PyData Yerevan 2022
2022-08-12..2022-08-13
Yerevan, Armenia
who it's for
data scientists, machine learning engineers, data engineers, MLOps engineers, NLP engineers, research scientists
topics
probabilistic programming with Pyro; GPU-accelerated dataframes with cuDF; AWS streaming data architectures; PyTorch Geometric graph neural nets; AutoML model creation; explainable AI techniques; Delta Lake and Trino lakehouse; Triton inference serving; multilingual large-scale ASR; BERT for healthcare; active learning for 3D segmentation; A/B testing in production
| title | author | length |
|---|---|---|
| Probabilistic Programming and Readable Models | Fritz Obermeyer | 01h 07m |
Argues that probabilistic programming produces readable, interpretable models and demonstrates it with real tools and results rather than platitudes. Uses Pyro… | ||
| Data, AI, Hype and Fear - Recipe for Disaster? | Wolfgang Weidinger | 56m |
Crisp claim: 85% fail to deliver and only 53% reach production, backed by practical experience at a large insurer where a fraud detection project at a global… | ||
| Accelerated Data Science Libraries | Ashot Vardanian | 39m |
Replaces Pandas/NumPy/NetworkX with concrete backends and shows measured outcomes: switching to cuDF or cuPy yields tabular 10–50x speedups and math kernels… | ||
| Building Data Pipelines on AWS | Rudolf Eremyan | 38m |
Practical claims and concrete recipes decide tonight: cost traps that produced an unexpected $500 bill, optimization rules for Athena like Parquet +… | ||
| Using Python for the PCR Design | Ricardas Ralys | 23m |
Claims a working Python pipeline that uses NumPy to build nearest‑neighbor energetic tables, parses up to ~1 million primer/probe candidates, and cuts that… | ||
| The Dangers of Mindless Forecasting | Aghasi Tavadyan | 32m |
Warning that maximization of metrics often produces brittle models; high‑AUC ensembles and neural nets can overfit noisy data, while a simple classification… | ||
| PyTorch Geometric for Graph Neural Nets | Dmitry Korobchenko | 47m |
Live code demo builds a 3-layer GCNConv model in PyTorch Geometric that maps 3-d node features to a 64-d embedding and uses global mean pooling to predict 6… | ||
| ML Platform for Insurance Conglomerate | Dmitry Mezhensky | 30m |
A cloud-native ML platform is built for a major US insurance conglomerate, migrating a legacy on‑prem stack to AWS and automating CI/CD so model delivery time… | ||
| Modern Data Stack: Optimising and Scaling Data in a Tech Company | Nacho Aranguren | 41m |
Argues that adopting DataOps and a dedicated Data+Analytics Engineering function saves engineering time and frees analysts for higher‑value work and shows a… | ||
| Recommendation Systems in Market Research | Davit Abgaryan | 43m |
Platform-level recommender replaces brittle rules with a production factorization approach: factorization machines via xlearn, trained on user survey attempts… | ||
| Cifar-10 Exploratory Data Analysis | Anna Shahinyan | 28m |
Claims that CIFAR-10 can be trimmed without hurting accuracy, citing both a paper that finds 10% subsets retain performance and results from semantic… | ||
| Best Practices for Coding in ML/DS - Techniques to Construct your Project | Mher Khachatryan | 37m |
Advice targets long‑lived ML projects: invest time upfront so future changes stay fast, enforce single responsibility principle and DRY, prefer descriptive… | ||
| Streamlit: A Faster Way to Build and Share Data Apps | Karen Javadyan | 42m |
Claims that turns Python scripts into shareable web apps with minimal code, demos a live app that loads images, runs Keras models and shows side‑by‑side… | ||
| The Structure and Interpretation of ML Metadata | Gevorg Soghomonyan | 28m |
Claims that ML metadata is the development tool that actually makes software 2.0 reproducible and governable, demonstrated by the Aim open source experiment… | ||
| Use AutoML to Create High-Quality Models | Aleksandr Patrushev | 41m |
Claims that AutoGluon plus Amazon SageMaker workflows build competitive models with three lines of code, stacking diverse learners and transformers, and often… | ||
| Tiling & Parallel Processing of Large Images | Anush Tosunyan | 24m |
Pipeline replaces whole-image processing with tile-level work using rasterio to read/write offset tiles, runs models per tile and averages class probabilities… | ||
| Active Learning for 3D Mesh Semantic Segmentation | Erik Harutyunyan | 38m |
Claims that Monte‑Carlo Dropout applied to PointNet++ meaningfully reduces annotation needs for 3D mesh/pointcloud semantic segmentation and shows where it… | ||
| Eating Humble Py: From Toy Problem to Real-World Solution in Predicting CLV | Katherine Munro | 44m |
Claims a pipeline that moves from toy data to real CLV estimates and shows what actually breaks: cohort averages, RFM clustering, and a simple… | ||
| Semantic Multimodal Multilingual Similarity Engine | Vladimir Orshulevich | 41m |
A working semantic multimodal multilingual similarity engine is described with a bias toward GPU‑accelerated inference and distributed deployment, built on a… | ||
| Building your own Multiskill AI Assistant with DeepPavlov | Daniel Kornev | 38m |
Concrete claim: DeepPavlov provides an open source conversational stack that runs a neurosymbolic architecture combining neural annotators and symbolic… | ||
| Moving Inference to Triton Servers | Marine Palyan | 25m |
Moves inference onto Triton Inference Server, splitting pre/postprocessing onto CPU nodes and running compiled models on GPU so a single G4 instance replaces… | ||
| Explainable AI as a Conventional Data Analysis Tool | Maria Sahakyan | 39m |
