrogen- and helium-dominated atmospheric compositions, and find similar bounds on the mass loss rate in both scenarios. Our hydrodynamical models indicate that if a lightweight atmosphere exists on 55 Cnc e, our observations would have easily detected it. Together with the non-detection of Lyman $α$ absorption by Ehrenreich et al 2012, our helium non-detection indicates that 55 Cnc e either never accreted a primordial atmosphere in the first place, or lost its primordial atmosphere shortly after the dissipation of the gas disk.

astro-ph.EP↗

arXiv2020

Using Computer Vision to Automate Hand Detection and Tracking of Surgeon Movements in Videos of Open Surgery

Open, or non-laparoscopic surgery, represents the vast majority of all operating room procedures, but few tools exist to objectively evaluate these techniques at scale. Current efforts involve human expert-based visual assessment. We leverage advances in computer vision to introduce an automated approach to video analysis of surgical execution. A state-of-the-art convolutional neural network architecture for object detection was used to detect operating hands in open surgery videos. Automated assessment was expanded by combining model predictions with a fast object tracker to enable surgeon-specific hand tracking. To train our model, we used publicly available videos of open surgery from YouTube and annotated these with spatial bounding boxes of operating hands. Our model's spatial detections of operating hands significantly outperforms the detections achieved using pre-existing hand-detection datasets, and allow for insights into intra-operative movement patterns and economy of motion.

cs.CV↗

arXiv2020

Soft Multicopter Control using Neural Dynamics Identification

Dynamic control of a soft-body robot to deliver complex behaviors with low-dimensional actuation inputs is challenging. In this paper, we present a computational approach to automatically generate versatile, underactuated control policies that drives soft-bodied machines with complicated structures and nonlinear dynamics. Our target application is focused on the autonomous control of a soft multicopter, featured by its elastic material components, non-conventional shapes, and asymmetric rotor layouts, to precisely deliver compliant deformation and agile locomotion. The central piece of our approach lies in a lightweight neural surrogate model to identify and predict the temporal evolution of a set of geometric variables characterizing an elastic soft body. This physics-based learning model is further integrated into a Linear Quadratic Regulator (LQR) control loop enhanced by a novel online fixed-point relinearization scheme to accommodate the dynamic body balance, allowing an aggressive reduction of the computational overhead caused by the conventional full-scale sensing-simulation-control workflow. We demonstrate the efficacy of our approach by generating controllers for a broad spectrum of customized soft multicopter designs and testing them in a high-fidelity physics simulation environment. The control algorithm enables the multicopters to perform a variety of tasks, including hovering, trajectory tracking, cruising and active deforming.

cs.RO↗

arXiv2020

Characterizing Policy Divergence for Personalized Meta-Reinforcement Learning

Despite ample motivation from costly exploration and limited trajectory data, rapidly adapting to new environments with few-shot reinforcement learning (RL) can remain a challenging task, especially with respect to personalized settings. Here, we consider the problem of recommending optimal policies to a set of multiple entities each with potentially different characteristics, such that individual entities may parameterize distinct environments with unique transition dynamics. Inspired by existing literature in meta-learning, we extend previous work by focusing on the notion that certain environments are more similar to each other than others in personalized settings, and propose a model-free meta-learning algorithm that prioritizes past experiences by relevance during gradient-based adaptation. Our algorithm involves characterizing past policy divergence through methods in inverse reinforcement learning, and we illustrate how such metrics are able to effectively distinguish past policy parameters by the environment they were deployed in, leading to more effective fast adaptation during test time. To study personalization more effectively we introduce a navigation testbed to specifically incorporate environment diversity across training episodes, and demonstrate that our approach outperforms meta-learning alternatives with respect to few-shot reinforcement learning in personalized settings.

cs.LG↗

arXiv2020

PLATON II: New Capabilities And A Comprehensive Retrieval on HD 189733b Transit and Eclipse Data

Recently, we introduced PLanetary Atmospheric Tool for Observer Noobs (PLATON), a Python package that calculates model transmission spectra for exoplanets and retrieves atmospheric characteristics based on observed spectra. We now expand its capabilities to include the ability to compute secondary eclipse depths. We have also added the option to calculate models using the correlated-$k$ method for radiative transfer, which improves accuracy without sacrificing speed. Additionally, we update the opacities in PLATON--many of which were generated using old or proprietary line lists--using the most recent and complete public line lists. These opacities are made available at R=1000 and R=10,000 over the 0.3--30 um range, and at R=375,000 in select near IR bands, making it possible to utilize PLATON for ground-based high resolution cross correlation studies. To demonstrate PLATON's new capabilities, we perform a retrieval on published HST and Spitzer transmission and emission spectra of the archetypal hot Jupiter HD 189733b. This is the first joint transit and secondary eclipse retrieval for this planet in the literature, as well as the most comprehensive set of both transit and eclipse data assembled for a retrieval to date. We find that these high signal-to-noise data are well-matched by atmosphere models with a C/O ratio of $0.66_{-0.09}^{+0.05}$ and a metallicity of $12_{-5}^{+8}$ times solar where the terminator is dominated by extended nanometer-sized haze particles at optical wavelengths. These are among the smallest uncertainties reported to date for an exoplanet, demonstrating both the power and the limitations of HST and Spitzer exoplanet observations.

