ML · Model evaluation · Research tools

Tomer Zipori

Playful questions. Rigorous answers.

I build models, experiments, and tools to understand how people—and machines—make sense of the world.

Selected work

Questions with somewhere to go.

Three finished projects where a claim becomes a tool, a workflow, or a picture.

01 · Visual explanation

OrdinalRegressionViz

Visual tools that connect model parameters to latent evidence, response thresholds, and uncertainty.

Open the package

Latent signal and noise distributions with ordered response thresholds
Evidence, criteria, response.

Completed experiment

Can gradient descent learn around padded zeros?

A BatchNorm implementation detail clearly changed internal activations. Across five paired seeds, its effect on intent-classification macro-F1 was smaller and uncertain.

Read the full experiment Study note Study complete · 27 runs

Across five seeds, current BatchNorm macro-F1 falls as training padding increases while masked BatchNorm stays level, but the interaction interval is wide.
Five paired seeds · best-checkpoint test macro-F1

Garden

Follow a question.

These paths connect my projects, notes, and longer articles. Pick a theme, or take an unexpected turn.

Explore the full garden

About me

Let’s talk about statistics, models, and deep neural networks.

My background spans psychology, experimental methods, statistics, data science, and machine learning. I’m facinated by statistics and machine learning. I explore them and turn research questions into models, visual explanations, and even sometimes tools that other people can use.

More about me CV

Tomer Zipori outdoors, looking toward the camera