🛍 Machine Learning at Shopify
Weekly news digest curated by the industry insiders
📝 Editorial
Building large machine learning (ML) architectures remains unexplored territory for most companies. Despite the massive adoption of ML frameworks and platforms, most companies still apply ML in constrained and relatively small-scale scenarios. As an industry, we are still figuring out the best practices for ML infrastructures that can run large numbers of ML models from experimentation to production. Not surprisingly, the best inspiration for large-scale ML architectures comes from technology giants that are running some of the largest ML infrastructures in the world. Companies like Uber, LinkedIn, Meta, and Airbnb have been very transparent about their architectures used to run ML workloads and have even open-sourced many of its components. This week, we have another technology powerhouse to draw inspiration from: Shopify.
A few days ago, Shopify published some details about Merlin, the platform powering its internal ML solutions. Merlin is based on a very modern architecture optimized for rapid experimentation and scale. At a high level, Merlin shares some similarities with architectures such as Uber’s Michelangelo or Airbnb’s Bighead but it also has some very unique characteristics. For instance, Merlin uses Ray as its fundamental engine for ML scalability. Merlin also uses Pano, a custom feature store that persists and enables features across all ML models. Another interesting area of innovation of Merlin is its native integration with Notebook environments and the consistency of its project structure. Even though it remains close-sourced, the initial details of the architecture can serve as inspiration to organizations building ML solutions at scale.
🔺🔻TheSequence Scope – our Sunday edition with the industry’s development overview – is free. To receive high-quality content about the most relevant developments in the ML world every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻
🗓 Next week in TheSequence Edge:
Edge#183: we explore data vs model parallelism in distributed training; discuss how AI training scales; overview Microsoft DeepSpeed, a training framework powering some of the largest neural networks in the world.
Edge#184: we look inside DALL-E 2 and learn how OpenAI upgraded its supermodel that can generate artistic images from text. Subscribe if you haven’t yet
Now, let’s review the most important developments in the AI industry this week
🔎 ML Research
Converse
Salesforce Research published a paper detailing Converse, a framework for building modular, task-oriented chatbots →read more on the Salesforce Research blog
Zero-Shot Task-Oriented Dialogue
Google Research published two papers outlining methods for task-oriented conversational agents that can transfer knowledge across different tasks →read more on the Google Research blog
Contrastive Learning
Stanford University published a detailed blog post explaining the underpinnings of contrastive learning →read more on the Stanford University blog
Contrastive Learning on Image-Text Data
Google Research published a paper proposing a contrastive learning method that matches text to pretrained images but does so in a way that can transfer knowledge across different tasks →read more on the Google Research blog
🛠 Real World ML
Shopify Merlin
Shopify published a blog post detailing Merlin, an internal platform that powers their ML pipelines →read more on the Shopify Engineering blog
Feathr
Linked open-sourced Feathr, a feature store used in their internal ML applications →read more on the LinkedIn Engineering blog
Presto on Kafka
Uber published a blog post illustrating their architecture for running SQL queries using Presto over Kafka data streams →read more in the Uber Engineering blog
✏️ A Survey: Data Labeling for ML, part 4
Please take a very simple survey to help us prepare an article about data labeling. It will take about 2-3 minutes.
As a thank you, we will send you a cheat sheet with 40+ free ML & data science books and courses! We appreciate your help.
🤖 Cool AI Tech Releases
MoViNets
TensorFlow open-sourced MoViNets, a collection of mobile optimized video classification models →read more on the TensorFlow blog
💸 Money in AI
ML&AI
Metadata store for MLOps Neptune.ai raised an $8 million in a Series A funding round led by Almaz Capital. Great job opportunities in Warsaw/Poland.
Open-source workflow orchestration platform Union.ai (Flute) raised $10 million in seed funding led by NEA. Hiring in Seattle/US, or remote.
AI platform for enterprises Noogata raised $16 million in Series A funding led by Eight Roads. Hiring in Tel Aviv/Israel and New York/US.
AI-powered:
Intelligent workforce platform Observe.AI raised $125 million as part of a Series C funding round led by SoftBank. Hiring in Bangalore/India and across the US.
Decision platform BlueOcean raised a $30 million Series B funding round led by Insight Partners. Hiring remote across the US.
Travel and expense management platform ITILITE raised a $29 Million Series C round led by Tiger Global and Dharana Capital. Hiring in Bangalore/India.