Weak supervision, active learning, and transfer learning work together in a data-centric workflow rather than acting as separate solutions.
2
Foundation models still need adaptation for specialist enterprise tasks, where accuracy, custom data, and governance matter.
3
Large foundation models can help create smaller production models that are cheaper, easier to govern, and better suited to a specific task.
Summary
Alex Ratner explains why Snorkel AI focuses on data-centric development instead of treating model architecture as the main source of progress. He distinguishes active learning, transfer learning, and weak supervision, then describes how they can be combined to decide where to label, how to label it, and how to get a useful starting point. Foundation models fit into this workflow as powerful but imperfect labeling and initialization tools. Ratner says enterprise adoption faces two gaps: adapting general models to specialist, high-accuracy tasks, and deploying them within cost, latency, and governance limits. His proposed approach uses foundation models to label data and help train smaller models for production. He also argues that real progress comes from working directly on customer problems, where data quality and labeling often block deployment, rather than only optimizing benchmark scores.
Data-centric development changes what machine learning teams improve
Alex Ratner contrasts model-centric development with data-centric AI. In the model-centric approach, teams iterate on features, algorithms, or model architectures while treating the data as someone else's responsibility. In the data-centric approach, the model class can be fixed or selected from a standard set, while the team spends its effort labeling, shaping, debugging, and improving the training data. Ratner says this shift came from seeing real projects spend most of their time labeling and relabeling data, checking model errors, and working with subject matter experts. The experts often explained why they assigned labels, but traditional supervised learning discarded that reasoning and retained only the data point and label.
Active learning, transfer learning, and weak supervision solve different parts of the labeling problem
Ratner describes the three methods as complementary tools. Active learning helps decide where to look next when developing or labeling data. Transfer learning reuses a prior dataset or model to create a stronger starting point instead of beginning from scratch. Weak supervision makes labeling more efficient by using labeling functions, heuristics, patterns, or other programmatic signals rather than clicking on every data point. He warns that none of these methods is a magic solution. Active learning can be difficult because knowing where the model needs more signal may be as hard as training the model itself. Snorkel's workflow combines the methods across repeated iterations.
Labeling functions preserve domain knowledge that manual labels usually discard
A labeling function lets a domain expert express a rule or heuristic for labeling data. Ratner traces this idea to work with genomicists and other subject matter experts who explained patterns while labeling examples. Traditional training data captured only the final label, so details such as negation, key phrases, patterns, or external references were lost. Programmatic labeling keeps more of that knowledge and can make labeling 10 to 100 times more efficient by accepting noisier signals and cleaning them algorithmically. The same approach can be used beyond text. For images, labeling functions can operate on metadata, feature or object detectors, and eventually natural-language prompts applied to image or multimodal foundation models.
Foundation models have an adaptation gap between general generation and specialist prediction
Ratner says foundation models are impressive for creative, human-in-the-loop work, where a person can select or edit an output. They are different from automated enterprise systems that must meet a high accuracy bar on custom data. He gives his own stable diffusion example: he tried about 30 samples before getting a useful Snorkel logo image, which is acceptable in a creative workflow but would be poor performance for production automation. General models trained on web text also do not automatically understand contracts, fraud data, medical material, or other specialist domains. Prompting and zero-shot methods can help, but difficult, high-value problems usually still require fine-tuning on labeled training data.
Enterprise deployment has cost, latency, and governance constraints
The second gap is deployment. Ratner says many enterprises cannot easily serve large foundation models because of their cost and latency, and because they need to understand and govern the models before exposing them to customers. He describes a large US bank that viewed submitting GPT-3 to its model risk management committee as almost impossible. Sensitive private data creates another barrier, since many enterprise problems involve custom, on-premises information that cannot simply be sent to a general model. Ratner's answer is to use foundation models where they help with development, then produce a smaller specialist model that fits the organization's production and governance requirements.
Large foundation models can help train smaller models for production
Ratner describes a workflow in which a foundation model helps label data, while a smaller model is trained for deployment. In one contract classification example, Snorkel's approach used less than one percent of the ground-truth labeled data and trained a model that was over a thousand times smaller and hundreds of times cheaper to run than the fine-tuned GPT-3 comparison. The foundation model is useful as a generalist source of information, while the deployed model handles one custom task on specialist data. Ratner presents this as a way to meet higher accuracy, cost, repeatability, and governance requirements without serving the largest model directly.
Prompts can be combined with other labeling signals
Ratner explains that Snorkel treats prompts as another form of labeling function. A team can use a foundation model for a zero-shot warm start, write domain-specific prompts, add ordinary programmatic labeling functions, and combine the resulting signals into a training set. He gives an anti-money-laundering and know-your-customer example in which a bank needed to extract information from large collections of customer documents. Zero-shot foundation-model labeling handled some simple fields and provided a useful starting point, but it did not reach the required accuracy by itself. The broader weak-supervision system combines imperfect signals so developers do not have to choose a single model or prompt in isolation.
Real production problems changed Ratner's research priorities
Ratner says his work with Christopher Ré's group taught him to balance formal research with direct work on real use cases. The Snorkel team initially expected to work on advanced modeling techniques, but customers repeatedly said that labeling data was what blocked them. That experience led them toward data-centric AI. He argues that teams should learn the fundamentals and develop abstractions, while also spending enough time with clinicians, genomicists, and other users to see a real problem through. Benchmark scores and creative demos can hide the practical work required to build a useful system.
"Foundation models are just another tool in the toolkit and immensely powerful one, but also an imperfect and tricky one that you have to tune an engineer and is not going to work magic out of the box on your specific problem."Alex Ratner43:37
Who should watch
You are deciding whether foundation models can support a high-accuracy enterprise workflow rather than a creative demo.
Your team spends more time labeling and debugging training data than changing model architectures.
You need a way to use a large foundation model during development while deploying a smaller model under cost, latency, privacy, or governance limits.
