Professor Włodzisław Duch: Good education is key to the success of every strategy connected with AI
Today’s most important artificial intelligence (AI) technologies are a combination of possibility to analyze signals (images) and reasoning based on heuristic reasoning methods; they are therefore a merger of techniques applied in the IBM Watson system and deep learning methods used for big data sets. This allows to create applications that use image analysis, medical diagnostics, analysis of satellite images, control over self-driving vehicles, etc. They require both recognition of structures (perception) and reasoning which uses the perception.
That is why good education is key to the success of every strategy connected with AI. Since software development is much more complicated than designing bridges, warranties for software differ from the ones for building structures. Yet, AI systems are sometimes created by people who have not even completed a basic course or read a good coursebook.
Here are some of the topics tackled by my team. All of them have a great commercial potential both as basic and applied technologies.
Meta-learning
Problem: on the market there are a lot of software packages for machine learning which can be used for analyzing data in millions of ways. Creating models with the use of typical learning systems requires a lot of knowledge and consists in manual construction of models, initial data preparation, feature selection, choice of a method and a detailed architecture of the learning system, and then in teaching its parameters on data collected.
About 20 years ago I proposed how to solve that problem with meta-learning. I suggested to search for the best models among all models available and, once the most promising one have been identified, to proceed with further learning of parameters in a standard way. To do that, high computing power is needed but fortunately we already have it at our disposal.
We have also created (main developers being Norbert Jankowski, PhD DSc, and Krzysztof Grąbczewski, PhD DSc) Intemi software, which defines a space and process of searching for the best models basing on the complexity analysis.
Benefits: significantly easier data modeling without deep machine learning knowledge. This approach can form the basis for further rapid development of machine learning applications in many artificial intelligence projects.
Basic concepts and interpretability of machine learning models
One of the most important problems today is to create interpretable models and to understand the basics of how learning machines work.
In particular, we have developed models suitable for analyzing difficult projects with internally complex logic, based on new learning objectives, feature support machines, as well as an alternative to fuzzy systems built on prototypes and similarity-based learning. Such approaches have numerous advantages over fuzzy logic methods commonly used in AI and have been rediscovered in recent years by machine learning experts due to their ability to interpret neural networks with new neural transfer functions.
Interpretation of machine learning models is not always possible. The two most important approaches are the extraction of logical rules, which describes the operation of such models, and the visualization of their operation.
In many cases, the functions performed by neural networks and other machine learning models are too complex to be presented in an understandable way. However, you can graphically represent the results of their work and make sure that such solutions are stable and safe. This solves the widely discussed problem of the “black box”, which limits the application of AI in areas requiring reliability and stability of predictions.
Artificial intelligence applications that use natural means of communication by referring to our cognitive abilities at the psychological level are called cognitive applications. IBM refers to its Watson technology as “cognitive computing”.
AI systems are sometimes created by people who have not even completed a basic course or read a good coursebook
A related discipline, referred to as “cognitive infocommunications”, focuses on communication between humans and information systems and on the applications which result therefrom. The level of neuropsychological functions is slightly deeper and the inspirations for AI that are derived from it can be described as neuropsychological informatics, which is much less developed but will become hugely important in the future. Such approaches are a vital part of artificial intelligence.
At present, most big companies offer so-called personal assistants, e.g. Amazon Alexa, Google Assistant, Apple Siri, Samsung Bixby, Microsoft Cortana and many more. This is an example of cognitive informatics, where artificial intelligence techniques are used to understand speech and to create simple models of knowledge relating to user’s expectations. This is often done with the use of cognitive architectures, i.e. developing software based on a certain model of natural cognitive systems, usually inspired by the general brain structure.
Due to the weak development of artificial intelligence in the field of natural language processing methods, there are still no personal assistants that would communicate in Polish. In this case it is necessary to understand the syntax and the questions (I delivered lectures on this subject in 2005).
In 2004 in Singapore I proposed to create an interface called HIT (abbreviation of “humanized interfaces”) for mobile devices that would include a graphical avatar and that could represent its owner in some specific situations. After recognizing to which category a question belonged, the program would run a relevant specialized application allowing for voice, touch or gesture-based interaction. The use of the Q/A (question/answer) technology, which is similar to the 20 questions game, made it possible to narrow down the scope of subjects. We planned educational applications, tests to assess knowledge at different levels, health advice solutions and much more. Unfortunately, the project was not funded back then.
Neurocognitive informatics
Since 2017 DARPA [Editor’s note: Defense Advanced Research Projects Agency] has been implementing the “Targeted Neuroplasticity Training” program, which is to support the learning of many skills, including foreign languages, cryptography and methods of intelligence analytics.
Detection of suspicious structures in images or during real-time observations is much more efficient if the neurofeedback method is used, which is designed to draw attention in a situation when a brain responds but no conscious action is taken. Such methods have also produced good results in the case of systems recommending information browsed on the Internet.
Many applications of neurocognitive informatics will pertain to neuromodulation methods. Currently, Parkinson’s disease and many other diseases are treated with numerous methods of deep stimulation and vagus nerve stimulation; various methods of stimulation are applied in depression or chronic pain (this is done by the Polish Neuromodulation Society and the Polish Society for the Study of Pain).
Due to the weak development of artificial intelligence in the field of natural language processing methods, there are still no personal assistants that would communicate in Polish
There are also new opportunities to diagnose and treat people with psychosomatic problems. A significant part of mental problems are caused by abnormal information flow in the brain, too weak or too strong connections between different areas (i.e. disorders of the connectome, a set of such connections). Soon it will be possible to reconstruct the connections by increasing neuroplasticity with the use of non-invasive methods such as DCS and TMS.
