The problem of meaning in AI: Still with us after all these years

I was invited to give a talk at the “Programs, minds and machines” workshop, which will be hosted jointly by the Mathematics and the Philosophy Research Institutes of UNAM, August 6-9, 2018.

The problem of meaning in AI: Still with us after all these years

Tom Froese

In recent years there has been a lot of excitement about the possibilities of advanced artificial intelligence that could rival the human mind. I cast doubt on this prospect by reviewing past revolutions in cognitive robotics, specifically the shift toward situated robotics in the 90s and the shift toward a dynamical approach in the 00s. I argue that despite claims to the contrary, these revolutions did not manage to overcome the fundamental problem of meaning that was first identified in the context of various theoretical and practical problems faced by Good Old-Fashioned AI. Even after billions of dollars of investment, today’s computers simply do not understand anything. I argue for a paradigm shift in the field: the aim should not be to replicate the human mind in autonomous systems, but to help it realize its full potential via interfaces.

Advertisements

New article on entraining chaotic dynamics

We show that it is possible for a participant to interactively control a chaotic system by entraining with its dynamics, with the effect that they become more regular while the participant becomes more chaotic.

This has implications both for researchers interested in controlling chaotic systems, and also for practitioners in movement rehabilitation.

Entraining chaotic dynamics: A novel movement sonification paradigm could promote generalization

Dobromir Dotov and Tom Froese

Tasks encountered in daily living may have instabilities and more dimensions than are sampled by the senses such as when carrying a cup of coffee and only the surface motion and overall momentum are sensed, not the fluid dynamics. Anticipating non-periodic dynamics is difficult but not impossible because mutual coordination allows for chaotic processes to synchronize to each other and become periodic. A chaotic oscillator with random period and amplitude affords being stabilized onto a periodic trajectory by a weak input if the driver incorporates information about the oscillator. We studied synchronization with predictable and unpredictable stimuli where the unpredictable stimuli could be non-interactive or interactive. The latter condition required learning to control a chaotic system. We expected better overall performance with the predictable but more learning and generalization with unpredictable interactive stimuli. Participants practiced an auditory-motor synchronization task by matching their sonified hand movements to sonified tutors: the Non-Interactive Predictable tutor (NI-P) was a sinusoid, the Non-Interactive Unpredictable (NI-U) was a chaotic system, the Interactive Unpredictable (I-U) was the same chaotic system with an added weak input from the participant’s movement. Different pre/post-practice stimuli evaluated generalization. Quick improvement was seen in NI-P. Synchronization, dynamic similarity, and causal interaction increased with practice in I-U but not in NI-U. Generalization was seen for few pre-post stimuli in NI-P, none in NI-U, and most stimuli in I-U. Synchronization with novel chaotic dynamics is challenging but mutual interaction enables the behavioral control of such dynamics and the practice of complex motor skills.

Psychological study on chaos control

Dobri Dotov and I have published an extended abstract in the Proceedings of the Artificial Life Conference 2018 about the study that he realized at UNAM. We suggest that the results have implications for how we should think about how to stabilize the behavior of complex adaptive systems with which we can interact.

We will present this work at the ALIFE conference in Tokyo as part of the special session on “ALife and Society: Transcending the artificial-natural divide”.

Mutual synchronization and control between artificial chaotic system and human

Dobromir Dotov and Tom Froese

Dexterous assistive devices constitute one of the frontiers for hybrid human-machine systems. Manipulating unstable systems requires task-specific anticipatory dynamics. Learning this dynamics is more difficult when tasks, such as carrying liquid or riding a horse, produce unpredictable, irregular patterns of feedback and have hidden dimensions not projected as sensory feedback. We addressed the issue of coordination with complex systems producing irregular behaviour, with the assumption that mutual coordination allows for non-periodic processes to synchronize and in doing so to become regular. Chaos control gives formal expression to this: chaos can be stabilized onto periodic trajectories provided that the structure of the driving input takes into account the causal structure of the controlled system.

