Location is built into the fabric of our daily experiences. Whether you’re exploring a city with Lonely Planet, sharing with friends on Snapchat, seeing if it’s going to rain on Weather.com, tracking breaking news on Bloomberg, or checking your ETA in Lyft — location is essential to every one of these applications. But how can we create precise maps of the world? In the old days you sent thousands of people to do surveys, measure and draw maps. Today, we use sensors to capture the world and AI to make sense of this data. I present industry cases from Mapbox, Facebook, Microsoft and others how machine learning and computer vision allows us to create planet scale maps from data.
So, you trained a model. Now, if you actually plan to use it in production system, you need to wrap in some sort of service and deploy it. Ebonite framework will do it for you in a couple lines of code. It allows you to persist, reproduce and turn into docker images virtually any machine learning model python object.
A Glimpse into Machine Learning Applications in Bioinformatics
Bioinformatics plays a vital role in different fields such as cancer genetics, waste cleanup, evolutionary biology, and many others. In this meetup, we will talk about what is bioinformatics and what questions bioinformatics trying to answer. Also, we will closely examine some problems in bioinformatics and how machine learning can be used to solve them.
How Google Flights, Kayak, and Hooper are using fare predictions. How to preprocess 1.5 billions of flight search data set using Databricks platform and PySpark. What strategy can be used at the preprocessing stage. What external data sources can be used to enhance the model performance. How to overcome Spark MLlib limitations. What ML models were applied to tackle the problem. Which strategy to select during the training process and how to access the results.
Chief Data Scientist, Senior Managing Director @ Sberbank
AI @ Cloud
Organization-wide adoption of AI applications and AI-first paradygm transforms financial business into competitive modern technological platform. Cloud AI solutions are the essential driver for increased speed of innovation and adoption of AI applications. In this talk we will share Sberbank experience with personal Cloud for AI and our thoughts on does an enterprise really need one in order to become AI-ready.
While technology tends to get blamed for failed data science projects, the uncomfortable truth is that management is often culpable. In this presentation, we consider how management, both of models and of humans, can make all the difference.
Overview of the trifecta of major global Cloud Vendors (AWS, Azure, GCP) and their pre-trained Cognitive AI NoOps services for: Vision Image, Vision Text OCR, Vision Video, Audio Transcription and Language Translation. Practical approaches and differences between the major global Cloud Vendors, and their rate of Governance. Hands-on detailed examples and demos as is shared and maintained on GitHub by the Author.
Audio data augmentation for automatic speech recognition (ASR) and keyword spotting
Overfitting of modern large and deep ASR and keyword spotting models remains the largest problem. Audio data augmentation is a very important task that can prevent overfitting and significantly increase accuracy of the target voice processing system. Recently, the simple augmentation method was proposed in the paper SpecAugment (https://arxiv.org/pdf/1904.08779.pdf). In this lecture, we observe these and several other efficient methods for audio data augmentation. One such useful method is a recording of frequency response of the target audio channel and transfer it to the training dataset. We explored a simple way to estimate a frequency response of speech channel and apply it to the sound recorded in a different environment. Finally, we observe a lightweight open-source project (https://github.com/iver56/audiomentations) to sequentially apply several augmentation methods on the fly. Jupiter notebook with demo examples will be available afterward.
PhD Research Scholar @ Thapar Institute of Engineering and Technology
Predicting change-prone source code components for enhanced software maintenance
An effective software change prediction mechanism predicts those code components that are likely to be employed with some change from one version of software to the next. In an actual scenario, change prediction models are capable of being directly incorporated in software developers' analytics dashboards, assisting in software maintenance procedures. I'll highlight the key machine learning techniques that have been useful for this purpose, what machine learning tools have been incorporated for the same, how the datasets have been pre-processed before analysis and how the techniques have been validated to assess their efficacy.
We will learn why object detection is tricky. Then we will review how YOLO (namely v3) framework tackle this problem. Finally we will demonstrate how to extend YOLO framework to solve domain specific tasks.
Step by step implementation of the method Deep CNN-Based Blind Image Quality Predictor (DIQA). Also, I will go through the following TensorFlow 2.0 concepts. Download and prepare a dataset using a tf.data.Dataset builder. Define a TensorFlow input pipeline to pre-process the dataset records using the tf.data API. Create the CNN model using the tf.keras functional API. Define a custom training loop for the objective error map model. Train the objective error map and subjective score model. Use the trained subjective score model to make predictions.
In 2019 it has become feasible to deliver and manage hundreds of ongoing Data Science or AI initiatives sustainably and properly tied to business. Recently a new term called "AI transformation" emerged, as a spiritual successor to "Data-Driven enterprise". But how can one find a proper path to succeed with this AI transformation? In this talk you will find an overview of the transformation processes, as well as the overall maturity model that reflects the success of Data Science function. We will talk about key barriers and potentially useful recipes for both design and implementation of a successful DS/AI strategy.
What is Application Score, how it works? How banks distinguish good clients from bad clients, how to predict customer default? What modelling technics do they use? What if there is no data for some client segment? What requirements does Central Bank put on the modelling technics? I touched these questions just a bit during my last talk at previous Meetup, but I saw big interest from the auditory. Hence, suggesting to talk about this.
Model Insights & Visualization in Financial Services
One of the most critical stages in data science projects is to identify, interpret and visualize the insights generated by machine learning and optimization models. I will outline the framework to be considered when data scientists implement model visualization and I will present use cases in the customer, operations and campaign analytics domain using profiling and visualization outputs which address business requirements for financial services sector.
From oil exploration to icebreakers route mapping, data is a key for effective operations in all kinds of geoscience-heavy business. Pavel will talk about main applications of Computer vision in O&G and other industries that help to cut the time and costs and increase effectiveness and, at the same time, make tradional operations more environment friendly.
17:30 Outro & networking
Что такое Minsk Data Science Major?
Major - крупное мероприятие сообщества, от команды Data Fest. Это что-то большее, чем ODS Meetup, но не такое масштабное, как Data Fest. И случается чаще, чм раз в году.
Все так, мы не берем денег за вход. За выход и материалы, кстати, тоже не берем. С вас только пройти регистрацию, получить приглашение (не бойтесь), и приехать.
Теперь всегда будут Major-ы?
Это новый член семьи наших мероприятий сообщества: ODS Meetup < Major < Fest. Мы продолжим проводить все эти форматы. Fest делаем раз в году, а остальные - почаще.
Что такое ODS Q&A?
Мы уже такое несколько раз делали, начиная с этого мероприятия. Участники сообщества задают разные вопросы про область, инструменты, карьеру, и т.д., мы их группируем, и затем выбранные эксперты на них на сцене отвечают.