Last week, I wrote about 4 key factors significantly impacting IDP adoption, starting with the ability to work in suboptimal conditions. This week is a deep-dive into the second factor, which addresses how modern Intelligent Document Processing (IDP) software can learn and improve — keeping in mind that different solutions tackle the problem in different ways — some better than others.
Stepping back, let’s talk about learning in general. Learning always involves instruction in some form: teachers, books and articles that explain, videos that instruct and so on. In computing, just as in life, there is no learning without instruction. Even so, many terms are often used to describe the learning process with three used most commonly: supervised, unsupervised and reinforcement. With all of these, the objective is to replace the grinding, arduous, complex and expensive tasks of configuring, testing, optimizing and maintaining IDP automation.
Supervised Learning
Within supervised learning, the system can learn through direction by a user, either interactively (as with a teacher) or through some basic instructions and providing an answer key (as with a student learning from a text book). Supervised learning is the most common, mostly because the platforms that use this type of learning are generally easier to construct than others. But even within this category, there are different ways to apply it. For instance, some IDP solutions offer supervised learning capabilities solely as an interactive process where a person, generally a subject matter expert, works through different documents, noting required data while the system records the actions.
Other systems can work independently using only a textbook approach where labeled information is input, and the system analyzes the most reliable means to reproduce the same output. And underneath, systems use different AI techniques. Some essentially use a rules-based approach, recording specific actions so that they repeat them when the same data is encountered. Others prefer to focus on deep learning neural networks for any task. Others, as with Parascript, use a variety of techniques, each selected based on their strengths for specific automated tasks.
Reinforcement Learning
Reinforcement Learning is a still-emerging area where the focus is on creating learning systems where relatively little input data is available. Instead, the training element is the environment with which the system interacts.
A real world analog could be a cat where a cat learns from a system of rewards or punishments. If a cat uses a scratching post, it is rewarded with a treat which, over time, teaches the cat to repeat this positive behavior. If the cat scratches on a couch, it is punished and over time, it learns to not behave in that way.
A computing example that uses rewards and punishments is the Alpha Go system that has beaten the world’s best Go champions. It can initially be taught, using supervised learning, what strategies to employ and then benefits from reinforcement learning during actual games where it can analyze positive and negative outcomes of its strategies and even come up with more efficient, rewarding strategies. Reinforcement learning has no real practical place in IDP since automation tasks can be pre-defined as inputs and proper outcomes.
Unsupervised Learning
The last is unsupervised learning and the fact is, it is not really learning at all. Just like humans, the ability to learn without instruction or feedback is not possible. However, there are ways to make inferences, just as we can make observations and conclusions based upon no instruction or feedback.
As an example, without having to understand different document types, I can easily sort documents by likeness. I don’t even need to understand the language to get them mostly correct.
In computing the “unsupervised learning” algorithms such as clustering and regression operate in a similar way, noting patterns and making actions based upon those patterns. Unlike other forms of learning, the system does not get better or more efficient since there isn’t anything providing input to do so.
Discovery of Key Input Attributes
In all the true forms of learning techniques, whether through supervised or reinforcement, the underlying behavior is the discovery of key attributes of the inputs that provide the most reliable paths to producing the ideal, correct output.
In IDP, it’s all about using machine learning to discover document features, whether these are text, handwriting, structural or graphical that are key to performing a task. The ultimate benefit is that you spend more time working on the requirements in the form of good input data and more time examining the results. And much less time with all the pain and hassle in-between.