Claims that explainable ML becomes a conventional analysis step by extracting a SHAP values matrix from a trained model (examples use XGBoost and the shap… | ||
| Grover’s Quantum Search for Data Science and Why should we Care | Tigran Sedrakyan | 42m |
Grover’s algorithm delivers a concrete quadratic speedup (O(√n)) for unstructured search by repeating an Grover iteration: oracle + diffuser roughly π/4·√n… | ||
| Empirical Determinacy of Posterior Location and Scale in Bayesian Hierarchical Models | Sona Hunanyan | 35m |
Claims a concrete diagnostic and provides code: the R package ed4bhm computes a power-weighted likelihood perturbation (w ≈ 1 ± Δ), estimates overlap via a… | ||
| Sequential Attention-Based Neural Machine Translation | Hadi Abdi Khojasteh | 50m |
Shows a working neural machine translation built implementation from scratch using TensorFlow, training on an English↔Portuguese dataset (prayer sentences) and… | ||
| How to Start Critical Thinking in Data Science | Arpi Sahakyan | 37m |
Begins with concrete examples: the birthday paradox (23 people → ~50% match), hospital survival numbers that flip when critical-case counts are revealed… | ||
| Using Few Shot Object Detection for Utility Pole Detection | Mark Hamazaspyan | 27m |
Claim: few shot object detection turns 5–10 annotated S‑pole images into a usable detector by base-training on 2500 T-pole images then two‑stage fine‑tuning… | ||
| The Explainability Problem: Towards Understanding AI | Nura Kawa | 41m |
Argues that explainability remains unsolved and gives concrete failures: off‑the‑shelf saliency maps that radiologists cannot interpret, post‑hoc explainers… | ||
| AI-Powered Solutions for Cybersecurity | Elina Israyelyan | 31m |
Claims a production phishing detector that holds 95% average accuracy since March 2022, trained on combined sources including PhishTank and OpenPhish to a… | ||
| NVIDIA NeMo: Toolkit for Conversational AI | Alex Laptev | 36m |
An open source toolkit that combines PyTorch and PyTorch Lightning with modular collections for ASR NLP and TTS provides pretrained models and step‑by‑step… | ||
| BERT Model for Real World Healthcare Data | Artem Terentyuk | 20m |
A modified BERT model adapts to electronic health records by turning visits into tokens and adding temporal positional encoding to handle sparse messy data… | ||
| Classical Texture Synthesis and Beyond | Hovhannes Margaryan | 28m |
Compares four concrete algorithms: Image Quilting with minimum-cost seam and patch-size/overlap/tolerance knobs (FFT or kd-tree for nearest neighbors), Texture… | ||
| Building a Streaming (E-health) Data Pipeline: When and How? | Zohreh Jafari | 42m |
Argues that streaming is the right choice when data is massive, unbounded, or needs timely answers, and shows an e‑health pipeline that actually runs: a… | ||
| Building a Lakehouse data platform using Delta Lake, PySpark, and Trino | Viacheslav Inozemtsev | 42m |
Clear claim: a practical lakehouse built on Delta Lake table format plus PySpark and Trino yields a single platform for analytics and ML. The demo shows… | ||
| Scaling Semi-Supervised Production-Grade ASR on 200 Languages | Luka Chkhetiani | 44m |
Describes a concrete pipeline that scales semi supervised, self supervised ASR to 200 languages using Wav2Vec2/XLSR style pretraining and downstream fine… | ||
| Large Scale Field Delineation | Hrach Asatryan | 41m |
Uses encoder‑decoder models (ResUNet34 backbone) on Mapbox imagery with three targets: boundary, extent, distance masks; trains on RPG and internal annotations… | ||
| NetworkX - your Unexpected Assistant for Clustering Analysis | Meirav Ben Izhak | 24m |
NetworkX is pitched as a concrete assistant: code converts k‑means cluster centers into a fully connected distance graph (pdist then squareform), prunes edges… | ||
| Bachelor theses in Deep Learning: Submitted to an Armenian University | Sergey Hayrapetyan | 28m |
Bachelor theses present an object‑detection pipeline for top‑down vehicle imagery using 32,000 labeled vehicles from six locations at ~15 cm/pixel so cars are… | ||
| Target-Based Sentiment Analysis with T5 | Liana Minasyan | 23m |
Claims a 20% leaderboard gain using T5 repurposed as a question‑answering model, fine‑tuned with prefix prompts and techniques like layer freezing and cosine… | ||
| A/B Testing in Production | Sona Hambaryan | 36m |
Covers concrete failures and fixes for running A/B tests in production: defines baseline conversion and detectable lift (example baseline 0.10 aiming for a… | ||
| Large Scale Representation Learning In-the-wild | Hadi Abdi Khojasteh | 41m |
Claims that unsupervised and self‑supervised methods match or beat supervised baselines on ImageNet and cut label needs by orders of magnitude, showing an… | ||
| Lightning Talks | Andrey Manoshin, Hayk Aprikyan | 35m |
EENLP presents a searchable index of 90+ datasets and 60+ models, plus newly compiled cross‑lingual commonsense datasets for five tasks and evaluation numbers… | ||
| Near-Duplicate Ad Detection in Online Classified Ad Services | Diyar Mohammadi | 36m |
Problem definition, datasets and concrete gains: models train on ~1M labeled pairs with 200–300 test pairs and production sees ~500k submissions/day and 60%… | ||
| How to use Pandas Efficiently | Hossein Mortazavi | 21m |
Claims that using chaining + assign and pandas' datetime and groupby tools produces clearer, more maintainable pipelines, and demonstrates them with live code… | ||
| A/B Testing in Production | Sona Hambaryan | 36m |
Starts with a concrete claim about experiment setup failures and shows numbers: a 1 000‑user baseline example, seeing 90 vs 128 conversions (≈0.028 lift) but… | ||
| nothing matches | ||