arXiv subjects

Bruce A. Bassett

Publications and source records attributed to Bruce A. Bassett.

At least 37 records · Page 2Linked to original sources

arXiv2021

Low-Resource Neural Machine Translation for Southern African Languages

Low-resource African languages have not fully benefited from the progress in neural machine translation because of a lack of data. Motivated by this challenge we compare zero-shot learning, transfer learning and multilingual learning on three Bantu languages (Shona, isiXhosa and isiZulu) and English. Our main target is English-to-isiZulu translation for which we have just 30,000 sentence pairs, 28% of the average size of our other corpora. We show the importance of language similarity on the performance of English-to-isiZulu transfer learning based on English-to-isiXhosa and English-to-Shona parent models whose BLEU scores differ by 5.2. We then demonstrate that multilingual learning surpasses both transfer learning and zero-shot learning on our dataset, with BLEU score improvements relative to the baseline English-to-isiZulu model of 9.9, 6.1 and 2.0 respectively. Our best model also improves the previous SOTA BLEU score by more than 10.

cs.CL↗

arXiv2021

Classification of Multiwavelength Transients with Machine Learning

With the advent of powerful telescopes such as the Square Kilometer Array and the Vera C. Rubin Observatory, we are entering an era of multiwavelength transient astronomy that will lead to a dramatic increase in data volume. Machine learning techniques are well suited to address this data challenge and rapidly classify newly detected transients. We present a multiwavelength classification algorithm consisting of three steps: (1) interpolation and augmentation of the data using Gaussian processes; (2) feature extraction using wavelets; and (3) classification with random forests. Augmentation provides improved performance at test time by balancing the classes and adding diversity into the training set. In the first application of machine learning to the classification of real radio transient data, we apply our technique to the Green Bank Interferometer and other radio light curves. We find we are able to accurately classify most of the 11 classes of radio variables and transients after just eight hours of observations, achieving an overall test accuracy of 78 percent. We fully investigate the impact of the small sample size of 82 publicly available light curves and use data augmentation techniques to mitigate the effect. We also show that on a significantly larger simulated representative training set that the algorithm achieves an overall accuracy of 97 percent, illustrating that the method is likely to provide excellent performance on future surveys. Finally, we demonstrate the effectiveness of simultaneous multiwavelength observations by showing how incorporating just one optical data point into the analysis improves the accuracy of the worst performing class by 19 percent.

astro-ph.IM↗

arXiv2020

Deep Evolution for Facial Emotion Recognition

Deep facial expression recognition faces two challenges that both stem from the large number of trainable parameters: long training times and a lack of interpretability. We propose a novel method based on evolutionary algorithms, that deals with both challenges by massively reducing the number of trainable parameters, whilst simultaneously retaining classification performance, and in some cases achieving superior performance. We are robustly able to reduce the number of parameters on average by 95% (e.g. from 2M to 100k parameters) with no loss in classification accuracy. The algorithm learns to choose small patches from the image, relative to the nose, which carry the most important information about emotion, and which coincide with typical human choices of important features. Our work implements a novel form attention and shows that evolutionary algorithms are a valuable addition to machine learning in the deep learning era, both for reducing the number of parameters for facial expression recognition and for providing interpretable features that can help reduce bias.

cs.NE↗

arXiv2020

Deep Learning improves identification of Radio Frequency Interference

Flagging of Radio Frequency Interference (RFI) is an increasingly important challenge in radio astronomy. We present R-Net, a deep convolutional ResNet architecture that significantly outperforms existing algorithms -- including the default MeerKAT RFI flagger, and deep U-Net architectures -- across all metrics including AUC, F1-score and MCC. We demonstrate the robustness of this improvement on both single dish and interferometric simulations and, using transfer learning, on real data. Our R-Net model's precision is approximately $90\%$ better than the current MeerKAT flagger at $80\%$ recall and has a 35\% higher F1-score with no additional performance cost. We further highlight the effectiveness of transfer learning from a model initially trained on simulated MeerKAT data and fine-tuned on real, human-flagged, KAT-7 data. Despite the wide differences in the nature of the two telescope arrays, the model achieves an AUC of 0.91, while the best model without transfer learning only reaches an AUC of 0.67. We consider the use of phase information in our models but find that without calibration the phase adds almost no extra information relative to amplitude data only. Our results strongly suggest that deep learning on simulations, boosted by transfer learning on real data, will likely play a key role in the future of RFI flagging of radio astronomy data.

