TARCISIO QUEIROZ CERQUEIRA
ADVOGADO/ATTORNEY IN BRAZIL
LEGAL ADVISER ON SOFTWARE AFFAIRS
www.tarcisio.adv.br
tarcisio@tarcisio.adv.br
Systems, in terms of computer software, are defined as a group of computer programs working interactively, in order to perform certain tasks. This definition includes all kinds of systems.
"Artificial Intelligence" can be defined as an expression commonly used to designate those kinds of computer systems that display certain capabilities associated with human intelligence, such as perception, understanding, learning, reasoning and problem-solving.
Artificial Intelligence (AI) systems are special computer software systems which advise people, doing the same as a recognized human expert could do.
AI systems can also be defined as a group of problem-solving computer programs which solve substantial problems generally conceded as requiring expertise. They are called knowledge based because their performance depends critically on the use of a data base of information used by experts.
An AI system solves a problem by 'symbolic' processing. In AI systems a program called an "inference engine" operates upon information contained in the knowledge base to solve a particular problem. The inference engine contains the control logic which browse through the knowledge base in response to a particular query input by the user.
Conventional systems usually perform the standard formula "input-process-output" of information. Ordinary computing uses 'procedural' processing to solve a particular problem. A computer is given an algorithm, which is a step-by-step sequence of those steps that must be taken to solve a given problem. Conventionally programmed computers start at point A and proceed in a linear fashion to point B. Conventional systems receive one input ("start-information") which starts one or more programs and produces one or several outputs ("exit-information"). This "modus faciendi" has a close relationship with the rights and obligations of each part of the agreement: who had developed and who had ordered the development.
Ronald S. Laurie comments that there is a demarcation problem in AI systems, as computer scientists do not agree on their rigorous definition. Moreover, apart from the nature of artificial intelligence systems, "...the experts even differ as to whether particular kinds of software systems, e.g., speech or image recognition, are properly characterized as AI or not. This is due, at least in part, to the fact that the boundaries of AI are continually expanding and contracting in different directions to include new forms of 'intelligent' software as they emerge and to exclude others previously considered to be within the domain of AI as they mature".
Artificial intelligence systems are a huge, grey and elastic area, and the EC Software Directive does not refer to copyright protection as extending to them.
Some authors take the view that there is no single definition of Artificial Intelligence, but there are directions, or goals, rather than devices, or technologies, or theories, while others state that artificial intelligence systems have to be treated in a special manner because they generate knowledge.
However, we may add, all systems, including those known as ordinary, or conventional, generate knowledge - as knowledge can be considered merely information, and, in fact, expert systems are not so much characterized by the fact that they use knowledge, as HOW this knowledge can be used.
In both systems we identify natural persons who develop the system, companies or natural persons who own the systems and users, or customers, or clients, who use the system. There is also, optionally, the distributor or reseller, and most systems - whether conventional or AI - have their development divided into phases, and payments from one part to another usually happen at the end of each phase.
Conventional or AI systems are groups of computer programs, they may work in the same hardware platform, information can be represented in the same manner, and they are written, or produced, by computer programmers.
On the other hand, Artificial Intelligence systems are different from their counterpart conventional systems in several respects, such as in terms of their complexity, the technology used for their development and the larger and more heterogeneous group of experts required for their production, including those from areas not usually related to computers. Furthermore, responsibilities and rights of both sides of the agreement can be different, depending on the system.
There are three points to be considered regarding intellectual property law and artificial intelligence systems:
- An AI system is a group of computer programs, written in any computer language;
- There is information recorded in order to permit an AI system to work properly. This information, depending on the contract, belongs to the person who supplied it, or who made it available and gave permission for its use. When in doubt the contract of development of an AI system should show who owns which.
- Computer programs, in AI or conventional systems, are protected by copyright in most countries, independently of the media where they are recorded. No matter where the shell is fixed, whether on a disk or on a ROM-read only memory.
- On the other hand, the knowledge base, which is a data collection - on which some countries have established different requirements of originality - requires copyright protection.
In contrast, J.H. Spoor states that sometimes the knowledge base contains ideas, or requires work of creation in terms of organization of data and storage of information, and, in these cases ideas can be patented, depending on theis nature. Also, in AI systems computer programs (in the shell) can be implemented in semiconductor chip form, especially if a large number have to be produced or if performance requires it. "...In such cases, Semiconductor Chips Acts, which are now in force in a number of countries, will apply".
