“Decisions acting at the level of individual patents, where tens or even hundreds of thousands of individual patents are potentially available as inputs, would be good candidates for the application of artificial intelligence.”
Contemporary AI technology of the kind one has increasingly heard about in recent years is based on machine learning and deep learning methodologies. These use large amounts of computing power to crunch thousands of sample input-output pairs to train adaptable data structure models. Eventually, they are able to produce their own correct outputs when presented with an nth + 1 input. These can be thought of as questions and answers. If an AI model is given, say, 10,000 sample questions with correct answers, it will be able to correctly answer the 10,001st question by itself. Once trained, computing requirements are low.
Due to the nature of the methodology, AI is appropriate for situations that involve repetitive decision-making processes. For one thing, many existing examples of correct decisions must be available during the training. Further, after the training phase, a system is applied to similar situations over and over again. Because of this, the application space for AI is sometimes overblown. However, once understood, this limitation usefully directs our attention to instances of decision-making that can be automated or made more efficient using AI.
Employing AI for Patent Portfolio Management
If we consider patent portfolio management in terms of constituent decision-making processes, we might be able to identify which of them are appropriate for the application of AI. Patent portfolio management typically includes tasks like full portfolio assessments, estimates of the potential value of individual patents, identification of weaker areas of technical coverage in a portfolio, identification of patents to cull, the evaluation of external portfolios, comparisons of such portfolios to internal portfolios, the decision to enforce IP with respect to external organizations, to acquire external portfolios, or to sell a portion of a portfolio, etc. Which of these processes are likely to benefit from AI?
Understanding the applicability of AI to repetitive processes in need of large amounts of previously existing sample inputs for training, we can narrow the field of action significantly. Strategic decisions acting at the level of portfolios as a whole, such as the decision to enforce IP with a given competitor, are poor candidates because they happen very rarely and are highly context-sensitive. Processes with low numbers of practically attainable input samples are in general difficult to automate. However, decisions acting at the level of individual patents, where tens or even hundreds of thousands of individual patents are potentially available as inputs, would be good candidates. Examples of this include which patents to cull, which patents have potentially high value, and automatically detecting the technology category of a patent. It should be remarked that properly applied automation at the level of the individual patent can ultimately also help decision-making at higher strategic levels, but by this time the decision-making is still made by human beings, albeit based on a richer data set provided by AI.
Training Samples and Subject Matter Experts
Another practical limitation impacts on the choice of applications: generation of training samples. The AI model in training requires thousands of appropriately formatted input samples. The simple existence of previously made decisions of a certain type is not enough. They must be available to the training team, properly vetted (i.e. considered to be correct), and then formatted for input. In the case of decisions made on patents, these are more or less always made by highly-trained technical staff. To generate samples into an AI for patent decision-making, one must have access to such technical experts during the sample-generation phase and have them dedicate their time to generating samples. In practice this means that Subject Matter Experts are asked to engage in an evaluation activity in the same way they would for a real-world project, but with targets selected to generate the best training samples possible.
Once enough decisions have been generated by human Subject Matter Experts, they can be used as a training set for an AI. After the compute-heavy training phase, the AI is then available to very rapidly answer further samples with no human intervention.
Through this process, decades of knowledge, as it exists in the minds of the Subject Matter Experts that are called upon during the training phase, are distilled into a computerized process which can then significantly accelerate decision-making in patent portfolio management.