Roch J. Ripley: Hi, everybody, my name is Roch Ripley. I am an intellectual property lawyer and partner in the Vancouver office of Gowling WLG. And I'm here today to talk about protecting AI based inventions with intellectual property. We find it a little ironic, I'm presenting this from my home office, where because of the COVID 19 pandemic, I imagine most of you at this present moment are confined to home. So it's not a renegade AI that has brought civilization to its knees like the Terminator movies told me what happened. Rather, it is a little virus that is neither artificial nor intelligent, that has died. Nonetheless, this too shall pass. And when it does, AI as a foundational technology will continue to be very important in the future, and hence the presentation today. So first, what I would like to talk about is, what do I mean by AI? AI in and of itself is a very general and fuzzy term. And I think we're talking about intellectual property patent protection, it really helps us to get a little more granular details and specific. I'm going to talk once we've defined exactly what we're talking about the various IP rights that apply to protect AI based systems, a particular copyright trade secret and patents, and then talk about some patents specific considerations. So for the purpose of this presentation, I'm going to define AI as machine learning. And by machine learning, I mean, you have an AI system that may be implemented using, for example, in their own network. And you've got some training data, which has been classified into various categories. So for example, you may have training data in the form of various images, the images may show different kinds of vehicles, cars, trucks, bicycles, what have you. And to train your system, you show the system, we image and you tell it, this is an image of a car, or this is an image of a bicycle, then when you get new analogous data that has not been classified, so for example, you may have a bunch of photographs that you know, show various vehicles, but no one has identified which particular vehicles, the various new photographs show, you then input that those new photographs to the system. And those new photographs are classified by the AI system based on our training data. And the output is a bunch of classified new photos. So you will know which of those new photos show a car versus a bicycle versus a truck. There are a lot of different models that can be used for machine learning, I just mentioned neural networks. Even within neural networks, there are lots of subtypes of neural networks. So for image processing, convolutional neural networks, or cn ns are very popular. For natural language processing, RN ends or recurrent neural networks are very popular. You don't need to use neural networks, decision trees support vector machines. And similar models can also be used to implement functionality we talked about.
The first of the IP rights I want to talk about is copyright. Now in the context of an AI based system, copyright, generally speaking is going to be used to protect the source code and the object code that you use to implement your system. Source code, of course, is what you pay employees or contractors create for you in ... human readable form, the object code like results after its compilation. Copyright is relatively limited in that it protects people from unsurprisingly, copying that source coding object code. So for example, if you distribute the object code to customers, and the customers then copy or pirate that object code and distribute it, that's a violation of copyright. Similarly, if you have an employee who without permission, copies your source code repository and takes it to a third party, that is something that's also protected by copyright, but it doesn't protect functionality. So if you create a wonderful AI based system and make a lot of money off it, and then a competitor sees that system and says, I want to copy what that does, because it's so great if they independently recreate that functionality. Copyright doesn't help you. In terms of ownership, previous slide was in terms of ownership. Copyright is automatically owned by a company or employer if the employee created in the course of his or her employment, but if it's created by an arm's length contractor, you need something explicit in the contract that says the company your company owns it. Moving on to trade secret trade secrets is a type of legal protection that unsurprisingly applies when you keep information secret and take reasonable steps to do so. So for example, we just talked about source code object code. If you store the source code and object code in a hardened IT system, you have a proper contractual arrangements in place to make it so that any employee who comes on board, anyone who joins your organization, even as a contractor is obliged to keep everything you're exposed to secret, then that's the infrastructure and this information can fall away. That is enough usually, to convince a court to grant you trade secret protection. So someone misappropriate your source code like an employee I just talked about who runs with third party, then in addition to a violation of copyright, you have a cause of action for misappropriation of trade secrets. Now, this is not going to work for all commercialization models, it may work for example, if you've got a SaaS based implementation, and you're keeping the source code and object code secret, and you only provide access to it through customers who log into a server, and can interact with the system's functionality, and wouldn't work. For example, you've got a system where you've got the object code and you're distributing it to third parties. Because you're distributing the object code, you can at the same time also keep it secret. So if someone were able to lawfully gain access to that program, by purchasing it, it's not a secret anymore, and you can rely on that protection. Now, one important part of an AI based system is the training data. And training data is something that you often can and want to keep secret. Now that's important because like I said earlier, the training data is something you need to make or train the AI based system for future use. But it's not something you necessarily need to distribute to customers to commercialize the system. And that's important because copyright may not apply to protect training data. Generally speaking, you need to exercise some level of skill and judgment or originality in order for copyrights to invest in something, and then may not be the case for training data, which may be mechanically created. Now moving on to patents, there has been a lot of patenting activity. Recently in the AI space, we can see some quotes here from articles that are relatively recent companies that despite pledging openness, have rushed to patent AI tech. And a comment there that is for AI technology, or patent applications for AI technology has skyrocketed. If you look at this graph, that is coming up now. You can see the USPTO from 2014 to 2017. There has been a lot as the author comments on exponential rise who can find the green curve there is for vehicles and navigation. So for example, that is an autonomous driving, you can see the orange curve below that, that is image analysis. And of course, there's a lot of AI that goes into image processing. Now that data is current to 2017. But I can confirm that the trends are continuing if not accelerating.
