PyData Prague #26 - Table Diffusion

👋 Let's diffuse into our next event and have great round-table interactions!

🗓️ The 26th Prague PyData meetup will take place at MSD. The talks will start a bit earlier this time at 18:00 and we encourage you to come as soon as 17:30 to enjoy the opportunity to socialise and refresh yourselves (which you can continue doing during the break and after the talks).

🤗 Our main goal is to build the community around Python and data and make it welcoming to people of various skills and experience levels.

⚡ If you are interested in giving a lightning talk (up to 5 minutes to present an idea, tool or results related at least to some degree to Python and/or data), please contact us before or during the event.

📢 Defect detection in X-ray images of solid tablets. Data augmentation with Stable diffusion
Zdeněk Morávek
Data augmentation is a standard method applied to improve the training of supervised machine learning systems. It performs transformation of existing data such as rotations, clipping, scaling etc. The method proved useful, still there are some treats of the original data that affect the efficiency and scalability of the augmentation.
Generative AI allows to create synthetic data from original dataset. The synthesis is virtually limitless and the synthetic data does not share any treats with the original data. This makes it a powerful extension for data augmentation, especially if the original dataset is limited. There is still a question whether the synthetic data represents well the original dataset.
We applied generative algorithm of stable diffusion to generate synthetic cracks in solid state tablets. The dataset is limited in size and the cracks are a low contrast objects with variable properties. We developed a Mask R-CNN classifier and trained it with available dataset as a baseline model.
We selected suitable images for training the stable diffusion generator and created a synthetic dataset. We investigated statistics of pixel properties of the real and synthetic datasets showing that the main features are conserved though details differ. We used the synthetic data to train an alternative model and compare its performance to the baseline. We demonstrated that in terms of accuracy, we can achieve improvement, but on the other hand we observed higher false positive ratio and also reduced applicability to qualitatively different datasets. We discuss reasons behind these observations and how to improve on them.

📢 AI-Generated Tabular Synthetic Data: What It Is, How It’s Created, and Its Applications
Ivona Krchová
Synthetic data has become an important tool in data science, offering a way to generate realistic data while preserving privacy. In this talk, we’ll explore AI-generated tabular synthetic data—what it is, how it’s created, and how it can be used effectively in various contexts.
I’ll begin with a short overview of synthetic data, explaining its key concept and how it differs from traditional data anonymization techniques. Next, I’ll briefly describe the algorithm developed by MOSTLY AI for generating tabular synthetic data. Finally, we’ll explore key use-cases, we’ll discuss how synthetic data can be used to enhance datasets, address missing values or mitigate bias in model outcomes.

The venue will open at 5.30pm but the intro won't take place sooner than at 6:00pm. There will be refreshments available, 🤗 sponsors.

Please, RSVP here.
See you soon,
PyData Prague team