Of course, depending on the development of the system, which normally is regulated by contract, ownership might be attributed to one or another party. The knowledge base can be owned by the experts who have provided the knowledge, and the computer programs usually belong to the company or persons who have developed them.
Thomas Dreier defines protection of Artificial Intelligence Systems merely as a question of copyrightability, and also focuses it on three points: collection of data, facts and rules; algorithms; and computer-output. According to him, a) algorithms are not protected by copyright, but we have to consider that in some cases the combination of algorithms form the structure of the program, and, in these cases, should not be excluded from copyright protection, and b) taking into consideration originality and creativity, "... any item contained in a knowledge base or stored in a data base may itself be copyrightable and its storage therefore be subject to the respective rightholder's consent".
NEURAL NETWORKS: AN EVOLUTION OF ARTIFICIAL INTELLIGENCE SYSTEMS. THE IDEA OF SPECIAL LEGISLATION
Neural networks systems are a new kind of artificial intelligence systems where computer programs are set up in such a way as to work like human neurons. These systems, like real humans neurons, can learn from presented data, correct errors, generate information, and be trained by presenting facts.
In a neural network, the "weights" are the means by which the network relates the input values to the correct output. In a conventional expert system, a person called a "knowledge engineer" specifies rules and search techniques to adjust input and output. In a neural network, the system itself assigns and arranges the weights in order to correctly correlate input and output.
Andy Johnson-Laird says that one of the curious characteristics of neural network systems is that "... you can take exactly the same structure of network - we call it topology (to avoid being to obvious) - for a completely different application". According to him, an important point is to determine where to identify the knowledge in a multiramificated neural network. He explains that the knowledge can be found "... in (these) mathematical weights between the neurons. They correspond in human physiological terms to synaptic junctions between one neuron and the next, which is (he suspects) where the knowledge is stored in human brains".
The author of the neural network, itself, or the author of the programs which constitute the neural system is the person who actually wrote the code that simulates the neural network.
The problem is the legal protection of facts used to train the neural network. Andy Johnson-Laird explains that he questions the situation when "... someone who produces a trained neural network, puts it onto the market and allows further training by the user. ... In the case of the neural networks, if you add a great body of knowledge to a network that has four thousand neural weights in it, at the end of all that training, you still have four thousands neural weights, albeit of different values. However, the way they have changed is by going from one pseudo-random numbered group to another pseudo-random numbered group".
The computer programs which constitute a neural network can be dowloaded and dumped as any object code of any conventional computer program, but in terms of infringement, "... it is going to be exceedingly difficult for an expert to look at the neural weights and tell you what they represent and whether there is plagiarism".
Randall Davis states that the problem is not with the neural networks, as they can be considered simply a variety of computer programs which work in a curious way, as a set of interconnections and weights, rather than traditional code. "...They are also created in a curious fashion, typically by being given a set of examples that exemplify what they are to compute, rather than an explicit description of what computation to perform. But neither the creation process nor its results are particularly mysterious". He suggests that to protect a neural network we need simply to protect three things: (i) the pattern of interconnectivity among the units, (ii) the weights on those connections, and (iii) the input and output categories, i.e., the labels that tell us what kind of numbers to put into each input.
As neural networks - and any sort of AI system - constitute new technology far beyond that contemplated by the drafters of present copyright, trade secret and patent laws, Gerald H. Robinson suggests that if the protectability of the intellectual property in neural networks seems doubtful at best, an alternative solution may be ad hoc legislation. "... Yet, formulating the contours of new legislation is a momentous task. On one hand, every incentive for technological advance should be given, i.e., the opportunity to make money, without closing the door so tightly that those with genuinely new ideas cannot enter the competition. The very nature of neural networks suggests that all protected neural networks elements should be required to be licensed by owners at reasonable rates. Moreover, the length of the monopolistic protection granted under new legislation should be quite short so that a developer could recoup investment and make profit, but not hold the ideas too closely and for too long. The very fact that much about the neural network cannot be defined in terms of product, process, or author, and that some of its qualities seem transient suggests that neural networks would be almost impossible to protect and virtually impossible to police for violations of any such law".
As observed by Thomas Dreier, "... It should be noted that the experts of WIPO meeting on a possible protocol to the Berne Convention concluded, with regard to both expert systems and artificial intelligence applications, that at the present stage further studies were needed before it could be determined whether any legal analysis was correct".
|
Artigos