Search

In recent years there has been increasing interest in the field of Open-Set Recognition, which allows a classification model to identify inputs as "unknown" when it encounters an object or class not in the training set. This ability to flag unknown inputs is of vital importance to many real world classification applications. As almost all modern training methods for neural networks use extensive amounts of regularization for generalization, it is therefore important to examine how regularization techniques impact the ability of a model to perform Open-Set Recognition. In this work, we examine the relationship between common regularization techniques and Open-Set Recognition performance. Our experiments are agnostic to the specific open-set detection algorithm and examine the effects across a wide range of datasets. We show empirically that regularization methods can provide significant improvements to Open-Set Recognition performance, and we provide new insights into the relationship between accuracy and Open-Set performance.

Overfitting describes a machine learning phenomenon where the model fits too closely to the training data, resulting in poor generalization. While this occurrence is thoroughly documented for many forms of supervised learning, it is not well examined in the context of unsupervised learning. In this work we examine the nature of overfitting in unsupervised contrastive learning. We show that overfitting can indeed occur and the mechanism behind overfitting.

Semi-supervised learning approaches train on small sets of labeled data along with large sets of unlabeled data. Self-training is a semi-supervised teacher-student approach that often suffers from the problem of "confirmation bias" that occurs when the student model repeatedly overfits to incorrect pseudo-labels given by the teacher model for the unlabeled data. This bias impedes improvements in pseudo-label accuracy across self-training iterations, leading to unwanted saturation in model performance after just a few iterations. In this work, we describe multiple enhancements to improve the self-training pipeline to mitigate the effect of confirmation bias. We evaluate our enhancements over multiple datasets showing performance gains over existing self-training design choices. Finally, we also study the extendability of our enhanced approach to Open Set unlabeled data (containing classes not seen in labeled data).

We present a model for the formation of the Magellanic Stream (MS) due to ram pressure stripping. We model the history of the Small and Large Magellanic Clouds in the recent cosmological past in a static Milky Way potential with diffuse halo gas, using observationally motivated orbits for the Magellanic Clouds derived from HST proper motions within the potential of the Milky Way. This model is able to reproduce the trailing arm but does not reproduce the leading arm feature, which is common for models of the stream formation that include ram pressure stripping effects. While our model does not outperform other models in terms of matching the observable quantities in the MS, it is close enough for our ultimate goal -- using the MS to estimate the MW mass. From analyzing our grid of models, we find that there is a direct correlation between the observed stream length in our simulations and the mass of the Milky Way. For the observed MS length, the inferred Milky Way mass is $1.5 \pm 0.32 \times 10^{12}$ $M_\odot$, which agrees closely with other independent measures of the Milky Way mass. We also discuss the MS in the context of HI streams in galaxy clusters, and find that the MS lies on the low-mass end of a continuum from Hickson groups to the Virgo cluster. As a tracer of the dynamical mass in the outer halo, the MS is a particularly valuable probe of the Milky Way's potential., Comment: Accepted for publication in MNRAS. 15 pages, 14 figures

Modern network-on-chip (NoC) systems face reliability issues due to process and environmental variations. The power supply noise (PSN) in the power delivery network of a NoC plays a key role in determining reliability. PSN leads to voltage droop, which can cause timing errors in the NoC. This paper makes a novel contribution towards formally analyzing PSN in NoC systems. We present a probabilistic model checking approach to observe the PSN in a generic 2x2 mesh NoC with a uniform random traffic load. Key features of PSN are measured at the behavioral level. To tackle state explosion, we apply incremental abstraction techniques, including a novel probabilistic choice abstraction, based on observations of NoC behavior. The Modest Toolset is used for probabilistic modeling and verification. Results are obtained for several flit injection patterns to reveal their impacts on PSN. Our analysis finds an optimal flit pattern generation with zero probability of PSN events and suggests spreading flits rather than releasing them in consecutive cycles in order to minimize PSN., Comment: 17 pages, 7 figures, submitted to FMICS 2021

Non-ideal magnetohydrodynamics (MHD) is the dominant process. We investigate the effect of magnetic fields (ideal and non-ideal) and turbulence (sub- and transsonic) on the formation of circumstellar discs that form nearly simultaneously with the formation of the protostar. This is done by modelling the gravitational collapse of a 1~M$_\odot$ gas cloud that is threaded with a magnetic field and imposed with both rotational and turbulent velocities. We investigate magnetic fields that are parallel/anti-parallel and perpendicular to the rotation axis, two rotation rates and four Mach numbers. Disc formation occurs preferentially in the models that include non-ideal MHD where the magnetic field is anti-parallel or perpendicular to the rotation axis. This is independent of the initial rotation rate and level of turbulence, suggesting that subsonic turbulence plays a minimal role in influencing the formation of discs. Aside from first core outflows which are influenced by the initial level of turbulence, non-ideal MHD processes are more important than turbulent processes during the formation of discs around low-mass stars., Comment: Accepted for publication in MNRAS. 12 pages, 8 figures

