Applications

• The Next Generation of Data Catalogs: Data Discovery Platforms
• Machine Learning in Action: The Approach of Booking.com
• Exploring the Netflix Cosmos Platform

Theory

• Deep Learning for Phenotype Compound Screening
• Gradient-Guided Dynamic Efficient Adversarial Training

Dataset for Mathematical Problem Solving

As an individual with over 13 years of experience in the data realm, I have observed the evolution of the "data-driven" movement firsthand. Prior to founding and selling my first data startup, I worked as a statistical analyst developing sales forecasting models in R, a software engineer implementing data transformation tasks, and a product manager analyzing user behavior through A/B testing. These roles taught me the critical importance of understanding the context surrounding data, including its origins, last update, and potential for integration with other datasets. This comprehension is vital for unlocking the true value of data and achieving successful outcomes.

Nevertheless, accessing this context can be challenging. Often, the nuances of data reside only in the minds of engineers or analysts who have interacted with it recently. When other data users seek to grasp the context, they typically must consult those who have worked with the data previously. This reliance on personal knowledge becomes increasingly problematic as organizations grow. Identifying the right individual with the necessary context can be time-consuming and may require multiple conversations to gain a complete understanding of the data in question.

The Rundown

Machine Learning in Production: The Booking.com Strategy

In the past five years, Booking.com has integrated machine learning as a fundamental tool for product development. Machine learning now influences every aspect of the customer journey, with hundreds of data scientists deploying and experimenting with numerous models that reach millions of users daily.

Supporting machine learning at scale presents various challenges, particularly in terms of deploying models to production reliably and quickly, while accommodating a wide range of model types and monitoring techniques. Emphasizing the core value of diversity as a strength, we developed a system that supports a plethora of machine learning approaches. Here, we introduce RS, our Machine Learning Productionization System.

The Netflix Cosmos Platform

Cosmos is a computing platform that marries the strengths of microservices with asynchronous workflows and serverless functions. It excels in managing resource-heavy algorithms coordinated through complex hierarchical workflows that can span from minutes to years. Cosmos accommodates both high-throughput services, utilizing vast CPU resources, and latency-sensitive tasks requiring prompt computation results.

This article will delve into the reasons behind the creation of Cosmos, its operational mechanisms, and key takeaways from our experience.

Deep Learning for Phenotype Compound Screening

Phenotype-based screening offers advantages over target-based drug discovery; however, it faces scalability challenges and lacks insights into drug action mechanisms. A chemical-induced gene expression profile can reveal mechanistic signatures of phenotypic responses, but the associated data is often sparse, unreliable, and has limited throughput. We present DeepCE, a novel approach utilizing a mechanism-driven neural network that incorporates graph neural networks and multihead attention to model chemical-gene interactions. Our innovative data augmentation technique leverages information from unreliable experiments, demonstrating that DeepCE outperforms existing methods. We also illustrate its application to drug repurposing for COVID-19, yielding promising novel compounds.

Dynamic Efficient Adversarial Training

Adversarial training is recognized as an effective yet resource-intensive method for training robust neural networks capable of withstanding adversarial attacks. In response to this inefficiency, we propose Dynamic Efficient Adversarial Training (DEAT), which gradually increases adversarial iterations throughout the training process. Our theoretical findings link the lower bound of a network's Lipschitz constant to the magnitude of its partial derivatives regarding adversarial examples, guiding us in refining training procedures.

Introducing the MATH Dataset

Mathematical problem-solving remains a skill that eludes current computer capabilities. To assess this ability in machine learning models, we unveil MATH, a dataset comprising 12,500 complex competition math problems, each accompanied by detailed solutions. This resource aims to enhance model training and improve accuracy, though our findings indicate that merely increasing resources and parameters may not suffice for achieving strong mathematical reasoning.

Stay Informed

That wraps up ML UTD #43. The landscape of academia and industry is ever-changing! To stay updated, be sure to visit the Life With Data blog, read articles on Medium, and follow us on Twitter.

The video titled "WWDC 2022 - June 6 | Apple" explores groundbreaking innovations and updates from Apple's developer conference, showcasing the latest advancements in technology.