In the data driven world, it is becoming harder to take decisions because of the information overload and data noise. What should we do if we are not sure about the impact of the change we are bringing on our platform? Here A/B test comes in between to help us decide if this change actually had an impact, the impact could be positive, negative or neutral.

Driven by rewards and guided by their own life experience, AI agents are much like us. They can be clueless or wise, playful or mature, newbies, or pros. In this session, you will understand how those attributes can be formalized as code using Reinforcement Learning. In this session, you will understand how such attributes can be expressed in code, as we explorer the fundamental ideas behind Q-learning and Deep RL. Eventually, we use TensorFlow to solve a well familiar problem in this area.

Extracting key-fields from a variety of document types remains a challenging machine learning problem. Services such as AWS and Google Cloud provide text extraction products to "digitize" images or pdfs. These return phrases, words and characters with their corresponding coordinate locations. Working with these outputs remains challenging and unscalable as different document types require different heuristics with new types uploaded daily. Furthermore, a performance ceiling is reached even if algorithms work perfectly equaling the accuracy of the service OCR.

We propose an end-to-end scalable solution utilizing deep learning and OCR architecture to automatically extract important text-fields from documents. Computer vision algorithms utilizing deep learning produce state-of-the-art classification accuracy and generalizability through training on millions of images. Region proposals are generated by off-the-shelf OCRs including Tesseract. We compare the in-house model accuracy with 3rd party OCR services.

Bill.com is working to build a paperless future. We parse through millions of documents a year ranging from invoices, contracts, receipts and a variety of other types. Understanding those documents is critical to building intelligent products for our users.

In this session Max will show you how to mix-and-match a TensorFlow model with SwiftUI to create a magical experience on iOS. While some familiarity with Swift might be nice to have, this session has been specifically built for Python programmers!

Pandas is a powerful library but there is a lot of misguided information and advice. This talk will discuss best-practices for your pandas code to make it easy to write, read, and debug.

This talk will describe how Gousto, a leading recipe box service based in the UK, is using python to build a personalisation ecosystem. Our menu planning optimisation algorithm allows us to create the perfect mix of recipes, ensuring a variety of dish types, cuisines and ingredients. Our recommendation engine sitting on top of this can then offer each customer a personally curated menu, making sure that all users have meaningful choice. All this while ensuring that we are also optimising for maximum performance from an operational point of view!

The talk will give an overview of our methods, our infrastructure, our results and everything that we have learnt along the way.

Introduction to NLP, Examples in real-life(Voice Recognition, Text-to-speech, Sentiment analysis, Perception analysis), Elements of NLP(Bag of words, Truncation Techniques, Term Document Matrix, Feature Extraction), an in-depth example of general sentiment analysis problem, peek into recent advances in NLP

There is obviously a need to combine multiple data sources (structured, unstructured or real time) for AI and ML training. In almost all the ML and Data tools currently in the industry, we bring data from each polyglot source individually to the consuming tool (e.g Python or Julia and R) and then do the joining of data. This causes considerable use of memory just to store data and process he join. We can create a apriori layer in front of these tools to do the heavy lifting of data and then consume through simple SQL - thus using our AI and ML platform for the compute and model training, rather than storing the data frames. The talk shows an indepth analysis of managing network, unification of polyglot persistence and seamless access of all kinds of data for the faster and efficient processing of ML applications.

Enterprise ML - challenges and solutions

Need a nice knowledge graph visualization? Graphviz is not interactive and is difficult to customize. D3 is interactive but I don't wanna write JavaScript. Having a Python library that makes interactive graph visualization in Jupyter notebook is like a dream and we will show you how it came true.

During a crisis such as COVID-19, through social media, effective community management has been realized by first identifying users with concerns and users that can provide support and then connecting (matching) them. The diverse user perspectives help individuals seek informative care through experience or action-oriented information. However, the matching of a user with a concern, a support seeker (SS), and a user with relevant experience, a support provider (SP) is the responsibility of human moderators, who scan the posts and use their medical expertise to find suitable matches. Thus, an automated system that captures medical knowledge implicit in the posts to match a support seeker with a support provider efficiently can significantly help both users and moderators, especially during a crisis. In this talk, I will describe the procedure for explainable data creation, contextualization, abstraction, and matching algorithm that would efficiently recommend SPs to SS in real-time.

Transformer-based decision-making recommenders for E-commerce are taking over older and obsolete algorithms such as RNNs. E-commerce is primarily based on supply chain management. When we purchase a product online, it immediately triggers a chain of events in production, warehouse, and delivery management.
The presentation starts by explaining why RNNs are obsolete and the architecture of Transformers. The presentation dives into Python source code to show how to generate MDP sequences in Python that take product delivery constraints into account. A transformer-based recommender Python notebook will predict decisions. Finally, a Python program will simulate real-time production as soon as a consumer purchases a production online.
By the end of the presentation, you will understand why transformers are replacing RNNs and how to use them in real-life.

We'll start with the importance of Hyperparameter in predictive modeling algorithms. Starting with the very basic HPO algorithms like grid search and random search, we'll jump to advanced Sequential Model based Bayesian Optimization(SMBO) algorithms. We'll discuss a little mathematics behind SMBOs and have a hands on session on libraries like Hyperopt and Optuna.

Apart from answering a question with data, data science projects often require to create a tool that uses machine learning models to do something useful.
Machine learning models that do not make it to production cannot provide enough business impact. Some of the challenges to operationalize a machine learning model are technical but others are organizational.

This talk begins with introducing some key challenges and how MLOps can help to simplify some of the problems. In this talk, more focus is given on the MLOps concepts and ideas rather than the technical specifics. Tools and technologies often evolve and change quickly, but the basic principles likely remain the same.

Specifically in this talk, Tajinder covers: What is deploying to production? What are the steps for deploying a model into production? We will then delve deeper into how MLOps processes can be infused in key steps of model life cycle - Build, Preproduction, Deployment, Monitoring, and Governance.

This presentation is ideal for machine learning engineers and data scientists who want to learn more about deploying machine learning models in production.
Attendees will walk away with a better understanding of the challenges while operationalizing machine learning models and how to apply MLOps processes to address some of the challenges