Artificial intelligence and machine learning have been revolutionizing the field of e-discovery in recent years. Not all machine learning is created equal, however.
Transfer learning is a specific kind of machine learning that has the ability to significantly reduce the time, cost and labor of traditional e-discovery, and NimbleSystems was the first provider on the market to offer it. If you want to learn how to capitalize on these savings, read on.
What Is Transfer Learning?
One specific kind of machine learning that’s having a significant impact on e-discovery is called transfer learning. In transfer learning, the system gains knowledge from a learning task and can store that knowledge and apply it again to solve other learning tasks in the future.
Transfer learning works with two pools of data: source data and target data. Source data is made up of the training documents used in a prior task, while the target data is the training documents from the new task. It’s very difficult to apply machine learning directly to text, so transfer learning systems often first convert text into numbers and vectors that computers can easily process.
There are two main types of transfer learning: data-focused and model-focused. In data-focused transfer learning, training models are built from scratch using vectors from past tasks and new vectors from the current task. In model-focused transfer learning, pretrained models are altered and adapted using vectors from the new task, with no need for vectors from past tasks.
Text can be analyzed with both data-focused and model-focused transfer learning, while images are only treated with model-centric transfer learning. In fact, model-focused transfer learning is commonly used for image classification. NimbleSystems offers a number of image models to identify pictures, maps, handwriting and logos.
How NimbleSystems Is Bringing Transfer Learning to the Law
In the practice of law, transfer learning can be used to harness the work that’s been expended on a previous case and adapt that work to future cases. Specifically, NimbleSystems implements instance transfer learning, a data-focused form of transfer learning that adjusts the data source for each iteration of active learning and then trains a new model first on the source data and then on the target data.
By capturing past learning and applying it in a new matter, lawyers are able to reduce the amount of time, labor and overall expenses that typically go into training new models from scratch. This brings efficiency to e-discovery, which can traditionally be a very inefficient and costly endeavor.
Some attorneys, however, have concerns about reusing data from previous cases, worried that doing so might implicate confidentiality, privilege and more. The methods underlying transfer learning eliminate those concerns. Without getting too technical, during transfer learning, documents are stripped of identifying information like names, dates, addresses and email addresses before being encoded as mathematical vectors. The original documents are then no longer used. This eliminates any possibility that information or documents from a previous case could be recreated or accessed by attorneys in a later case.
Other e-discovery providers start every machine learning model from scratch, and all past machine learning is lost when a case is over. Firms should examine their cases and practice areas and decide if they want to provide source materials. Doing this will build up a library from material as diverse as environmental law, antitrust, commodities and financial instruments and defective drugs and medical devices. NimbleSystems has worked in all these areas and has experience turning coding from old cases into source models. The past legal work we’ve done in these areas can be encapsulated into the transfer learning models that we offer our customers.
E-discovery is about more than just reviewing documents – you should be gaining knowledge from those documents that you can use to better handle cases in the future. With transfer learning from NimbleSystems, lawyers can master new learning tasks much quicker than starting from scratch with other AI-based tools.
Contact us today to learn more.