Direct brain stimulation with microelectrodes offers even more possibilities. The effects of such stimulation in the premotor cortex have already been demonstrated. And although the stimuli were too weak to cause muscle contraction, they contributed to faster learning of specific skills. Synchronization of processes in the frontoparietal network with TMS significantly improves the performance of the working memory. We are also working on memory implants that would replace certain areas of the hippocampus, which is of great importance in the treatment of people suffering from mild dementia and other diseases causing memory disorders.
The analysis of brain activity will allow for better learning environments adapted to individual differences in learning preferences. In our laboratory we conduct research focused on development of phonemic hearing, which is essential for learning a language and for the working memory in the case of babies and for mathematical skills in the case of kindergarteners.
Monitoring the development with non-invasive measuring devices, occulometry, EEG, and video image analysis will allow to create environments and interactive toys to achieve best possible conditions for the development of children’s intelligence.
The matter is of great social importance and, in this case, artificial intelligence is used in many stages of data analysis. Two spin-offs were created in our Neurocognitive Laboratory: Neurodio and PerKog Technologies. We have won four gold medals at the following invention fairs: Lépine in Paris, INPEX in Pittsburgh, INTARG in Kraków and INNOVA EUREKA 2015 in Brussels; we have been honored with a distinction awarded by the Governor of Kujawsko-Pomorskie Province and with a Diploma of the Minister of Science and Higher Education for the design of an intelligent cradle and cognitive toys that are capable of providing early diagnosis, detecting developmental disorders, ensuring continuous monitoring of babies and guiding their development.
Stimulation of phonemic and musical hearing development will make it possible for children to learn any language, including tonal languages. Stimulating the development of working memory by posing challenges and using brain reward mechanisms will rouse children’s curiosity, make them more active and more willing to learn and explore.
We have also proposed projects to assess the mental state of people responsible for other people’s lives (e.g. pilots or operators of dangerous devices), based on the analysis of physiological parameters and neuropsychological tests. We have developed an automatic system for analyzing psychometric tests, e.g. MMPI, which detects mental disorders and can be used to evaluate employees of large companies. The system has been trained on examples provided by psychometricians and uses fuzzy rule-based reasoning. So far, it has been used mainly by clinical psychologists to test students.
The idea of the DISCOVERITY was to detect fraudulent attempts and false answers to questions asked, which was possible due to the analysis of voice signals, breathing, pulse, thermal imaging, video-recorded facial microexpressions, body and eye movement, as well as EEG (if a more detailed analysis was required).
Our Alter Ego project was designed to collect and store memories of people suffering from dementia
In 2005 we presented DISCOVERITY to Defence Science Organization in Singapore, but it turned out that it was too early for such projects. I also showed it to psychiatrists working in the veterans’ hospital in Cincinnati and presented it as a project to conduct a preliminary patient interview capable of summarizing changes in patients’ health condition. If there are grounds to believe that some information has been withheld or that false information has been provided, the dialogue system asks additional questions and refers to previous conversations. It can be combined with telemedicine projects, reducing the time it takes for doctors to diagnose their patients.
Within the scope of the LifeNaut project many people create repositories that allow them to build personal avatars which their descendants can talk to. Our Alter Ego project was designed to collect and store memories of people suffering from dementia.
A system collecting information about the lives of older people with memory problems, asking detailed questions about their family relationships, showing them photographs and videos, and reminding them of the people they can see in such photos and videos helps them to preserve their personality and strengthens relationships with their family members and caregivers. Evoking memories, analyzing relationships between people, encouraging patients to talk about missing details, and showing them videos and photos is particularly important in the case of Alzheimer’s disease, because it prolongs the period of time during which it is still possible to communicate with sufferers.
Although the projects described above were developed more than ten years ago, they have not been greeted with understanding. Implementation of some of them was possible only after setting up the Neurocognitive Laboratory in a new research center.
Creative computing
In one of my first articles on intuition and computational creativity, neurobiological inspirations were used to build a model for word creation based on descriptions of products and services. The model requires to train associative neural networks on known cases (e.g. on a dictionary of a given language), initially stimulate such networks with product descriptions, search for strongly interconnected fragments of representation and selection based on semantic and phonological filters.
This is the first implementation of the theory of creativity, known in cognitive psychology as Blind Variation Selective Retention (BVSR). Not only does it make it possible to create new names (such services are offered by many specialist companies), but also to analyze neologisms and to give them a meaning.
It is also connected with word games, such as the 20 questions game, which allows to clarify the meaning of the questions in the natural language dialogue systems. The key part of all applications requiring text and speech comprehension is semantic memory. Bots operating in virtual environments and tasked with displaying information on corporate websites use a primitive semantic memory and have very limited capabilities.
Large projects implemented in the past decade by such companies as Cyc Corporation and Microsoft, or MIT ConceptNet and Open Mind Common Sense projects, aimed at creating systems whose reasoning is based on large semantic networks, were not successful.
However, a significant progress in the development of dialogue systems in natural language may be expected from the company called Semantic Machines, which was acquired by Microsoft in 2018. Their project uses speech recognition methods, deep machine learning and reinforcement learning, and creates large data corpora needed to train models. Much of the technology is to be language-independent. The system is planned to be first developed in English and Chinese. Inflected languages, such as Polish, pose specific problems.
A dialogue system based on the above will take into account the context and intentions of the interlocutor; it should also show some creativity. If such a system is created, it will revolutionize the way computers, mobile devices and the Internet of Things are used. In recent years, computational creativity has become an important topic taken up by many new groups. If I am not mistaken, Poland’s BVSR implementation is still the only one in the world.
This paper is an adaptation of the first part of the study by professor Włodzisław Duch entitled “Comments on the strategic artificial intelligence program in Poland. Introduction to machine learning, cognitive informatics and neurocognitive technologies” (Toruń 2018), prepared for the National Information Processing Institute.