PyData Warsaw 2019
2019-12-12..2019-12-13
Warsaw, Poland
| title | author | length |
|---|---|---|
| Showreel | 03m | |
Claims that AI systems amplify existing prejudice with concrete examples of biased AI increasing societal bias, shows available trusted AI tooling from cloud… | ||
| The Ethics of Artificial Intelligence - What... | Vince Madai | 39m |
Argues for concrete evidence not abstract warnings: cites Google labeled people as gorillas and later fixed translation bias, shows a speech-based Alzheimer… | ||
| Trusted AI – Building Reproducible, Unbiased and... | Romeo Kienzler | 28m |
Presents concrete methods and working Python code for reproducible, auditable machine learning by capturing experiment provenance and dataset lineage, using… | ||
| Transfer Learning - Entering a new era in NLP | Malte Pietsch | 45m |
Transfer learning is now practical: BERT and Transformer variants move from papers to production, Google reports BERT serving ~10% of US search queries, and… | ||
| Machine Learning can't do the thinking | Inga Strumke | 41m |
The core claim is that no AGI yet and urgent attention belongs to data and ethics not speculation about sentient machines The talk names concrete systems and… | ||
| Disease Modeling with Scipy and PyMC | Dean Langsam | 30m |
Uses SciPy ODE solvers and PyMC MCMC to fit interpretable compartmental models to Israeli Ministry of Health data, estimating the force of infection (beta)… | ||
| The NLU Orchestra | Amit Beka | 28m |
Claims that linear pipelines fail for real conversations and replaces them with a DAG of micro-models that each expose a small, standardized output contract so… | ||
| Reproducible Machine Learning | Mateusz Opala | 30m |
Argues that reproducibility fails from missing data versions, hidden randomness, and untracked experiments and gives concrete fixes: Netguru uses mlflow and… | ||
| TrashAsistant: A kivy App, which uses Deep Neural Networks... | Olgun AYDIN, Krystian Zieliński | 29m |
Prototype mobile app built with kivy app that classifies trash photos using VGG16 transfer learning and a custom top, trained on about 4,000 images. Two… | ||
| Posterior Collapse in Deep Generative models | Michał Jamroż | 28m |
Evidence shows vanilla VAE on CelebA 64×64 (≈200k images, 150k train) with a 32‑D latent often posterior collapses into the prior, so reconstructions preserve… | ||
| Learning to rank with the Transformer | Tomasz Bartczak, Radoslaw Bialobrzeski | 29m |
Applies the Transformer encoder with self‑attention to score whole item lists and reports measurable NDCG gains: on Web30k (30,000 queries) the model beats… | ||
| Analysing Russian Troll Tweets data with Python | Mia Polovina | 38m |
Analysis uses the FiveThirtyEight/Clemson dataset (about 2,98 million tweets from 2,848 handles) with Python for cleaning, UMAP clustering and time-series… | ||
| Anyone can Build Great Deep Learning Applications - Deep Numpy | Cyrus Vahid | 30m |
Claims that existing NumPy code runs unchanged but on GPU via Deep Numpy, with MXNet NDArray compatibility and multi‑GPU parallelism. Demonstrates transfer… | ||
| How to manage data-related projects and not fail (too often)? | Jakub Nowacki | 28m |
Compares CRISP‑DM and Scrum, argues for Kanban over Scrum for exploratory data work, breaks roles into data analyst/data scientist/machine‑learning‑engineer… | ||
| Visual Search | Marcin Tuszyński | 28m |
Walks through building Allegro's visual search engine with concrete results and tools: ResNet embeddings extracted from the penultimate layer, perceptual DCT… | ||
| Football video analysis using Deep Learning | Jacek Komorowski | 35m |
Combines classical CV and tailored CNNs (inspired by SSD) to detect ball and players from CCTV‑style long shots, trains on ~1,000 manually labeled frames and… | ||
| Transfer learning for image recognition in healthcare... | Michał Kierzynka | 30m |
Experiments use ImageNet‑pretrained Inception v3 and ResNet plus a small custom CNN on four datasets (histopathology patches for cancer, retinal scans ~4k… | ||
| keras-fsl: Fast Model Builder for Production-Ready Few... | Clément Walter | 24m |
A reusable framework built on tf.keras that implements metric-learning pipelines for few-shot problems, providing model implementations, dataset loaders, and… | ||
| How to numerically represent semi-structured log data for... | Marcin Kowiel | 24m |
Pipeline converts raw logs into embeddings and normalization rules and evaluates anomaly detection on real data rather than promises. Uses a character‑ngram… | ||
| Sound Modelling - parametric methods and deep learning... | Pawel Cyrta | 30m |
Compares concrete parametric pipelines — source‑filter models, SMS tools and the WORLD model — with learned generators like WaveNet/Magenta and Melotron… | ||
| Generative Text Modelling: Scratching the surface | Tomasz Dziopa | 25m |