astro-ph.EP↗

arXiv2019

A Hubble PanCET Study of HAT-P-11b: A Cloudy Neptune with a Low Atmospheric Metallicity

We present the first comprehensive look at the $0.35-5$ $μ$m transmission spectrum of the warm ($\sim 800$ K) Neptune HAT-P-11b derived from thirteen individual transits observed using the Hubble and Spitzer Space Telescopes. Along with the previously published molecular absorption feature in the $1.1-1.7$ $μ$m bandpass, we detect a distinct absorption feature at 1.15 $μ$m and a weak feature at 0.95 $μ$m, indicating the presence of water and/or methane with a combined significance of 4.4 $σ$. We find that this planet's nearly flat optical transmission spectrum and attenuated near-infrared molecular absorption features are best-matched by models incorporating a high-altitude cloud layer. Atmospheric retrievals using the combined $0.35-1.7$ $μ$m HST transmission spectrum yield strong constraints on atmospheric cloud-top pressure and metallicity, but we are unable to match the relatively shallow Spitzer transit depths without under-predicting the strength of the near-infrared molecular absorption bands. HAT-P-11b's HST transmission spectrum is well-matched by predictions from our microphysical cloud models. Both forward models and retrievals indicate that HAT-P-11b most likely has a relatively low atmospheric metallicity ($<4.6 \; Z_{\odot}$ and $<86 \; Z_{\odot}$ at the $2 σ$ and $3 σ$ levels respectively), in contrast to the expected trend based on the solar system planets. Our work also demonstrates that the wide wavelength coverage provided by the addition of the HST STIS data is critical for making these inferences.

astro-ph.EP↗

arXiv2019

Mass Loss from the Exoplanet WASP-12b Inferred from $\textit{Spitzer}$ Phase Curves

The exoplanet WASP-12b is the prototype for the emerging class of ultra-hot, Jupiter-mass exoplanets. Past models have predicted---and near ultra-violet observations have shown---that this planet is losing mass. We present an analysis of two sets of 3.6 $μ$m and 4.5 $μ$m $\textit{Spitzer}$ phase curve observations of the system which show clear evidence of infrared radiation from gas stripped from the planet, and the gas appears to be flowing directly toward or away from the host star. This accretion signature is only seen at 4.5 $μ$m, not at 3.6 $μ$m, which is indicative either of CO emission at the longer wavelength or blackbody emission from cool, $\lesssim$ 600 K gas. It is unclear why WASP-12b is the only ultra-hot Jupiter to exhibit this mass loss signature, but perhaps WASP-12b's orbit is decaying as some have claimed, while the orbits of other exoplanets may be more stable; alternatively, the high energy irradiation from WASP-12A may be stronger than the other host stars. We also find evidence for phase offset variability at the level of $6.4σ$ ($46.2^{\circ}$) at 3.6 $μ$m.

astro-ph.EP↗

arXiv2019

Patient-Specific Effects of Medication Using Latent Force Models with Gaussian Processes

Multi-output Gaussian processes (GPs) are a flexible Bayesian nonparametric framework that has proven useful in jointly modeling the physiological states of patients in medical time series data. However, capturing the short-term effects of drugs and therapeutic interventions on patient physiological state remains challenging. We propose a novel approach that models the effect of interventions as a hybrid Gaussian process composed of a GP capturing patient physiology convolved with a latent force model capturing effects of treatments on specific physiological features. This convolution of a multi-output GP with a GP including a causal time-marked kernel leads to a well-characterized model of the patients' physiological state responding to interventions. We show that our model leads to analytically tractable cross-covariance functions, allowing scalable inference. Our hierarchical model includes estimates of patient-specific effects but allows sharing of support across patients. Our approach achieves competitive predictive performance on challenging hospital data, where we recover patient-specific response to the administration of three common drugs: one antihypertensive drug and two anticoagulants.

stat.ML↗

arXiv2019

Modules for Experiments in Stellar Astrophysics (MESA): Pulsating Variable Stars, Rotation, Convective Boundaries, and Energy Conservation

We update the capabilities of the open-knowledge software instrument Modules for Experiments in Stellar Astrophysics (MESA). RSP is a new functionality in MESAstar that models the non-linear radial stellar pulsations that characterize RR Lyrae, Cepheids, and other classes of variable stars. We significantly enhance numerical energy conservation capabilities, including during mass changes. For example, this enables calculations through the He flash that conserve energy to better than 0.001 %. To improve the modeling of rotating stars in MESA, we introduce a new approach to modifying the pressure and temperature equations of stellar structure, and a formulation of the projection effects of gravity darkening. A new scheme for tracking convective boundaries yields reliable values of the convective-core mass, and allows the natural emergence of adiabatic semiconvection regions during both core hydrogen- and helium-burning phases. We quantify the parallel performance of MESA on current generation multicore architectures and demonstrate improvements in the computational efficiency of radiative levitation. We report updates to the equation of state and nuclear reaction physics modules. We briefly discuss the current treatment of fallback in core-collapse supernova models and the thermodynamic evolution of supernova explosions. We close by discussing the new MESA Testhub software infrastructure to enhance source-code development.

astro-ph.SR↗