Richard Bright: Can we begin by you saying something about your background?

Antonio Chella: I graduated in computer engineering in 1992. My Ph.D. thesis concerned neural networks for robot vision.  At that time, I was amazed by the vast difference between the richness of the human perception of the world and the trivial, too simplified perception of a robot. Since then, progress has been made, but this difference remains the same: there is something that is still missing in robot perception. So, I started to study cognitive architectures, driven by the idea that perception cannot be studied as a process working alone but as tightly embedded process in human cognition. Eventually, around 2001, after becoming a full professor in robotics, I started to study the general problem of consciousness in humans and robots. I saw it as a natural evolution of my intellectual journey.

In 2005 I organized a meeting on artificial consciousness in Agrigento, in the south of Sicily, where I invited a small group of researchers with similar curiosities. It was a pioneering meeting devoted to this argument. At that time, we were fewer than 20 people. Now the situation is different: there are workshops, journals, books dedicated to artificial consciousness. Very recently, the European Commission published a call dedicated to this topic.

I have to say that I always committed part of my work time to support the community of artificial consciousness researchers: I remember a very productive symposium sponsored by AAAI that I organized in 2007 in Arlington. Since then, I founded the International Journal of Machine Consciousness, now the Journal of Artificial Intelligence and Consciousness, and at the same time, I organized workshops and meetings.

RB: What is the focus of your research?

AC: The general focus of my research concerns the study of cognitive architectures for artificial consciousness. Consciousness is what makes us human, and therefore, to build a human-like robot, it is inescapable to face the problem of consciousness. The approach I am following is the architectural one: the processes which are tightly related to consciousness, like attention, planning, interaction, self-recognition, self-monitoring, and so on, could be implemented and integrated as pieces of complex architecture.  Here, the implementation of every single process of consciousness may also be feasible, but a natural integration of these processes is a formidable task.

In general, building a conscious robot is a real challenging task that may be considered seriously only by the joint action of many universities and public and private research centers. I would be surprised if a research lab alone would implement a conscious robot: we need a critical mass.

RB: In humans, inner speech plays an important role in self-reflection and self-concept emergence. Can you say something about the development of cognitive architecture and robotic self-recognition methods?

AC: Together with my colleague Arianna Pipitone we are building computational models for robot self-recognition and inner speech. Many people, like myself, use inner speech every day and on many occasions. It is a way of thinking, and it is surprising how the mainstream of AI research did not consider inner speech. Some linguists hypothesize that inner speech allows us to acquire consciousness of our thoughts, and in any case, is a way to order and better define our lines of thinking. A source of inspiration is the famous 2000 movie What Woman Want, where the character interpreted by Mel Gibson can literarily hear the inner speech of all the women close to him. So, we asked ourselves what would happen if we could listen to the inner speech of a robot? Indeed, we would understand the robot’s operations better and possibly establish a more trustworthy interaction with it.

Inner speech is undoubtedly a critical ingredient for self-consciousness, and it is deeply studied in psychology. For example, Alain Morin, a psychologist from Canada and a reference scholar for inner speech studies, analyzed the situations that trigger the generation of inner speech, like mirroring themselves or perceiving their name in a conversation. In our model, we followed Alain Morin, and also Alan Baddley. In his three-component working memory model, Baddley hypothesized a central executive module, a phonological loop, which is at the basis of inner speech, and a visuospatial sketchpad. Then, following Baddeley’s model, we designed a cognitive architecture able to generate many forms of inner speech. Another pillar of our cognitive architecture is the associative memory, based on a semantic memory with associative capabilities. So, for example, when the robot perceives an orange, it recognizes it. Now, according to the associative memory, oranges are also fruits. Still, also apples are fruits, and then the robot triggers a search for apples in the scene. At the same time, the robot generates inner speech that summarizes this reasoning process.

Currently, our inner speech system is based on ACT-R, a general-purpose cognitive architecture. Then we linked the architecture to our robot using ROS, the Robot Operating System. We are using Pepper, a humanoid robot developed by SoftBank Robotics.

RB: What is the mirror test task?

AC: The mirror test task is an important test proposed by the psychologist Gordon Gallup in 1970. He hypothesized that an animal recognizes itself in a mirror if it can recognize a mark painted on its body. The mirror test is a standard test for self-recognition, and only a few animal species, like the great apes, can recognize themselves. The test is obviously tightly related to self-consciousness capabilities.

