🚩 Machine Learning’s Last Mile Problem
Weekly news digest curated by the industry insiders
📝 Editorial
The MLOps space is thriving with innovations and continues to get incredibly crowded. Almost every week, we hear about MLOps startup raising sizable financial rounds while, at the same time, incumbents like Amazon, Microsoft, and Google are also innovating in the space. The result is that areas such as feature stores, ML monitoring, and data labeling have become incredibly competitive. One area of MLOps that hasn’t gotten the same level of attention is model serving and deployment. Every trend in enterprise software, such as cloud computing or big data, has created a generation of companies that solve the deployment aspects of that technology. ML deployment solutions are still nascent compared to the magnitude of the challenge. Many experts refer to this as ML’s last-mile problem.
Deploying a large number of ML models efficiently in ways that optimize computation costs is incredibly cumbersome. The default standard in the industry has become container platforms and ML serving frameworks in stacks like TensorFlow or PyTorch, but that remains quite limited. Problems such as testing against batch or streaming data, analysis of computation costs or deployment across large compute clusters often require orchestrating different technology stacks in very fragile ways. This problem is well-known in MLOps, and only a handful amount of startups innovate in this area. A few days ago, New York’s ML startup Wallaroo announced a new funding round to tackle precisely this problem. Wallaroo has been around for a few years and has built a fairly complete stack, but model deployment is definitely the area their offering excels at. If ML deployment follows the same pattern as other areas of MLOps, we are likely to see more funding and innovation flowing into the space.
🔺🔻 TheSequence Scope is our Sunday free digest. To receive high-quality educational content about the most relevant concepts, research papers, and developments in the ML world every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻
🗓 Next week in TheSequence Edge:
Edge#165: we discuss AutoRegressive Networks; +DeepMind’s PixelRNN and PixelCNN, two of the most important autoregressive models for image generation; +MMGeneration, a new toolkit for simplifying the implementation of generative models.
Edge#166: we deep dive into DeepMind’s new super model: AlphaCode can generate programming code at a competitive level.
Now, let’s review the most important developments in the AI industry this week
🔎 ML Research
Hierarchical Transformers
Google Research published a paper detailing a hierarchical structure to improve transformer models for computer vision →read more on Google Research blog
Hardware Optimized NAS
Google Research published a paper unveiling a neural architecture search (NAS) method able to adapt the produced models to a target hardware architecture →read more on Google Research blog
QMO
IBM Research published a paper unveiling Query-based Molecular Optimization (QMO), a framework that uses generative models to find the best variants of molecules →read more on IBM Research blog
🤖 Cool AI Tech Releases
iModels
Berkeley AI Research (BAIR) released imodels, a Python toolkit with implementations of many state-of-the-art interpretable modeling techniques →read more on BAIR blog
TensorFlow Runtime (TFRT)
The TensorFlow team provided an update in their ambitious TFRT project that tackles areas such as deployment, execution and experimentation →read more on TensorFlow blog
🛠 Real World ML
DeepETA
Uber unveiled some details about DeepETA, the architecture used to predict arrival times in the popular transportation app →read more on Uber blog
Improving TF-GAN
The TensorFlow team provides details about a small project used to improve the popular TF-GAN framework →read more on TensorFlow blog
💸 Money in AI
ML&AI&Data
Superconductive, a company behind the open-source tool for data quality Great Expectations, raised a $40 million Series B funding round led by Tiger Global. Hiring remote.
AI model management platform Wallaroo raised $25 million in a Series A round led by M12, Microsoft’s venture arm. Hiring in New York/US.
MLOps startup Qwak raised $15 million in a funding round co-led by Leaders Fund and StageOne Ventures. Hiring in Tel Aviv/Israel.
Cloud agnostic database ApertureData raised a $3 million seed round led by Root Ventures. Hiring in Bay Area/US.
AI-enhanced
Smart data capture platform Scandit raised $150 million in a series D funding round led by Warburg Pincus. Hiring globally.
Micropsi Industries raised $30 million in Series B funding round co-led by Metaplanet, VSquared and Ahren Innovation Capital. Hiring in Berlin/Germany. INTERN positions available.
Logistics startup Leaf Logistics raised $37 million in a Series B round led by Sozo Ventures. Hiring in New York, Chicago, remote/US.
Automotive simulation platform Morai raised $20.8 million in a Series B funding round led by Korea Investment Partners. Hiring in Seoul/S.Korea.
Entertainment localization startup Deepdub raised $20 million in Series A funding led by Insight Partners. Hiring in Tel Aviv/Israel.
No-Code Decision Intelligence solution Cerebra raised $15 million in a Series A round led by Notion Capital. Hiring in San Francisco Bay Area/US.