Observation
Something in a paper, a training run, or a conversation raises a question that does not have an obvious answer. That question becomes the seed.
About the author
Teodora Vašíček
Deep Learning Researcher & Editorial Voice at swienc
A researcher who writes, not a writer who researches
Teodora Vašíček has spent over a decade working at the intersection of applied mathematics and machine learning, with most of that time focused on the mechanics of how deep networks actually learn.
Her work began in numerical optimisation, specifically studying why certain gradient descent configurations converge faster on ill-conditioned loss surfaces. That narrow focus gradually widened. She became interested in the architectural decisions that practitioners make almost by habit — residual connections, batch normalisation placement, dropout scheduling — and whether those decisions hold up under scrutiny when you change the data distribution.
At swienc, she writes about these questions directly. Each piece starts from a specific technical observation rather than a broad topic, and the argument builds from there. She does not write tutorials. She writes analysis.
How each piece gets made
The editorial process at swienc is deliberately slow. A single post typically takes three to four weeks from first observation to publication — most of that time spent running experiments and questioning the initial framing.
Experiment
Small-scale tests on controlled setups. The goal is not to produce publishable results — it is to understand whether the intuition holds in practice.
Argument
The writing frames a position. Not a neutral summary of existing work, but a specific claim that can be agreed with or challenged.
Review
A second pass looking specifically for claims that rest on weak evidence. Anything that cannot be defended gets cut or reframed.
Current research threads
Several topics are actively being investigated for upcoming pieces. Some are nearly ready; others are still in the early observation stage where the question is clearer than any answer.
- Layer normalisation placement in encoder-only transformers — draft complete
- Weight initialisation sensitivity in deep residual networks — published
- Dropout placement effects on recurrent architectures — experiments done
- Loss landscape geometry in overparameterised models — argument drafted
- Sparse attention patterns — when they degrade and why — in review
- Gradient clipping thresholds across optimiser families — early observation
- Mixed-precision training failure modes — question stage