Can we learn chaos control in a sensorimotor task? Three groups practiced an auditory-motor synchronization task by matching their continuously sonified hand movements to sonified tutors: a sinusoid served as a Non-Interactive Predictable tutor (NIP), a chaotic system stood for a Non-Interactive Unpredictable tutor (NI-U), and the same system weakly driven by the participant’s movement stood for an Interactive Unpredictable tutor (I-U). We found that synchronization, dynamic similarity, and causal interaction increased with practice in I-U. Our findings have implications for current efforts to find more adequate ways of controlling complex adaptive systems.

UNISON

Keynote at From Animals to Animats 15 (SAB 2018)

I will be a keynote speaker at FROM ANIMALS TO ANIMATS 15: The 15th International Conference on the Simulation of Adaptive Behavior (SAB 2018), which will take place 14-17 August 2018, in Frankfurt, Germany, and is organized by the International Society for Adaptive Behavior (ISAB).

Here is my title and abstract:

Searching for the conditions of genuine intersubjectivity: From robotics to HCI

Tom Froese

Many our most valued experiences are experiences that we share with others. Yet the basis for this sense of we-ness remains mysterious. Could it really be possible that two people share one and the same experience? How so? Two lines of research are providing important insights. First, complex systems analyses of social robotics and agent-based models have demonstrated that there is nothing mysterious about the possibility of cognitive activity being distributed in a multi-agent system. Second, experimental investigations of real-time embodied social interaction mediated by human-computer interfaces demonstrate that co-regulation of interaction dynamics makes a difference to experience. This formal and empirical research on social interaction supports the possibility of genuine intersubjectivity: we can directly participate in the unfolding of each other’s experience.

Research page updated

I finally found some time to update the research page of my website. Here is the opening paragraph:

I am a cognitive scientist interested in understanding the complexities of the human mind on the basis of embodied, embedded, extended, and enactive approaches to cognition (so called “4E cognition”). For me this means systematically investigating how our minds are shaped by being alive, by being sensorimotor animals, and by us leading socially, technologically, and culturally constituted ways of life (Froese and Di Paolo 2011; Torrance and Froese 2011). One of the most promising approaches to better appreciate the role these different facets can play is to try to understand their origins and the qualitative changes their appearance implies.

The rest can be found here on the research page.

Cognitive science course next semester

Here is the information about the course I will teach at UNAM next semester.

The course will introduce ongoing debates in cognitive science about our changing understanding of the mind. Instead of being thought of as a digital computer inside the brain, mind is now widely considered to be an embodied, embedded and extended activity in the world. These ideas will be illustrated based on case studies of research in agent-based models, complex systems and human-computer interfaces, with special emphasis on demonstrating how social interactions and technologies shape our mind.

Students are not expected to program models nor to design interfaces, but to understand the implications of the new cognitive science and to apply them to their own research interests.

The course will be taught mainly in English to better prepare students for the special terms used by leading researchers in cognitive science.

For an introduction to this field, see this video: http://vimeo.com/107691239

Here is the official course information:

Posgrado en Ciencia e Ingeniería de la Computación (PCIC)

Plan: Maestría en Ciencia e Ingeniería de la Computación (Clave 80-4014)
Actividad académica: Temas Selectos de Inteligencia Artificial
Tema: Agentes autónomos y multiagentes (o: “Agentes Autónomos, Sistemas Sociales, y la Nueva Ciencia Cognitiva”)
Horarios: Lunes y Miércoles, 11:30 – 13:00
Profesor: Dr. Tom Froese

The course program can be downloaded here.

IEEE Haptics Podcast on the Enactive Torch

Our paper on the Enactive Torch, entitled The Enactive Torch: A New Tool for the Science of Perception, which was published in IEEE Transactions on Haptics, is discussed in that journal’s latest podcast. The coverage starts at 11:00.