This article was originally published by our ayfie CTO Johannes Stiehler in Law Technology Today on October 2nd, 2018.
Artificial intelligence is one of today’s hottest buzzwords. As the technology becomes more integrated into our everyday lives, it will become as essential in casework. A basic understanding of AI and its various forms will prove helpful when building a fact-pattern and a strong legal case.
Much of the hype around artificial intelligence (AI) has centered on one kind of technology: deep learning. But AI actually comes in a variety of forms, and many AI-based technologies tend to rely on more than one type. For attorneys, understanding what type of AI is in question, and how that technology functions, can be the linchpin to a legal dispute.
Different products utilize different forms of AI—this is dependent on the technology and how it is used. Oftentimes, one product relies on a number of forms of AI. A basic breakdown of AI and its various use-cases are outlined below:
Deep learning is a special form of machine learning, (i.e., data representation learning). This is just computer speak for “a program that initially knows nothing.” The program in question is a recurrent neural network. These algorithms can achieve near-human performance and, in some cases, even surpass it. But in order for a deep learning algorithm to function, it needs massive amounts of data.
This required amount of data has limited deep learning applications. So while there may be a lot of buzz around deep learning, it’s very unlikely that you’ll interact with a technology that leverages deep learning any time soon.
For the day that encounter does happen, understanding the functionality of deep learning is essential to building a case. Unlike other forms of AI, deep learning is mostly a black box. There is no specific “reason” for an action. So, for example, if a self-driving car was based on pure deep learning and it caused a fatal car crash, it would be impossible to identify an exact “who did what when.”
Rules may sound boring, but this is the most direct way to make a machine think like a human. There is a multitude of business processes on the market that depend on rules, such as many expert systems, important modules of what is known as “digital assistants” and some forms of document categorization.
Rules exist in many different applications and on many different levels. This form of AI can also reach very high levels of complexity by applying rules iteratively. This means the output of one rule can serve as the input to another.
If something goes wrong, it is usually easy to find the rule that triggered the problem. If a self-driving car was based on rules (as most are today) and it caused a fatal crash, it would be fairly easy to identify the rule that caused that error. The “who did what when” is a question more easily answered in this scenario.
Many AI technologies rely on some sort of rule-based system. Many aspects of “digital assistants” and automated customer service bots are all examples of rule-based technology.
Supervised Machine Learning
Classical (supervised) machine learning (as opposed to deep learning) requires significantly less amount of training data, which is why it’s a more commonly used form of AI. While deep learning can process unstructured input, a lot of thought needs to go into preparing a classic machine learning model.
For a moderately modern learning algorithm, such as support vector machines or logistic regression, y a couple of dozen positive and negative examples are needed to build a model. A human must answer the question: What representation should the algorithm now train on? This decision is an important component of how this sort of AI functions.
There is a lot of work that goes into this “feature extraction” process, which is often language-dependent, to make machine learning function. Most machine learning algorithms are based on the idea of identifying commonalities in the training documents and compiling those into a “model.” These models are the result of complex mathematical transformation and therefore hard to untangle. For many modern approaches—such as support vector machines—it is not easy to identify which training document contributed to the outcome in which way. Hence, a machine learning model can usually neither be inspected nor corrected directly, except by feeding more training input. Whether this has the desired effect depends on many factors that the user cannot control. The feature engineering and the training process are more transparent as compared to deep learning, so “classical” machine learning algorithms are still less black box than deep learning approaches.
As AI technology evolves, the products we interact with that utilize this technology will change as well. And the more intertwined it becomes in our lives, the more likely it is to show up in legal disputes. A basic understanding of AI, it’s various forms and how they function, will be a helpful knowledge base for the future work of attorneys.
Foto credit header image: bakhtiarzein via Fotolia