Several scientists studied the mirror test for a robot. Several years ago. I proposed a cognitive architecture for a robot that generated its movements by a simple reactive system without any form of sophisticated intelligence. Then, the robot was able to look at itself, like in a mirror, and create a narrative of its actions.

There is a tight connection between inner speech and self-recognition in front of a mirror, which is a typical trigger for inner speech. So, we employed the same cognitive architecture for inner speech to implement the robot capability to recognize itself. The robot wears a face mask, and it is in front of a mirror. The mask in this experiment has the role of the mark used in the mirror test. Then, the robot by looking at the mirror, it recognizes another robot with a mask. So, it checks whether it may be itself, by moving its arm and controlling if it wears a mask. Finally, the robot looks sideways to check if there is another similar robot close to it. In case there are no other robots, then the robot infers that the robot in the mirror is itself. This computational process captures the main ingredients of the mirror test, as the robot recognizes a mark, and it moves its arm to see if the mark actually is on its own body.

RB:  Is it possible to construct computer/robotic systems that can emulate neural correlates of consciousness, and would that computer/robotic system be conscious?

AC: Consciousness is part of the physical world, so we must study consciousness as a physical phenomenon by using the scientific tools of physical investigations.

However, I am not convinced that mimicking the brain at a lower level, i.e., at the neuronal level, is a thriving research plan. I think that investigating the brain at a higher level of abstraction could be more fruitful. Keeping in mind the late David Marr, I think the computational level is the correct level of abstraction. Consciousness could emerge by the integration of several different processes. As I mentioned before, a formidable task is the integration of all the subtasks that generate consciousness. Moreover, exciting insights for consciousness may emerge by investigating altered states of consciousness, for example, during meditation.

Consciousness is a graded phenomenon, so there is not all or none. Then, a robot may be more conscious than some animals but incomparably less conscious than a human.

RB: Following on from the previous question, given what we know about our own minds, can we expect to intentionally create artificial consciousness? Or will AI become conscious by itself?

AC: It is reasonable to think that consciousness may emerge from the careful integration of many complex processes, but I do not believe this kind of emergence may happen unintentionally. Of course, the exciting idea of building the artificial mind of a child and then letting the system evolve until a full adult mind, is an old one and it dates to Alan Turing. However, we must carefully create all the correct conditions that allow a robot to become conscious, also by considering its development.

RB: How does a machine ‘learn’ without human input?

AC: There are two ways for a machine to learn without human input. The simplest one is related to curiosity, i.e., the machine’s capability to seek new, exciting entities in the environment. Thus, for example, a robot may explore its environment and look for new, exciting things.  Another intriguing possibility concerns the capability of the machine to “dream.” In this case, the machine explores the inner model of its external environment. There is a free exploration in both cases, but the robot explores the actual environment in the first case. In contrast, in the second case, the robot explores the internal, virtual environment.

I conducted experiments of robot dreaming in the Cicerobot project, a robot able to provide guided tours in the Archaeological Museum in Agrigento. The robot had an internal model of the external environment of the museum. During the rest periods, it continued to explore the internal environment, for example, to find new shortcuts or more efficient paths.

RB: Does consciousness require embodiment?

AC: The problem of the embodiment is a critical one for consciousness. I tend to believe that embodiment is a necessary condition for a conscious entity. The concept of embodiment is complex, and it does not simply mean having something like a physical body. Actual embodiment means that the embodied entity has a body with many sensors and actuators that allow it to close contact with the external environment.

Then, an actual embodied entity is an entity that is tightly linked with the environment as one part of a more complex system. Consider our bodies: there is a continuous tight interaction with the external environment through our skin, our eyes, our ears, and so on. Then, maybe we should need to study new kinds of integration of sensors to reach a full embodiment.

RB:  What do you see as the potential implementations of artificial consciousness?

AC: A potential groundbreaking implementation of artificial consciousness would concern all the fields where a robot must perform real ethical choices. Nontrivial ethical questions are typically complex questions where there is only partial prior knowledge. Then it is questionable whether an AI-based system, which is generally based on previous knowledge, would present an ethical behavior. In many cases, an effective moral choice needs many of the ingredients necessary for consciousness, like attention, planning, interaction capabilities, self-recognition, self-monitoring, and some form of complex integration from many different sources. Then, I am convinced that an effective Artificial Moral Agent should have some form of consciousness, even simplified.

Artificial Intelligence, Brain, Computing, Consciousness, Mind, Neuroscience, Robotics, Science, Technology