Feedback Driven Learning at NeuralSpace
If you train NLP models, in fact, any AI models, from scratch, you have definitely asked yourself this question:
“What is the most efficient strategy to collect data and train models, so they produce the best possible results for your unique use case?”
Even if you had access to a dataset, how would you maintain it over time and make sure it gives the best results when circumstances change?
For software development, there are frameworks like Test Driven Development, Continuous Integration, and Continuous Deployment, which all guarantee agility and quality over the entire software lifecycle.
| But what about NLP models?
User requirements for software applications change and you can add new features that adapt to these changes. For NLP models, the language how people speak, the words they use, especially in an everyday language, evolves, and you need new data to adapt to these changes.
In this article, we will talk about Feedback Driven Learning: A framework for continuous development, testing, and deployment of NLP models.
There are three core principles in Feedback Driven Learning:
Here’s a short three-step guide about how you can deliver your best possible models over their entire lifecycle.
Build a minimum viable model with a small amount of data.
Exactly how an entrepreneur would build a minimum viable product, build a minimum viable model (MVM).
So what do we mean by MVM? For example: If you are starting from scratch and you have 5 classes, start by adding 60 examples: 10 training examples and two test examples for each class.
Train your model, it might not be perfect, but still, put it in production. Use AutoNLP and AutoMLOps for the training and deployment of your model in a few clicks.
Give data annotators, manual testers, and, of course, real users access to your new model. Use NeuralSpace’s Feedback Studio to collect all of these test data and carefully observe how they use your NLP model in the overall product.
This will also test the whole workflow of your app and if something fails you will know immediately. Once data are collected, take a few hours and correct some of the test sentences that were classified wrongly. Then, you add them back to your training dataset.
You can probably guess what comes now: you retrain your model on the new and old data with AutoNLP, deploy it with AutoMLOps and collect all the users’ input data again.
You can also choose to AB test your new model by using a “fire and forget mechanism”. This means, your old model, which is in production, is not interrupted but some of the data that are sent to the old model are also sent to the new model.
In this way, you can monitor the performance of your new model continuously and directly compare it to the performance of your old model.
The NeuralSpace Platform is live, test and try it out by yourself! Early sign-ups get $500 worth of credits — what are you waiting for?
Join the NeuralSpace Slack Community to connect with us. Also, receive updates and discuss topics in NLP for low-resource languages with fellow developers and researchers.
Check out our Documentation to read more about the NeuralSpace Platform and its different services.