astro-ph.IM↗

arXiv2020

Multidimensional RFI Framework for Characterising Radio Astronomy Observatories

Radio Frequency Interference (RFI) has historically plagued radio astronomy, worsening with the rapid spread of electronics and increasing telescope sensitivity. We present a multi-dimensional probabilistic framework for characterising the RFI environment around a radio astronomy site that uses automatically flagged data from the array itself. We illustrate the framework using about 1500 hours of commissioning data from the MeerKAT radio telescope; producing a 6-dimensional array that yields both average RFI occupancy as well as confidence intervals around the mean as a function of key variables (frequency, direction, baseline, time). Our results provide the first detailed view of the MeerKAT RFI environment at high sensitivity as a function of direction, frequency, time of day and baseline. They allow us to track the historical evolution of the RFI and to quantify fluctuations which can be used for alerting on new RFI. As expected we find the major RFI contributors for MeerKAT site are from Global Positioning System (GPS) satellites, flight Distance Measurement Equipment (DME) and the Global System for Mobile (GSM) Communications. Beyond characterising RFI environments our approach allows observers access to the prior probability of RFI in any combination of tracked variables, allowing for more efficient observation planning and data excision.

astro-ph.IM↗

arXiv2020

Climate & BCG: Effects on COVID-19 Death Growth Rates

Multiple studies have suggested the spread of COVID-19 is affected by factors such as climate, BCG vaccinations, pollution and blood type. We perform a joint study of these factors using the death growth rates of 40 regions worldwide with both machine learning and Bayesian methods. We find weak, non-significant (< 3$σ$) evidence for temperature and relative humidity as factors in the spread of COVID-19 but little or no evidence for BCG vaccination prevalence or $\text{PM}_{2.5}$ pollution. The only variable detected at a statistically significant level (>3$σ$) is the rate of positive COVID-19 tests, with higher positive rates correlating with higher daily growth of deaths.

q-bio.PE↗

arXiv2019

A Flexible Framework for Anomaly Detection via Dimensionality Reduction

Anomaly detection is challenging, especially for large datasets in high dimensions. Here we explore a general anomaly detection framework based on dimensionality reduction and unsupervised clustering. We release DRAMA, a general python package that implements the general framework with a wide range of built-in options. We test DRAMA on a wide variety of simulated and real datasets, in up to 3000 dimensions, and find it robust and highly competitive with commonly-used anomaly detection algorithms, especially in high dimensions. The flexibility of the DRAMA framework allows for significant optimization once some examples of anomalies are available, making it ideal for online anomaly detection, active learning and highly unbalanced datasets.

cs.LG↗

arXiv2019

Bayesian Anomaly Detection and Classification

Statistical uncertainties are rarely incorporated in machine learning algorithms, especially for anomaly detection. Here we present the Bayesian Anomaly Detection And Classification (BADAC) formalism, which provides a unified statistical approach to classification and anomaly detection within a hierarchical Bayesian framework. BADAC deals with uncertainties by marginalising over the unknown, true, value of the data. Using simulated data with Gaussian noise, BADAC is shown to be superior to standard algorithms in both classification and anomaly detection performance in the presence of uncertainties, though with significantly increased computational cost. Additionally, BADAC provides well-calibrated classification probabilities, valuable for use in scientific pipelines. We show that BADAC can work in online mode and is fairly robust to model errors, which can be diagnosed through model-selection methods. In addition it can perform unsupervised new class detection and can naturally be extended to search for anomalous subsets of data. BADAC is therefore ideal where computational cost is not a limiting factor and statistical rigour is important. We discuss approximations to speed up BADAC, such as the use of Gaussian processes, and finally introduce a new metric, the Rank-Weighted Score (RWS), that is particularly suited to evaluating the ability of algorithms to detect anomalies.

stat.ML↗

arXiv2018

DeepSource: Point Source Detection using Deep Learning

Point source detection at low signal-to-noise is challenging for astronomical surveys, particularly in radio interferometry images where the noise is correlated. Machine learning is a promising solution, allowing the development of algorithms tailored to specific telescope arrays and science cases. We present DeepSource - a deep learning solution - that uses convolutional neural networks to achieve these goals. DeepSource enhances the Signal-to-Noise Ratio (SNR) of the original map and then uses dynamic blob detection to detect sources. Trained and tested on two sets of 500 simulated 1 deg x 1 deg MeerKAT images with a total of 300,000 sources, DeepSource is essentially perfect in both purity and completeness down to SNR = 4 and outperforms PyBDSF in all metrics. For uniformly-weighted images it achieves a Purity x Completeness (PC) score at SNR = 3 of 0.73, compared to 0.31 for the best PyBDSF model. For natural-weighting we find a smaller improvement of ~40% in the PC score at SNR = 3. If instead we ask where either of the purity or completeness first drop to 90%, we find that DeepSource reaches this value at SNR = 3.6 compared to the 4.3 of PyBDSF (natural-weighting). A key advantage of DeepSource is that it can learn to optimally trade off purity and completeness for any science case under consideration. Our results show that deep learning is a promising approach to point source detection in astronomical images.

astro-ph.IM↗

arXiv2018

Painting galaxies into dark matter halos using machine learning

We develop a machine learning (ML) framework to populate large dark matter-only simulations with baryonic galaxies. Our ML framework takes input halo properties including halo mass, environment, spin, and recent growth history, and outputs central galaxy and halo baryonic properties including stellar mass ($M_*$), star formation rate (SFR), metallicity ($Z$), neutral ($\rm HI$) and molecular ($\rm H_2$) hydrogen mass. We apply this to the MUFASA cosmological hydrodynamic simulation, and show that it recovers the mean trends of output quantities with halo mass highly accurately, including following the sharp drop in SFR and gas in quenched massive galaxies. However, the scatter around the mean relations is under-predicted. Examining galaxies individually, at $z=0$ the stellar mass and metallicity are accurately recovered ($σ\lesssim 0.2$~dex), but SFR and $\rm HI$ show larger scatter ($σ\gtrsim 0.3$~dex); these values improve somewhat at $z=1,2$. Remarkably, ML quantitatively recovers second parameter trends in galaxy properties, e.g. that galaxies with higher gas content and lower metallicity have higher SFR at a given $M_*$. Testing various ML algorithms, we find that none perform significantly better than the others, nor does ensembling improve performance, likely because none of the algorithms reproduce the large observed scatter around the mean properties. For the random forest algorithm, we find that halo mass and nearby ($\sim 200$~kpc) environment are the most important predictive variables followed by growth history, while halo spin and $\sim$Mpc scale environment are not important. Finally we study the impact of additionally inputting key baryonic properties $M_*$, SFR and $Z$, as would be available e.g. from an equilibrium model, and show that particularly providing the SFR enables $\rm HI$ to be recovered substantially more accurately.

astro-ph.GA↗

arXiv2018

Automated Classification of Text Sentiment

The ability to identify sentiment in text, referred to as sentiment analysis, is one which is natural to adult humans. This task is, however, not one which a computer can perform by default. Identifying sentiments in an automated, algorithmic manner will be a useful capability for business and research in their search to understand what consumers think about their products or services and to understand human sociology. Here we propose two new Genetic Algorithms (GAs) for the task of automated text sentiment analysis. The GAs learn whether words occurring in a text corpus are either sentiment or amplifier words, and their corresponding magnitude. Sentiment words, such as 'horrible', add linearly to the final sentiment. Amplifier words in contrast, which are typically adjectives/adverbs like 'very', multiply the sentiment of the following word. This increases, decreases or negates the sentiment of the following word. The sentiment of the full text is then the sum of these terms. This approach grows both a sentiment and amplifier dictionary which can be reused for other purposes and fed into other machine learning algorithms. We report the results of multiple experiments conducted on large Amazon data sets. The results reveal that our proposed approach was able to outperform several public and/or commercial sentiment analysis algorithms.

