>_ Tibor Sloboda

Research

My research in multimodal systems and generative models, earlier work in medicine and chemistry, and selected publications.

I am a PhD candidate in artificial intelligence at FIIT, Slovak University of Technology, working on multimodal systems and generative models. I am interested in how models represent and combine different kinds of information, and in the mathematical foundations of those relationships. My work on cross-modal compatibility takes up that mathematical side.

Earlier work in medicine & chemistry

Much of my earlier research focused on biomedical ML. In histology, I worked on transforming stain images with unpaired GANs, including editable transformations and attention-enhanced models. I also contributed to PCD quant, which uses automatic image analysis to assess ciliary ultrastructure.

In chemistry, I worked on predicting phototoxicity with machine learning and investigating descriptors for that prediction. That research also involved building and maintaining MLTox, an online application through which the methods could be used.

Research & engineering

My interests extend into LLM systems, including co-authored work on guarded query routing. Alongside research, I work on model optimization and AI infrastructure for heterogeneous hardware. I care about what a method can do and what is needed to make it usable under actual data, compute, and deployment constraints.

From a model score to a decision

Prediction tasks such as phototoxicity assessment also raise a practical question: how should a model’s output become a decision? For a binary classifier, a threshold determines which scores count as positive. Lowering it catches more positives, but also produces more false alarms.

The synthetic example below makes that tradeoff visible. Drag the red threshold τ to the left, then to the right, or focus it and use the arrow keys. Watch recall and the false positive rate change together, and the operating point move along the same ROC curve.

Decision threshold explorer: score distributions and ROC curveLeft: two Gaussian classifier score distributions - negatives at mean 0, positives at mean 1.8, both unit variance - with a movable decision threshold at tau = 0.9. Scores above the threshold are called positive; the shaded tails are the true positives and false positives that choice creates. Right: the ROC curve those same two distributions trace as the threshold sweeps, AUC = 0.898, with the current operating point marked. Dragging the threshold moves the operating point and recomputes recall, false positive rate, precision, and F1 from the normal CDF.-2024classifier score snegatives · N(0, 1)positives · N(1.8, 1)τ = 0.90000.50.511false positive ratetrue positive rateAUC = 0.898

recall 0.816fpr 0.184precision 0.655F1 0.727prevalence 30%d′ = 1.8

fig. 1 · decision-threshold explorer. Synthetic Gaussian scores with unit variance, separated by d′ = 1.8. Left: score distributions with the mass above τ shaded · teal is true positives, red is false positives. Right: the ROC curve those distributions imply (AUC = 0.898, analytic), with the current operating point. All metrics follow from Φ: recall = Φ(d′−τ), fpr = 1−Φ(τ), precision from 30% prevalence. Without JavaScript the threshold parks at τ = 0.90.

The model and its ROC AUC stay the same as you move the threshold. The decision rule changes, along with the errors it accepts. Choosing that rule for an application requires evaluation on relevant data, calibration checks, and an understanding of what each kind of error costs. This is the connection between evaluating a predictive model and deciding how to use it; the Gaussian example is illustrative, not a result from a deployed medical system.

Publications & preprints

Google Scholar has my publication record and citation information.

For a research collaboration or help evaluating a model, get in touch.