Non-ideal magnetohydrodynamics (MHD) is the dominant process. We investigate the effect of magnetic fields (ideal and non-ideal) and turbulence (sub- and transsonic) on the formation of protostars by following the gravitational collapse of 1~M$_\odot$ gas clouds through the first hydrostatic core to stellar densities. The clouds are imposed with both rotational and turbulent velocities, and are threaded with a magnetic field that is parallel/anti-parallel or perpendicular to the rotation axis; we investigate two rotation rates and four Mach numbers. The initial radius and mass of the stellar core are only weakly dependent on the initial parameters. In the models that include ideal MHD, the magnetic field strength implanted in the protostar at birth is much higher than observed, independent of the initial level of turbulence; only non-ideal MHD can reduce this strength to near or below the observed levels. This suggests that not only is ideal MHD an incomplete picture of star formation, but that the magnetic fields in low mass stars are implanted later in life by a dynamo process. Non-ideal MHD suppresses magnetically launched stellar core outflows, but turbulence permits thermally launched outflows to form a few years after stellar core formation., Comment: Accepted for publication in MNRAS. 12 pages, 9 figures

Many spiral galaxies have extended outer H~I disks and display low levels of star formation, inferred from the far-ultraviolet emission detected by {\it GALEX}, well beyond the optical radius. Here, we investigate the supernova (SN) rate in the outskirts of galaxies, using the largest and most homogeneous set of nearby supernovae (SNe) from the Lick Observatory Supernova Search (LOSS). While SN rates have been measured with respect to various galaxy properties, such as stellar mass and metallicity, their relative frequency in the outskirts versus the inner regions has not yet been studied. Understanding the SN rate as a function of intragalactic environment has many ramifications, including the interpretation of LIGO observations, the formation of massive stars, and the puzzlingly high velocity dispersion of the outer H~I disk. Using data from the LOSS survey, we find that the rate beyond the optical radius of spiral galaxies is $2.5 \pm 0.5$ SNe per millennium, while dwarf galaxies host $ 4.0 \pm 2.2$ SNe per millennium. The rates of core-collapse SNe (that may collapse to form the massive black holes detected by LIGO/Virgo) in the outer disks of spirals is $1.5 \pm 0.15$ SNe per millennium and in dwarf galaxies is $2.6 \pm 1.5$ SNe per millennium. Core-collapse SNe in spiral outskirts contribute $7600 \pm 1700$\,SNe\,~Gpc$^{-3}$\,yr$^{-1}$ to the volumetric rate, and dwarf galaxies have a rate of $31,000 \pm 18,000$\,SNe\,Gpc$^{-3}$\,yr$^{-1}$. The relative ratio of core-collapse to Type Ia SNe is comparable in the inner and outer parts of spirals, and in dwarf galaxies., Comment: accepted to Astrophysical Journal Letters

We present the results of eighteen magnetohydrodynamical calculations of the collapse of a molecular cloud core to form a protostar. Some calculations include radiative transfer in the flux-limited diffusion approximation while others employ a barotropic equation of state. We cover a wide parameter space, with mass-to-flux ratios ranging from $\mu = 5$ to $20$; initial turbulent amplitudes ranging from a laminar calculation (i.e. where the Mach number, $\mathscr{M} = 0$) to transonic $\mathscr{M} = 1$; and initial rotation rates from $\beta_\mathrm{rot} = 0.005$ to $0.02$. We first show that using a radiative transfer scheme produces warmer pseudo-discs than the barotropic equation of state, making them more stable. We then `shake' the core by increasing the initial turbulent velocity field, and find that at all three mass-to-flux ratios transonic cores are weakly bound and do not produce pseudo-discs; $\mathscr{M} = 0.3$ cores produce very disrupted discs; and $\mathscr{M} = 0.1$ cores produce discs broadly comparable to a laminar core. In our previous paper (arXiv:1701.08741), we showed that a pseudo-disc coupled with sufficient magnetic field is necessary to form a bipolar outflow. Here we show that only weakly turbulent cores exhibit collimated jets. We finally take the $\mathscr{M} = 1.0$, $\mu = 5$ core and `stir' it by increasing the initial angular momentum, finding that once the degree of rotational energy exceeds the turbulent energy in the core the disc returns, with a corresponding (though slower), outflow. These conclusions place constraints on the initial mixtures of rotation and turbulence in molecular cloud cores which are conducive to the formation of bipolar outflows early in the star formation process., Comment: Accepted for publication in MNRAS. 16 pages, 13 figures

