Active learning: How to accelerate AI model training
Active learning: How to accelerate AI model training
Let’s face it: training an AI model isn’t much fun.
Imagine you were training an intern who’d just joined your company. This intern is extremely bright and can work at lightning speed for 24 hours a day without any breaks. This should be a dream come true, right? Except, they don’t know a thing about your business. They can’t tell the difference between a ‘thank you’ email and an urgent customer complaint. They have zero common sense and get even the most basic things wrong.
Anyone who’s ever started training an AI model for their business can probably relate to this example. The good news is that AI can be trained to understand your business accurately and carry out your most important processes. But it takes time and effort, and, usually, lots of data annotation.
Let’s face it: training an AI model isn’t much fun.
Imagine you were training an intern who’d just joined your company. This intern is extremely bright and can work at lightning speed for 24 hours a day without any breaks. This should be a dream come true, right? Except, they don’t know a thing about your business. They can’t tell the difference between a ‘thank you’ email and an urgent customer complaint. They have zero common sense and get even the most basic things wrong.
Anyone who’s ever started training an AI model for their business can probably relate to this example. The good news is that AI can be trained to understand your business accurately and carry out your most important processes. But it takes time and effort, and, usually, lots of data annotation.
The data annotation bottleneck
Data annotation, also called data labeling, is the manual process of tagging raw data with relevant classifiers or ‘labels’. In business, it’s a crucial part of the process for training AI models to accurately recognize and correctly respond to patterns in your data. For example, helping a model tell the difference between a ‘thank you’ email and an urgent complaint. Or helping it correctly extract important data from a message, like a delivery address or customer number, which can be crucial for many valuable automations.
You could say that annotation has become the new programming. Increasingly, instead of programming what we want machines to do, we label examples for them to replicate. But that doesn’t stop it being a long and boring process for the people who do it!
Data annotation consumes about 80% of the time dedicated to any AI project. Subject matter experts (SMEs), often involving teams of employees, will typically spend hundreds of collective hours labeling thousands of individual examples. But add human error to the mix. Some labels will inevitably be wrong, impacting the AI’s understanding of the data and probably requiring even more time to repair the damage.
Active learning: better AI in less time and lower costs
Supervised learning produces models that act in a more consistent and reliable way. It’s the only kind of model that should be used in a real business environment without supervision. Supervised learning is key to building Specialized AI models, which are built to understand and carry out a specific task. Yet the data annotation bottleneck means these models are slower to train and deploy than those made with unsupervised learning.
But what if we could combine the accuracy of supervised learning with the speed of unsupervised learning?
Active learning is a mature training method for AI, but it’s only recently been used to train enterprise AI models. It combines elements of both supervised and unsupervised learning to create better AI models in less time.
Like supervised learning, active learning requires annotated samples for model training. However, instead of simply learning from a dataset, the model makes unsupervised decisions about what it wants to learn next.
It then actively queries the SME but, crucially, only asks them to annotate examples it’s most unsure of or thinks will be most useful for its training. Just like in unsupervised learning, the model identifies patterns itself and decides what information it needs to learn better.
Active learning helps create an intelligent workflow for annotation. Remember the AI intern from our example at the beginning of this blog post? With active learning they could complete most of the training by themselves, deciding what they should learn next and only asking for assistance when they get stuck. Active learning is much closer to the learning patterns of humans, and it means a lot less handholding and work from the SME.
What makes active learning so useful for businesses struggling to train their own AI? You need much fewer annotated examples to train a model from beginning to end. The AI does the heavy lifting in terms of training and will work collaboratively with your SMEs to improve its understanding—both as you build the model and later while you consume and refine it.
AI models built with active learning can be trained faster, with fewer labeled examples, and without sacrificing accuracy or performance. Another advantage of active learning is that it leaves fewer opportunities for human error and bias to creep in. That’s why it’s the ideal method to help businesses train Specialized AI models that are reliable and get to work quickly.
Putting AI to work—faster
What’s the secret ingredient for AI success? Is it the models you use? Or how many data scientists and SMEs you hire to train them?
What really separates the AI leaders from the laggards is how fast they can ‘operationalize’ the technology. How quickly they can deploy AI in their business and how fast it starts creating value for them. Traditionally, this has been a real challenge for intelligent document processing (IDP). Training AI models to reliably understand and process documents and messages usually demands a big investment in time and effort.
That’s why UiPath uses active learning to accelerate time to value for customers using our leading AI capabilities for IDP.
UiPath Document Understanding and Communications Mining (both available via the UiPath Platform) enable users to quickly build custom AI models that can understand and automate documents and business communications. Thanks to active learning, these UiPath Platform capabilities start training with only a few annotated examples. SMEs and AI then work together to finetune model understanding by labeling only the most informative and valuable examples.
Our active learning approach—combined with the no-code, fully-guided user interface of the UiPath Platform—produces accurate, high-performing AI models in hours rather than weeks or months. For instance, the introduction of active learning in UiPath Document Understanding has seen 80% faster model training according to our internal tests. Models that used to take a week to train now need only a day before they’re ready for business.
Summary
In business and in life, time is the most precious thing we have. And, right now, data annotation is taking too much of it. Stretching out time to value and putting pressure on our people. Fortunately, active learning offers a better approach. By using both supervised and unsupervised methods, active learning cuts down on data annotation to focus on only the most important examples.
Active learning drastically reduces the labeling effort needed to train and deploy accurate, high-performing AI that really understands your business. It means less labeling, happier employees, and faster time to value for AI.
UiPath is a pioneer in active learning for the enterprise, reducing time and effort while boosting the performance and accuracy of Specialized AI models. To learn more about the latest UiPath advancements in active learning, join the UiPath Insider Program and try out building a modern Document Understanding project using the preview capabilities in UiPath Automation Cloud.
Compiled By Pushpendra Maurya
Profession Data Scientist
Source:UI Path
Comments
Post a Comment