Başak Ersöz

PhD candidate · Politecnico di Torino

Where machine learning meets light

I study how machine learning and data science can make sense of imperfect physical measurements. My research brings together machine learning, computational imaging and experimental photonics to turn real optical data into reliable information.

Explore my research View publications ↗
Machine learningComputational imagingExperimental photonics

01 / Research

Understanding the signal.
Revealing the structure.

I have worked on appliance sensors, on vehicle networks and now on a semiconductor laser. In each case the useful question was the same: which measurements are reliable enough to interpret, and how do they turn into a decision?

01Scientific machine learning

Reliable learning from real measurements

Experimental data rarely arrive in ideal conditions. I extract statistical and derivative features from raw waveforms, then use PCA, t-SNE and DBSCAN to separate stable operating regimes from unstable ones before any classification. The recurring finding is that unstable measurements dominate the error budget: choosing which data to trust often matters more than choosing a more complex model.

A t-SNE scatter plot of twelve thousand self-mixing measurements, coloured by target class, showing six separated groups.
Twelve thousand self-mixing measurements projected with t-SNE and coloured by target. The six targets separate reasonably well in this view.

Feature extraction (RMS, variance, peak-to-peak and derivatives) · PCA · t-SNE · DBSCAN · Regime selection · Classification · Per-class evaluation

02Applied machine learning

Learning across signals & systems

My earlier work explored LSTM-based electricity-price forecasting under volatile market conditions and ensemble methods for ECG classification. These projects connect time-series modelling with careful evaluation, from energy-market data to imbalanced signal datasets.

A line chart of electricity price in dollars per megawatt hour over 168 hours, with the measured series drawn as a solid navy line and the LSTM forecast as a dotted amber line.
One week of day-ahead locational marginal prices against the 24-hour-ahead LSTM forecast. The model follows the baseline closely and underestimates the sharpest spikes, which are the hours a trader actually cares about. Figure from our GPECOM 2023 paper, © 2023 IEEE.

Time-series forecasting; recurrent networks; ensemble learning

Explore related work ↗
03Computational imaging

Images from a single detector

I reconstruct target images from scalar measurements and experimentally recorded speckle patterns. My workflows connect sensing-matrix construction and pseudoinverse or sparsity-based reconstruction with quantitative analysis of pattern count, alignment error and model–measurement mismatch. More patterns help only while the measurement model stays consistent; under misalignment, extra data can make the reconstruction worse.

Three panels side by side: a noisy compressed-sensing reconstruction of the characters ID cm, the same image after a U-Net denoiser, and the clean ground truth.
A single-pixel target recovered from compressed-sensing measurements and then denoised by a U-Net. The network restores the overall shape but smooths the finest strokes. Figure from our paper in Proc. SPIE 13375, AI and Optical Data Sciences VI.

Compressed sensing · Sensing-matrix design · Moore–Penrose pseudoinverse · TVAL3 and OMP · SSIM, PSNR and RMSE evaluation

Explore related work ↗
04Experimental photonics

Optical feedback & self-mixing

Using a semiconductor laser as both source and sensing element, I study how light reinjected into the laser cavity carries information about a target and its environment. Within MIRABILIS I record self-mixing waveforms and photodiode signals on an optical feedback bench with a nanofiber scatterer, a motorized stage and a CMOS camera, and I control the acquisition conditions that decide whether a measurement is usable at all.

The experimental self-mixing set-up on an optical table: laser, attenuator, beam splitters, motorized translation stage, CMOS camera and target, with the components numbered and the nanofiber scattering element shown in a close-up inset.
The experimental self-mixing set-up I work on. The numbered components are: (1) 830 nm semiconductor laser, (2) attenuator, (3) 90/10 beam splitter, (4) motorized translation stage, (5) CMOS camera, (6) nanofiber scattering element (shown in the bottom inset), (7) 50/50 beam splitter, and (8) target. Back-reflected light from the target is reinjected into the laser cavity, producing the measured self-mixing waveform.

Optical feedback and self-mixing · 830 nm semiconductor laser · Speckle illumination · Nanofiber scattering elements · Controlled experimental acquisition and alignment

Explore related work ↗

Current doctoral project

MIRABILIS

Optical feedback, single-pixel imaging and machine learning for compact photonic sensing. I work on an optical feedback platform built around an 830 nm semiconductor laser, a nanofiber scatterer and speckle illumination, where light reflected from the target is reinjected into the laser cavity. A PRIN 2022 collaboration between Politecnico di Torino, Università di Bari Aldo Moro and Politecnico di Bari.

Project website ↗Project GitHub ↗

02 / Selected work

Questions explored.
Findings shared.

All publications & presentations ↗

03 / About

A researcher.
An engineer.
A curious mind.

Başak ErsözExplore my background ↗

I received the B.Sc. and M.Sc. degrees in electrical engineering from Yildiz Technical University, Istanbul, Turkey. I am currently pursuing the Ph.D. degree in Electrical, Electronics and Communications Engineering with the Politecnico di Torino, Turin, Italy.

From 2018 to 2020, I was an R&D Engineer in Sensors and Mechatronics with BEKO. From 2020 to 2022, I was a Multimedia and Connectivity R&D Engineer with Ford Motor Company, where I worked on automotive connectivity systems and intelligent vehicle technologies.

With more than seven years of industrial and research experience, my research interests include AI for scientific measurement and experimental systems, data science for experimental sensing and imaging, optical sensing systems, sensor system development, smart automotive and consumer dedicated engineering applications.

Mar 2024 – present

PhD candidate · Electrical, Electronics and Communications Engineering

Politecnico di Torino

2021 – 2023

MSc · Electrical Engineering

Yıldız Technical University

2014 – 2019

BSc · Electrical Engineering

Yıldız Technical University

Scientific computing

MATLAB · Python · Simulink · SQL · C (basic) · Numerical modelling · Reproducible data processing · Scientific visualisation

Machine learning & data

Data preprocessing and feature engineering · Linear and nonlinear dimensionality reduction and visualization · Unsupervised learning and density-based filtering/anomaly detection · Supervised classification and ensemble learning · Deep learning and convolutional architectures · Recurrent neural networks and temporal modeling · Autoencoders for representation learning, reliability, and anomaly scoring · Domain adaptation and transfer learning · Model selection and validation · Per-class and statistical performance evaluation

Photonics & imaging

Self-mixing interferometry · Semiconductor lasers · Speckle patterns · Single-pixel imaging · Compressed sensing · Sensing-matrix construction · Inverse reconstruction · SSIM and PSNR evaluation · Experimental acquisition and alignment

Engineering systems

Sensor technologies · Embedded sensing · CAN · I²C · Bluetooth · Automotive connectivity · Requirements engineering · System integration and validation

07 / Connect

Good research starts
with a conversation.

Interested in machine learning for experimental systems, computational imaging or optical sensing? I welcome conversations about research and collaboration.

Google Scholar ↗LinkedIn ↗GitHub ↗ORCID ↗Download CV (PDF) ↗