cs.CL↗

arXiv2017

zBEAMS: A unified solution for supernova cosmology with redshift uncertainties

Supernova cosmology without spectra will be an important component of future surveys such as LSST. This lack of supernova spectra results in uncertainty in the redshifts which, if ignored, leads to significantly biased estimates of cosmological parameters. Here we present a hierarchical Bayesian formalism -- zBEAMS -- that addresses this problem by marginalising over the unknown or uncertain supernova redshifts to produce unbiased cosmological estimates that are competitive with supernova data with spectroscopically confirmed redshifts. zBEAMS provides a unified treatment of both photometric redshifts and host galaxy misidentification (occurring due to chance galaxy alignments or faint hosts), effectively correcting the inevitable contamination in the Hubble diagram. Like its predecessor BEAMS, our formalism also takes care of non-Ia supernova contamination by marginalising over the unknown supernova type. We illustrate this technique with simulations of supernovae with photometric redshifts and host galaxy misidentification. A novel feature of the photometric redshift case is the important role played by the redshift distribution of the supernovae.

astro-ph.CO↗

arXiv2017

EDEN: Evolutionary Deep Networks for Efficient Machine Learning

Deep neural networks continue to show improved performance with increasing depth, an encouraging trend that implies an explosion in the possible permutations of network architectures and hyperparameters for which there is little intuitive guidance. To address this increasing complexity, we propose Evolutionary DEep Networks (EDEN), a computationally efficient neuro-evolutionary algorithm which interfaces to any deep neural network platform, such as TensorFlow. We show that EDEN evolves simple yet successful architectures built from embedding, 1D and 2D convolutional, max pooling and fully connected layers along with their hyperparameters. Evaluation of EDEN across seven image and sentiment classification datasets shows that it reliably finds good networks -- and in three cases achieves state-of-the-art results -- even on a single GPU, in just 6-24 hours. Our study provides a first attempt at applying neuro-evolution to the creation of 1D convolutional networks for sentiment analysis including the optimisation of the embedding layer.

stat.ML↗

arXiv2017

Text Compression for Sentiment Analysis via Evolutionary Algorithms

Can textual data be compressed intelligently without losing accuracy in evaluating sentiment? In this study, we propose a novel evolutionary compression algorithm, PARSEC (PARts-of-Speech for sEntiment Compression), which makes use of Parts-of-Speech tags to compress text in a way that sacrifices minimal classification accuracy when used in conjunction with sentiment analysis algorithms. An analysis of PARSEC with eight commercial and non-commercial sentiment analysis algorithms on twelve English sentiment data sets reveals that accurate compression is possible with (0%, 1.3%, 3.3%) loss in sentiment classification accuracy for (20%, 50%, 75%) data compression with PARSEC using LingPipe, the most accurate of the sentiment algorithms. Other sentiment analysis algorithms are more severely affected by compression. We conclude that significant compression of text data is possible for sentiment analysis depending on the accuracy demands of the specific application and the specific sentiment analysis algorithm used.

cs.NE↗

arXiv2017

MeerKLASS: MeerKAT Large Area Synoptic Survey

We discuss the ground-breaking science that will be possible with a wide area survey, using the MeerKAT telescope, known as MeerKLASS (MeerKAT Large Area Synoptic Survey). The current specifications of MeerKAT make it a great fit for science applications that require large survey speeds but not necessarily high angular resolutions. In particular, for cosmology, a large survey over $\sim 4,000 \, {\rm deg}^2$ for $\sim 4,000$ hours will potentially provide the first ever measurements of the baryon acoustic oscillations using the 21cm intensity mapping technique, with enough accuracy to impose constraints on the nature of dark energy. The combination with multi-wavelength data will give unique additional information, such as exquisite constraints on primordial non-Gaussianity using the multi-tracer technique, as well as a better handle on foregrounds and systematics. Such a wide survey with MeerKAT is also a great match for HI galaxy studies, providing unrivalled statistics in the pre-SKA era for galaxies resolved in the HI emission line beyond local structures at z > 0.01. It will also produce a large continuum galaxy sample down to a depth of about 5\,$μ$Jy in L-band, which is quite unique over such large areas and will allow studies of the large-scale structure of the Universe out to high redshifts, complementing the galaxy HI survey to form a transformational multi-wavelength approach to study galaxy dynamics and evolution. Finally, the same survey will supply unique information for a range of other science applications, including a large statistical investigation of galaxy clusters as well as produce a rotation measure map across a huge swathe of the sky. The MeerKLASS survey will be a crucial step on the road to using SKA1-MID for cosmological applications and other commensal surveys, as described in the top priority SKA key science projects (abridged).

