Ashkan Tashk — Postdoctoral Researcher in Machine Learning, AI, Eye-Tracking & Human-Computer Interaction at DTU Compute, Copenhagen

"Reading the Reader" (RtR)

How does typeface design influence the cognitive effort and efficiency of reading? The RtR project combines psychophysics experiments with advanced eye-tracking analysis to quantify how low-level typographic features — font family, style, and spacing — affect perceptual thresholds, visual short-term memory capacity, and processing speed.

Funding: The RtR project is funded by the Novo Nordisk Foundation (NNF) — Data Science Collaborative Research Programme 2022, Grant No. NNF22OC0076877.

Explored Research Topics

• Native Language Identification from Reading (NLIR): Leveraging eye-tracking gaze data to predict readers' native languages from their L2 reading patterns. Using the MECO-L2 dataset with linguistically motivated n-gram features (First Fixation, First Pass, and Total Fixation durations), we evaluate both classical logistic regression and a novel enhanced BERT-based classifier with progressive layer unfreezing. Monte Carlo simulations with 20 random seeds ensure robust generalization across 7- and 12-language classification tasks, with hierarchical clustering analysis validating linguistic coherence of the learned representations.
• Eye-Tracking & Pupillometry: Using Tobii eye-trackers to capture gaze patterns, fixation dynamics, and Task-Evoked Pupillary Responses (TEPR). Full preprocessing pipeline including blink detection, Hampel filtering, Gaussian smoothing, adaptive interpolation, and baseline correction for cognitive load analysis.
• TVA Parameter Estimation: Applying Bundesen's Theory of Visual Attention (TVA) to model visual cognition during reading. Per-participant Bayesian MCMC estimation of K (VSTM capacity), C (processing speed), and t₀ (perceptual threshold) across font conditions using PyMC.
• Font Legibility & Psychophysics: Investigating how font families (Cambria, Roboto, Garamond) and styles (regular vs. italic) modulate reading performance through word superiority effects, P(t) curves, and Bayesian posterior difference distributions.
• Bayesian Hierarchical Models: Growth Curve Analysis (GCA) using orthogonal polynomials for temporal pupil data, hierarchical models for Peak Pupil Dilation (PPD) curves, with emphasis on LOO-CV diagnostics and model identifiability.
• Machine Learning & Deep Learning: Development and application of supervised and unsupervised methodologies including CNNs, LSTMs, Random Forests, and transformer-based architectures (BERT, RoBERTa) for classification, segmentation, and prediction tasks across biomedical and cognitive domains.

• Biomedical Image Segmentation: Designing innovative deep convolutional architectures (AID-U-Net) for semantic segmentation of biomedical images, as well as polyp detection and segmentation from colon capsule endoscopy (CCE) images using combinations of RCNN and DRLSE methods.
• Electronic Health Records (EHR) Modeling: Time-variant event learning from electronic health records using LSTM and Random Forest models for predicting clinical outcomes such as metastatic prostate cancer progression beyond bio-markers.
• Multivariate Data Imputation: Development of the MIPLS2 algorithm for predicting and imputing missing values in two-way large-scale metabolomics and foodomics datasets using PLS2-based approaches with spectral data (H¹NMR, NIR, FTIR).

Multimodal Signal Processing for Attentional Resource Cognition

I co-organize the M-SPARC workshop at IEEE ICASSP 2026 in Barcelona, bringing together researchers on multimodal signal processing — eye-tracking, EEG, physiological sensing — for cognitive state estimation, adaptive interfaces, and human-centered AI systems.