Now in terms of patents, you've got a few criteria that you need to satisfy the criteria that is common to all inventions, the invention, your AI based system is new, so no one else could have done the same thing has to be invented. So if no one else has done the same thing, what you've done has to be sufficiently different from what's coming for to justify being granted a patent. Now that, of course is a little bit fuzzy, which is where a lot of the argument examiner's and sometimes courts takes place. It can be a fundamental change to an underlying type of AI model, for example, perhaps you did create a new kind of neural network. But it doesn't have to be most of the time. It's not it can be for example, for a new application or an existing model or network. It can be for iterative or incremental improvement to an existing application. Though those incremental improvements can satisfy the inventiveness threshold I just discussed. Now, something that is particular generally speaking, in this space, the software based inventions and certainly it applies to AI based inventions is the concern of what's called subject matter eligibility. So even if your AI system is new, even if it's invented, is it simply the right stuff for you to get a patent application? Or excuse me to get a patent in? Is it too abstract? Is it too mathematical to get a patent, you can't get a patent just for an abstract way of thinking, or for a mathematical formula or for a general concept, it's got to be sufficiently tangible and concrete to justify or to be able to convince an examiner in the court that you should, in fact, be able to get a patent. Now, generally speaking, in order to satisfy this threshold, you want to show that your AI based system has some kind of technical benefit over previous systems, so what do I need to that, but what do I mean by that you can include in your application data, showing how your AI based system is superior in systems that came before. So for example, you may have an incremental improvement in the way you train an image processing system. Perhaps instead of using only positive training data, you also use negative training data for perhaps instead of just using one kind of image to train your network using multiple kinds of images with different classifications there you found that decreases the false negative rate and or the false positive rate. It's data like that you can include in your application to show some kind of technical benefit. And that helps you not only with subject matter eligibility, but it also helps you show that your system is in fact invented. Because you're not just changing things for no reason at all, you've got to change it actually results in some kind of practical, different. Different patent offices have published different guidelines that are specific to machine learning, and AI. So for example, in Europe, you can look up the guidelines for examination chapter 3.3, point one on artificial intelligence and machine learning. In the US, they've got October 2019 guidance, where they published examples showing how that guidance applies. In this particular example, 39 is how a method for training a neural network for facial protection can actually be found eligible.
So to conclude, copyright patent and trade secret can be used to protect the system, all the way can be used, and they can work in ways that complement each other. So for example, you can patent a system and at the same time, you don't have to disclose the training data you use to train the system. So you want to be able to describe how you're using these IP rights in a coherent way to anybody who may have an interest in investment or purchasing the product. patent filings as we showed in the graph are increasing exponentially. So certainly patents are an important part in protecting the systems. And including data supporting a technical effect in the application is going to help you in a lot of different jurisdictions despite the fact there are differences in patent law between those jurisdictions to help you with establishing subject matter eligibility and establishing that your system is inventive. So that is the presentation. If you have any questions or feedback, please feel free to contact me using that information on the screen. Thank you