Let $K$ be a simply connected compact Lie group and $T^{\ast}(K)$ its cotangent bundle. We consider the problem of "quantization commutes with reduction" for the adjoint action of $K$ on $T^{\ast}(K).$ We quantize both $T^{\ast}(K)$ and the reduced phase space using geometric quantization with half-forms. We then construct a geometrically natural map from the space of invariant elements in the quantization of $T^{\ast}(K)$ to the quantization of the reduced phase space. We show that this map is a constant multiple of a unitary map., Comment: Final version. Minor typographical corrections. To appear in Quarterly Journal of Mathematics

We report results from twelve simulations of the collapse of a molecular cloud core to form one or more protostars, comprising three field strengths (mass-to-flux ratios, {\mu}, of 5, 10, and 20) and four field geometries (with values of the angle between the field and rotation axes, {\theta}, of 0{\deg}, 20{\deg}, 45{\deg}, and 90{\deg}), using a smoothed particle magnetohydrodynamics method. We find that the values of both parameters have a strong effect on the resultant protostellar system and outflows. This ranges from the formation of binary systems when {\mu} = 20 to strikingly differing outflow structures for differing values of {\theta}, in particular highly suppressed outflows when {\theta} = 90{\deg}. Misaligned magnetic fields can also produce warped pseudo-discs where the outer regions align perpendicular to the magnetic field but the innermost region re-orientates to be perpendicular to the rotation axis. We follow the collapse to sizes comparable to those of first cores and find that none of the outflow speeds exceed 8 km s$^{-1}$. These results may place constraints on both observed protostellar outflows, and also on which molecular cloud cores may eventually form either single stars and binaries: a sufficiently weak magnetic field may allow for disc fragmentation, whilst conversely the greater angular momentum transport of a strong field may inhibit disc fragmentation., Comment: Accepted for publication in MNRAS. 14 pages, 14 figures. Animations are other details can be found at http://www.astro.ex.ac.uk/people/blewis/research/outflows_weak_fields.html

Obtaining a stable magnetohydrodynamical (MHD) formalism in SPH - i.e. smoothed particle magnetohydrodynamics (SPMHD) - has proven remarkably difficult. To implement MHD requires two steps: a modification to the momentum equation and an induction equation, and both present challenges. We first provide an overview of how SPMHD is implemented, and then discuss how this implementation fails and the limitation of various corrective methods - with particular reference to the effects of particle disorder. Although there are many problems for which, with careful choice of corrective measures, good results can be obtained, we then show that, at the very limits of the state of the art, the ability to perform stable MHD calculations in SPH is curtailed by numerical issues., Comment: 8 pages, 6 figures. Proceedings of the "11th international SPHERIC workshop", TUM, Munich, Germany, 13-16 June 2016

The equations of smoothed particle magnetohydrodynamics (SPMHD), even with the various corrections to instabilities so far proposed, have been observed to be unstable when a very steep density gradient is necessarily combined with a variable smoothing length formalism. Here we consider in more detail the modifications made to the SPMHD equations in LBP2015 that resolve this instability by replacing the smoothing length in the induction and anisotropic force equations with an average smoothing length term. We then explore the choice of average used and compare the effects on a test `cylinder-in-a-box' problem and the collapse of a magnetised molecular cloud core. We find that, aside from some benign numerical effects at low resolutions for the quadratic mean, the formalism is robust as to the choice of average but that in complicated models it is essential to apply the average to both equations; in particular, all four averages considered exhibit similar conservation properties. This improved formalism allows for arbitrarily small sink particles and field geometries to be explored, vastly expanding the range of astronomical problems that can be modeled using SPMHD., Comment: 7 pages, 4 figures. Proceedings of the "10th international SPHERIC workshop", Parma, Italy, 16-18 June 2015

We have developed a modified form of the equations of smoothed particle magnetohydrodynamics which are stable in the presence of very steep density gradients. Using this formalism, we have performed simulations of the collapse of magnetised molecular cloud cores to form protostars and drive outflows. Our stable formalism allows for smaller sink particles (< 5 AU) than used previously and the investigation of the effect of varying the angle, {\theta}, between the initial field axis and the rotation axis. The nature of the outflows depends strongly on this angle: jet-like outflows are not produced at all when {\theta} > 30{\deg}, and a collimated outflow is not sustained when {\theta} > 10{\deg}. No substantial outflows of any kind are produced when {\theta} > 60{\deg}. This may place constraints on the geometry of the magnetic field in molecular clouds where bipolar outflows are seen., Comment: Accepted for publication in MNRAS, 13 pages, 14 figures. Animations can be found at http://www.astro.ex.ac.uk/people/blewis/research/outflows_misaligned_fields.html