Overview gives concrete results not hype it compares 3-layer LSTM character model trained on 16 MB Polish Wikipedia on Google Colab producing readable Polish… | ||
| In the service of the history. AI in archivistics. | Adrian Boguszewski, Natalia Ziemba-Jankowska | 30m |
A pipeline labels and recognizes faces on archival photos and ships as working prototypes, not slides. Trained on a long‑tailed set with 250k published and 37… | ||
| What's coming in Apache Airflow 2.0 | Jarek Potiuk | 32m |
Apache Airflow 2.0 is not backwards compatible and moves to Python 3 only; a team generates about 130 operators with a code-generator that speeds productivity… | ||
| Command line language – where NLP and cyber security meets | Zuzanna Kunik | 26m |
Built a pipeline called CMDLang framework that tokenizes shell input with bashlex plus a custom splitter, trains POS and NER models and surfaces anomalous… | ||
| How to effectively extract image features – let’s play with... | Filip Geppert | 22m |
Concrete pipelines and working scripts using OpenCV 4+ pythonic API: one project builds HSV + color masks, erosion/dilation, Canny edges and HoughLinesP to… | ||
| Predicting flight compensation - a case study | Adam Witkowski | 30m |
A machine learning pipeline verifies EU compensation claims using purchased flight and METAR weather data, trained with about 250 features on LightGBM and… | ||
| Machine Learning on big data in security applications | Yury Kasimov, Olga Petrova | 27m |
Avast says it runs Python with Apache Spark and Kafka over HDFS to process hundreds of terabytes and ~450 million samples, extracting 500+ static features… | ||
| Adding Narrative to BI Dashboards with Natural Language... | Žygimantas Medelis | 26m |
Claims that adding generated sentences to dashboards makes data actionable and speeds work by automating repetitive reports. Uses Accelerate Text (NLG) paired… | ||
| Geospatial analysis made easy with PostGIS and Geoalchemy | Nicolas Pierre, Ioana Gherman | 30m |
Shows how PostGIS stores shapefiles and OpenStreetMap extracts, how PostGIS spatial index and bounding boxes cut million‑polygon queries down to milliseconds… | ||
| How to structure PySpark application | Przemek Chrabka | 30m |
Shows a concrete PySpark app layout: extract input/output and app name to a config file, add a main entrypoint that loads a job from a jobs package, move UDFs… | ||
| Modern Machine Learning flow with Quilt and Polyaxon | Robert Kostrzewski | 34m |
Shows a concrete ML workflow: Quilt (S3‑backed data versioning) as an alternative to DVC for dataset packages, metadata and revision switching; Polyaxon on… | ||
| Automation: Build A Training Pipeline | Varun Kochar | 23m |
www.pydata.org
Many of us knows how to train & deploy ML models in cloud, but doing so have we become redundant. Running multiple experiments in single… | ||
| ML model from an idea to production with the help... | Martyna Urbanek-Trzeciak | 25m |
Python pipeline moves a video‑ad completion idea into production using pandas, NumPy, scikit‑learn and TensorFlow with an in‑house Sroka data connector. A… | ||
| TDD shouldn't be TDDious | Chris Sidebottom | 29m |
Shows concretely how notebook exploration becomes a tested pipeline: extracts a Pokemon notebook into small functions, uses pandas and CSV fixtures, stitches… | ||
| Unsupervised learning for news summarisation | Jakub Kubajek | 26m |
Unsupervised pipeline extracts topics from ~400 articles per day, built for Polish and running in production at the Chancellery for eight months. It scrapes… | ||
| How we personalized onet.pl with multi-armed bandits | Artur Bujak, Dominik Rusiecki | 27m |
Personalization on onet.pl uses multi-armed bandits to pick ten articles from hundreds for ~11 million monthly users, with implementations of epsilon‑greedy… | ||
| How to efficiently model learner’s knowledge with... | Mikolaj Olszewski, Mateusz Otmianowski | 31m |
Recreates Deep Knowledge Tracing with recurrent nets and entity embeddings and reports concrete results and engineering lessons the model uses an embedding… | ||
| Sentiment analysis of tweets in Polish language using deep... | Joanna Piwko | 27m |
Detects five emotions — happiness anger sadness fear disgust — for Polish tweets using a pipeline that goes from collection to model training. Data come from… | ||
| Machine Learning Spacecraft Designing for Cybersecurity | Marina Volkova | 25m |
Claims near perfect detection on validation sets using scikit-learn SVM and MLP, trained on about 660000 generated samples collected from pen test labs and… | ||
| Detection of solar panels based on aerial images of the city of... | Michel Voss | 18m |
Claims automated detection yields a citywide lower bound of 476 panels found across the sample area, about 2% of buildings, using high‑resolution aerial TIFFs… | ||
| nothing matches | ||