astro-ph.CO↗

arXiv2017

Automated Problem Identification: Regression vs Classification via Evolutionary Deep Networks

Regression or classification? This is perhaps the most basic question faced when tackling a new supervised learning problem. We present an Evolutionary Deep Learning (EDL) algorithm that automatically solves this by identifying the question type with high accuracy, along with a proposed deep architecture. Typically, a significant amount of human insight and preparation is required prior to executing machine learning algorithms. For example, when creating deep neural networks, the number of parameters must be selected in advance and furthermore, a lot of these choices are made based upon pre-existing knowledge of the data such as the use of a categorical cross entropy loss function. Humans are able to study a dataset and decide whether it represents a classification or a regression problem, and consequently make decisions which will be applied to the execution of the neural network. We propose the Automated Problem Identification (API) algorithm, which uses an evolutionary algorithm interface to TensorFlow to manipulate a deep neural network to decide if a dataset represents a classification or a regression problem. We test API on 16 different classification, regression and sentiment analysis datasets with up to 10,000 features and up to 17,000 unique target values. API achieves an average accuracy of $96.3\%$ in identifying the problem type without hardcoding any insights about the general characteristics of regression or classification problems. For example, API successfully identifies classification problems even with 1000 target values. Furthermore, the algorithm recommends which loss function to use and also recommends a neural network architecture. Our work is therefore a step towards fully automated machine learning.

cs.NE↗

arXiv2016

Application of Bayesian graphs to SN Ia data analysis and compression

Bayesian graphical models are an efficient tool for modelling complex data and derive self-consistent expressions of the posterior distribution of model parameters. We apply Bayesian graphs to perform statistical analyses of Type Ia supernova (SN Ia) luminosity distance measurements from the joint light-curve analysis (JLA) data set. In contrast to the $χ^2$ approach used in previous studies, the Bayesian inference allows us to fully account for the standard-candle parameter dependence of the data covariance matrix. Comparing with $χ^2$ analysis results, we find a systematic offset of the marginal model parameter bounds. We demonstrate that the bias is statistically significant in the case of the SN Ia standardization parameters with a maximal 6 $σ$ shift of the SN light-curve colour correction. In addition, we find that the evidence for a host galaxy correction is now only 2.4 $σ$. Systematic offsets on the cosmological parameters remain small, but may increase by combining constraints from complementary cosmological probes. The bias of the $χ^2$ analysis is due to neglecting the parameter-dependent log-determinant of the data covariance, which gives more statistical weight to larger values of the standardization parameters. We find a similar effect on compressed distance modulus data. To this end, we implement a fully consistent compression method of the JLA data set that uses a Gaussian approximation of the posterior distribution for fast generation of compressed data. Overall, the results of our analysis emphasize the need for a fully consistent Bayesian statistical approach in the analysis of future large SN Ia data sets.

astro-ph.CO↗

arXiv2016

Bayes Factors via Savage-Dickey Supermodels

We outline a new method to compute the Bayes Factor for model selection which bypasses the Bayesian Evidence. Our method combines multiple models into a single, nested, Supermodel using one or more hyperparameters. Since the models are now nested the Bayes Factors between the models can be efficiently computed using the Savage-Dickey Density Ratio (SDDR). In this way model selection becomes a problem of parameter estimation. We consider two ways of constructing the supermodel in detail: one based on combined models, and a second based on combined likelihoods. We report on these two approaches for a Gaussian linear model for which the Bayesian evidence can be calculated analytically and a toy nonlinear problem. Unlike the combined model approach, where a standard Monte Carlo Markov Chain (MCMC) struggles, the combined-likelihood approach fares much better in providing a reliable estimate of the log-Bayes Factor. This scheme potentially opens the way to computationally efficient ways to compute Bayes Factors in high dimensions that exploit the good scaling properties of MCMC, as compared to methods such as nested sampling that fail for high dimensions.

astro-ph.IM↗