The demand for improvements in autonomous technology is accelerating. Memory and processing power has continued to grow exponentially cheaper, but the volume of data to process has exploded making it nearly impossible for traditional data analysis techniques to provide timely and cost effective guidance. Whether the data arrives in real time, or is acquired and stored for later analysis, the need for systems that can generate valuable insights from mountains of raw information will only increase. Consequently, neural network technology is increasingly valuable to organizations large and small making it a prime target for intellectual property protection. However, a number of misconceptions have arisen regarding the patentability of neural network technology.
Software isn’t patentable, so neural nets aren’t either: Software inventions are patentable, but the Patent Office and the Courts have narrowed the scope of what is patentable by requiring that the patent claims must be directed to something more than a well-known or abstract concept implemented on a computer. This is especially interesting where neural networks are involved because in some cases, the network itself is not new. The network topology (i.e. number of nodes, number of layers, the connections between them, etc.) may not be new, and perhaps the activation functions or backpropagation techniques used by the network are also not new. The training data sets or data preprocessing techniques may not be new either, but the outcome of using such a neural network may be truly revolutionary. Thus it is important to plan ahead in the drafting process to include aspects like tangible data sources and physical sensor input/output, control of physical objects or machines, and information about what technical problems are being overcome and how.
I didn’t invent a neural network so this is probably not patentable: Keeping the right focus on the invention is a fundamental issue that sometimes hampers patentability for software inventions, and it can be particularly problematic where neural networks are involved. Is the “magic” in the neural network or is the magic in how that network is used? For example, is the concept a new topology for a neural net that is more efficient, yields better results, or solves a particular problem? Is it a new activation function, or a new type of backpropagation scheme? Is the invention a new gradient descent algorithm that is optimized for a particular problem space? If any of these are the case, then the claims and disclosure should focus more on the network itself and how it is configured. On the other hand, is the invention a system that works better because it uses a neural network? If the invention is an improvement on neural networks, then more details about the network itself will be needed to show the technical problem and solution. If the invention is an improvement in some other field of endeavor that happens to involve a neural network, then more information about the inputs, outputs, and operation of the device will be needed, and perhaps less information about the neural network itself. Determining what the invention is will drive what kind of disclosure is needed in order to obtain a patent.
I’ll file the application but keep the real invention secret: The patent system grants the right to stop others from making, using, or selling patented inventions. In return, inventors are required to teach the world how to make the invention. In the case of neural networks, this can be tricky because many of the details about how a neural net reaches a given result are unknown until it is put to use, or they may be different from one execution to the next, or in some cases they are simply unknowable without extreme effort. In some cases, billions of permutations of inputs, outputs, and the corresponding weights for each node in the network could exist, but only after the network is put to use. That said, trying to patent a concept while keeping it secret is not permitted. The invention must be disclosed in such a way that a person of ordinary skill in the field could make and use the invention. The claims may be allowed to describe inputs and outputs at a high level, but at least some explanation is required as to how the system uses them and how they interact with other components of the system. Usually, more disclosure is better than less because failure to adequately explain the workings of a neural network, or the use of it, may cost both the opportunity to patent the concept, and the opportunity to protect the invention some other way, such as by trying to keep it a trade secret. With a little careful planning, both of these negative outcomes can usually be avoided.