Electrical Engineering Systems Seminar

Speaker: Almog Hershko

M.Sc. student under the supervision of Prof. Dovi Poznanski

Wednesday, 31^st July 2024, at 15:30

Room 011, Kitot Building, Faculty of Engineering

Noise Agnostic Outlier Detection on Galaxy Spectra

Abstract

The field of astronomy, like many other scientific fields, is deep inside the age of big data. Thanks to technological advancements in sensors and computers, numerous large astronomical datasets already exist containing billions of observations. One such dataset is the Sloan Digital Sky Survey (SDSS), an ongoing sky survey of more than 20 years that includes (among other data) several million galaxy spectra.

Naturally, data-driven algorithms play a key role in helping scientists extract new insights from these datasets. Modern learning algorithms can handle vast amounts of data, extract trends and direct scientists to uncover the underlying physics driving them.

Alternatively, a learning algorithm can detect outliers that stand out from the rest of the data due to some unique or rare physical phenomena. Outlier detectors are unsupervised learning algorithms, that require only data and no labels and produce a ranking of the data according to some learned outlier score. Such outlier detectors can point the attention of researchers to unique objects that potentially hold key to new insights.

A unique feature of astronomical datasets is the fact that the vast majority of the data is noisy. First, because the sources are intrinsically faint, and we cannot change that. Second, because most of the volume of the Universe is far from earth, and objects grow fainter the further away from us they are. Consequently, a survey down to some sensitivity limit will usually detect most of its sources near the largest distance it can reach. This means that machine learning tools for astronomical datasets should be more robust to noise than in other domains.

This thesis builds upon an existing outlier detector for galaxy spectra that is based on unsupervised random forest (URF). URF trained on SDSS spectra has been shown previously to produce meaningful outliers but is prone to false alarm due to low signalto-noise ratio (SNR). The proposed algorithm in this work tries to preserve the good performance for high SNR data while training it to be noise agnostic in addition, by combining RF distillation with denoising in the training process.

השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר

Electrical Engineering Systems Seminar

Speaker: Roy Agmon

M.Sc. student under the supervision of Prof. Arie Yeredor

Monday, 29^th July 2024, at 15:00

Room 011, Kitot Building, Faculty of Engineering

Semi-Blind Separation of Complex-Valued Gaussian Sources

Abstract

Blind Source Separation (BSS) is a well-known problem in signal processing which aims to recover unobserved statistically independent sources (signals) based on observations of their mixtures. The term ``blind" reflects the facts that the sources are not observed and that there is no prior information available about the mixture or about the distribution of each source. If prior information regarding the sources' joint distribution is available, the problem is termed ``semi blind". Independent Component Analysis (ICA) is a common technique to separate independent sources in a single set. If there are multiple sets such that each set contains independent sources but the sources are correlated across sets, the problem is termed Joint BSS (JBSS), or semi-blind JBSS, and the common separation technique is Independent Vector Analysis (IVA). As part of this work we show that using IVA we can exploit the correlation between sets in order to separate sources that are not necessarily separable in a single set using ICA. The quality of separation is quantified by the interference-to-signal ratio (ISR), which measures the residual energy of a source in the reconstruction of another source.

This work addresses the semi-blind JBSS of a particular type of sources, which are complex-valued and Gaussian distributed. Our interest in complex-valued sources requires us to address the two types of complex-valued distributions, “circular” and “non-circular”, and to examine their statistical properties, mainly their covariance and pseudo-covariance matrices. The prior knowledge regarding the sources' distributions, which are Gaussian with known (zero)-mean, covariance, and pseudo-covariance matrices, gives rise to a Maximum Likelihood Estimation (MLE) - based separation approach, which exploits the prior information regarding the sources' joint distribution.

We present the mathematical derivation of the MLE – based separation approach, including performance bounds analysis and a comparison of the resulting ISR to its lower bound, the induced Cramér Rao Lower Bound (iCRLB). We also demonstrate by our simulation results the ability to use IVA in order to separate sources that are not separable in a single set.

השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר