In the first part of this series of articles, I argued that invention is inherently a creative act, and that since machine learning systems – however impressive or surprising their achievements – are incapable of human-like intelligence, reasoning, agency, or creativity, they therefore cannot ‘invent’. I acknowledged, however, that machines are certainly increasingly involved as ‘assistants’ in the process of invention. In the second part of the series I argued that where the result is a patentable invention there must always be at least one human inventor. In this final part, I want to look at how we should go about identifying the inventor(s) in any given case of machine-assisted invention. There are, in fact, three main aspects of machine learning technology in relation to which inventions may arise. The most rarefied of these is in the underlying machine learning algorithms and architectures themselves. Comparatively speaking, very few people work in this area, and for the most part they are to be found in universities and research centres.
Secondly, inventions may arise through the application of machine learning technology to solving problems and/or producing new results, products, and services. I expect that this is currently the most common type of machine-learning-related invention, particularly given the wide range of software tools now available to assist programmers in implementing the underlying algorithms.
Thirdly, there are ‘machine-assisted’ inventions, which are generated wholly or in-part by machine learning systems. Currently, such inventions are relatively rare, considering that few applications of machine learning are actually directed to the generation of new technologies. However, as machine learning increasingly finds its way into computer-aided design and engineering applications, this may change.
In all cases, however, I would argue that it is possible – and, indeed, necessary – to identify one or more human inventors (and no machine inventors). To suggest otherwise is, in my view, to misunderstand the true nature, and limitations, of machine learning systems. In a letter to shareholders, published in 12 April 2017, Amazon CEO Jeff Bezos wrote what is possibly the most succinct and jargon-free summary of what distinguishes machine learning systems from ‘traditional’ programming:
Over the past decades computers have broadly automated tasks that programmers could describe with clear rules and algorithms. Modern machine learning techniques now allow us to do the same for tasks where describing the precise rules is much harder.
The evolution from ‘traditional’ programming to machine learning is not as dramatic as some of the hype might lead us to believe. Instead of coding the ‘rules’, machine learning developers now build systems that are able to capture and generalise from patterns that exist in their input data. These systems still operate using rules and algorithms – but now these algorithms determine how they go about doing the capturing and the generalisation, rather than how they produce the final result. Viewed at an appropriate level of abstraction, then, little has changed, except for the power and scale of our machines.
In view of this, I argue that when it comes to machine-assisted inventions, inventorship will generally arise from successfully designing and applying a machine learning system on the path to achieving an inventive result, even in cases where the underlying software and/or hardware employed may have been developed or supplied by someone else.
Inventors of Machine Learning Technologies
How do you program a computer to play a board game, such as chess or Go? Traditionally, you might devise some kind of scoring system, so that for any game state you can compute which player is in a ‘stronger’ or ‘weaker’ position. By searching through the results of possible future moves, and calculating the projected scores, a computer program can try to find the current move that has the greatest prospect of strengthening its future position, and/or weakening its opponent’s position. For complex games, searching all possible future moves rapidly becomes infeasible within a reasonable time-frame. Furthermore, the performance of such a program is ultimately limited by the quality and flexibility of the scoring system. Place too high a premium on retaining the queen in chess, for example, and the computer player will never sacrifice its queen, even if doing so might put it on a path to victory.But is it really necessary to codify the ‘rules’ of strength and weakness in a mathematical formula devised by a human programmer? Just because the critical patterns in a single board position, or in a sequence of moves, are too complex to be easily summarised by an equation does not mean that they do not exist.
An alternative approach employed in many machine learning systems is to employ networks of simple calculation units to transform inputs (e.g. the current state of a game board) into outputs (e.g. a current move). Each calculation unit may have one or more parameters that can be varied. ‘Training’ is the process of finding values for all these parameters that (hopefully) will result in the most optimal output for any given input. A great deal of research within the field of machine learning is directed to devising and improving the architecture and operation of the calculation units, and to the algorithms used to optimise the parameters during the training process. This is one aspect of machine learning in which inventions may arise, and it is clear that the inventors in such cases are the researchers responsible for developing the improved machine learning technologies.
Inventors of Applied Machine Learning Systems
The number of people working with applications of machine learning is far greater than the number involved in developing and improving the underlying technologies. For anybody wishing to get involved in applying machine learning to the solution of problems, there are numerous software libraries and frameworks available that provide implementations of state-of-the-art techniques and algorithms in a range of programming languages (see, for example, this ‘Awesome Machine Learning’ list).The ‘MarI/O’ system configured to learn how to play the old Nintendo game Super Mario World (also discussed in Part 1 of this series) is a good example of such an application. As a reminder – or in case you did not already watch it – here again is the accessible and fun video demonstrating how the system works.
As the system’s creator, ‘SethBling’, explains in his narration, he did not develop the machine learning technologies underlying the MarI/O system, which are described in the research paper ‘Evolving Neural Networks through Augmenting Topologies’ [PDF, 445kB] by University of Texas researchers Kenneth Stanley and Risto Miikkulainen (in Evolutionary Computation, vol. 10 no. 2, 2002, pp 99-127). Nonetheless, the video illustrates how building practical and effective machine learning systems using such technologies has itself become a distinct discipline, requiring experience and creativity.
The nature of this creativity is plainly apparent in SethBling’s MarI/O demonstration. While the neural networks and genetic algorithms employed in the system were already well-known, the choice of these particular algorithms, the selection of parameters, the design of the input data (i.e. the simplified two-dimensional ‘view’ of the screen), and the formula for the performance measure (i.e. ‘fitness’) were all contributed by the human designer. Prior to commencing development of the system, none of these individual choices – let alone the specific combination of all of them – was necessarily ‘obvious’ or inevitable. Notably, the computer demonstrates no creativity of its own in learning to be an expert player of Super Mario World. It performs only as well in the learning task as SethBling’s design dictates.
A concrete machine learning system developed to address a particular need, or to solve a particular problem, may thus be a ‘combination’ in the same way that many mechanical systems are ‘combinations’, i.e. an assembly of known components into a novel configuration such that they work together to provide some new and useful function or result. It is well-established that such systems can be patentable, so long as they satisfy the substantive requirements of novelty and inventive step. It is equally clear that the inventors in such cases are those responsible for making the design decisions and/or conducting the necessary research and experimentation that resulted in the inventive combination.Inventors of Machine Assisted Inventions
The third category of invention relating to machine learning is the one I have called ‘machine assisted’ inventions, of which the proportional-integral-derivative (PID) controller that I described in Part 2 of this series of articles is an example. In this category, the invention is not in the underlying machine learning technologies, or the specific configuration of those technologies employed in the system, but rather the actual output of the system. In the case of the PID controller, the invention (described in US patent no. 6,847,851) took the form of new methods for calculating the parameters of the controller that resulted in improved performance. These methods – essentially being mathematical equations – were generated by a genetic-programming-based machine learning system developed by John R Koza and two co-inventors.From the perspective of patentability, the primary difference between a system such as SethBling’s MarI/O and Koza’s controller-generator is that, unlike a ‘solution’ to Level 1 of Super Mario World, a PID controller design is patent-eligible subject matter. In both cases, however, the only potentially inventive activity involved lies in the design of the machine learning system. Beyond that, the computer does nothing more than follow its programming to generate an ultimate ‘evolved’ output – either a sequence of moves resulting in successful completion of the level, or a set of equations that can be used to compute suitable parameters of a PID controller.
In both cases, the software could be handed over to someone else with no understanding of how it works, who would be able to use it to achieve the same, or similar, results. Clearly, this would not make the person who initiated execution of the software, or the owner of the computer on which it executes, an inventor of any patentable output that may result. Indeed, these people are completely unnecessary and irrelevant to the final result. The software is able to recognise improvements in its own outputs without further human intervention precisely because it is configured with an ‘objective’ (e.g. the ‘fitness’ function, in the case of MarI/O) that is used to evaluate its performance on each iteration.Accordingly, in the event that the output of such a system comprises a patentable invention, the inventors must be the system’s designers.
Conclusion – Creating Intellectual Property Requires an Intellect!
This series of articles began as a critique of recent suggestions by some people in the field of intellectual property that computers can be ‘creative’ or ‘inventive’ and that it may be necessary to consider the possibility that a computer could be named as an inventor on a patent application – or, conversely, that in some cases humans should be disentitled from inventorship on the basis that their computers, rather than themselves, were the true inventors. In an extreme case, Professor of Law and Health Sciences at the University of Surrey, Ryan Abbott, has argued not only that computers can be inventors, but that patents have already been granted on ‘computational inventions’ (‘I Think, Therefore I Invent: Creative Computers and the Future of Patent Law’, Boston College Law Review, Vol. 57, No. 4, 2016, available at SSRN).In my view, such claims arise from an inadequate understanding of how machine learning technologies actually work, and of their limitations. All existing machine learning systems are built to transform a discrete and limited number of input parameters into a corresponding discrete and limited number of output parameters via a specific set of computational functions. This is not a process of invention, even if the ultimate output may represent something that qualifies as a patentable invention. Such a result can only be achieved because that was the objective of the system’s designers. No credit can be attributed to the machine.
The term ‘intellectual property’ suggests, quite rightly, something that is intended to protect the products of the conscious mind. For every invention, therefore, a key question in identifying inventorship must be, ‘whose mind?’ There must always be at least one conscious mind involved, whether in devising, designing, directing, or discovering the invention. I have argued that, for every invention related to existing machine learning technologies, it is always possible to identify one or more minds, i.e. human inventors, to whom the invention can, and should, properly be ascribed. It is unnecessary, and wholly premature, to start attributing inventorship to machines. Perhaps the day will come when an analysis of inventorship, based upon a full and proper understanding of the technology involved in making an invention, fails to identify any responsible human mind. But that day has certainly not come yet, and I am not holding my breath!
Advanced computational tools will certainly continue to change the process of inventing. They will probably also change the nature of what we would regard as an invention – this would certainly not be the first time in history that technology made routine work of activities that would once have been regarded as inventive. However, machines do not invent. And while it is generally a bad idea to say ‘never’ when it comes to technology, I do not expect to see a ‘conscious’, and therefore potentially inventive, machine intelligence during my lifetime.
0 comments:
Post a Comment