 |
 |
|
|||||
 | |||||
| |||||
|
Understanding Humans from Robot Processes of Learning and Behavior
|
|||
|
|||
|
"For example, how is this coffee cup recognized and understood in our brain?" Reaching out his hand to the coffee cup on the desk, Tani says, "We may say that the coffee cup is recognized visually by its shape, color and the like. But it seems that we recognize the existence of the coffee cup on the basis of daily physically experienced patterns, such as the sensation of our fingertips touching the cup's handle and the coffee aroma rising from the cup as it comes into contact with our lips. We are highly interested in such questions as how the spaces for 'meaning' and 'ideas' are formed in the brain from the memory of repeatedly experienced behavioral patterns of interaction with the outer world, and we are working to clarify the mechanisms behind these processes." Raising the coffee cup, he says, "The action of raising a coffee cup is thought to be achieved by unconsciously combining basic motor schemes such as 'reaching out a hand toward the object,' 'grasping it,' and 'raising it' in a time series of behavioral patterns. This process of unconscious combination represents the core of cognition. Additionally, the use of multiple words in combination enables the diverse expression of a wide variety of meanings and ideas. Then, how are they expressed in relation to the memory of experienced behavioral patterns in the brain? Why can we speak fluently using multiple combinations of words without being conscious of the rules of grammar? The brain's neuronal circuits essentially exhibit analogue patterns, and seem to be relatively unsuitable for logical operations, such as 'obeying explicit rules,' which are performed well by computers. It seems that the brain unconsciously acquires something like tacit knowledge, which can never be expressed explicitly in writing, and acts accordingly; what, then, is its reality?" To answer these questions, the members of the Laboratory of Behavior and Dynamic Cognition are engaged in experimental studies on the learning and behavior of a robot incorporating a neuronal circuit model, in search of the brain mechanism behind the sequence from behavior to cognition. |
|||
| A robot that recognizes language and behaves as such | |||
|
The first to introduce here is a joint research program undertaken by team member Yuuya Sugita and laboratory head Tani, which concerns a robot's learning about the relationship between language and behavior. Although knowing nothing at the beginning, the robot becomes able to behave as commanded by repeatedly exercising behavioral patterns corresponding to variously combined words. A red object is placed on the left anterior side of the robot, a blue object at the center, and a green object on the right side (cover page). The language taught to the robot comprises combinations of three verbs and six object nouns. The robot is educated to behave as follows. For example, when commanded to "hit red," the robot approaches the red article and extends its arm to hit the red object, and when directed to "push blue," the robot approaches the blue object and pushes the blue object with its body. "In this process of learning, a key resides in the concept of distributed memory," says Tani. "For example, multiple combinable behavioral patterns are memorized in a single neuronal circuit. In distributed memory, individual behavioral patterns are not memorized independently, but the relationship among such diverse behavioral patterns is learnt. I think that through that process, even higher levels of meaning and ideas may emerge." |
|||
| Exploring the structure of meaning | |||
|
The neuronal circuit used by the robot to learn the relationship between language and behavior is configured with a "language module" to input words serially in due order and recognize sentences, and a "behavior module" to generate behavioral patterns in a time series (Figure 1). "In both modules, the process of 'prediction' plays a key role. Dr. Masao Ito, special adviser to the Brain Science Institute of RIKEN, proposed the hypothesis, 'the cerebellum predicts the sensation obtained as a result of motion,' early in the 1970s, providing a hint to our work. The language module predicts sequences of words, whereas the behavior module predicts time series flows of sensor input and motor output. It is the parametric bias (PB) neuron in our concept that bridges the two modules." What are the functions of the PB neuron? "The PB neuron accepts information from both the language and behavior modules during learning, in which situation the two modules mutually interact with and restrict each other. Hence, a structure allowing the correspondence of sentences and behavioral patterns has been arranged in the networks with the PB neurons. Following learning, inputting the two words 'hit red' in the language module results in the generation of a corresponding firing pattern for a group of PB neurons. This signal is then sent to the behavior module, which in turn predicts the time series flow of sensor input and motor output during the execution of the command 'hit red,' resulting in an actual behavioral pattern (the left panel of Figure 2)." The right panel of Figure 2 shows a plot of the two-dimensional principal components of the firing patterns of a group of PB neurons obtained when 18 combinations of verbs and object nouns were input. "This revealed an interesting fact, and we were able to see a regular structure in this plot," says Tani. Because the red object is always located on the left side of the robot, the commands "push red" and "push left" have the same meaning in its behavior; the PB plots for the two cases are close to each other. This is also true for the relationship between blue and center, and between green and right. Furthermore, individual sentences are arranged regularly on the two-dimensional lattice concerning verbs and object nouns. "Here, it is important that the robot has learnt only 14 of the 18 combinable sentences. Even without learning the remaining four sentences, the robot can recognize the sentences and behave correctly; the PB plot points corresponding to the four sentences (within the dotted circle in the left panel of Figure 2) appear at seemingly the right positions on the two-dimensional lattice." Tani explains: "This result demonstrates that the robot is capable of inferring unlearned behavioral patterns from learned patterns. We can say that this estimation has become possible as a result of the self-organization of a structure that combines verbs and objects in the neuronal circuit, as shown in the right panel of Figure 2. This is an important feature of distributed memory. Meaning cannot exist independently of individual things. Our experiments have provided good evidence that 'meaning' emerges in the structure of the relationship between words and behavioral patterns established according to their interactions during robot learning."
|
|||
| Interactive behavior of humans and a humanoid robot | |||
In cooperation with SONY Corporation, the Laboratory of Behavior and Dynamic Cognition is implementing a research program concerning the interactive behavior of humans and QRIO, a humanoid robot. Then, why was QRIO chosen? "We cannot create a robot with such a high degree of freedom. Additionally, QRIO is best suited for trial-and-error-based experimental studies because it is of moderate size, and also because it is unlikely to break, and hence is not dangerous to humans even if it falls."A behavior module incorporating PB neurons was installed in QRIO to allow the robot to mimic and learn a number of motor patterns of the upper half of the human body. Data on the positions of the arms of the facing man are input to the behavior module via a vision camera, and QRIO predicts and learns the motions of the man's arms from the repeated input of time series patterns. At the same time, QRIO's arms are moved with the support of human hands to teach the arm motor time series to enable its arms to mimic the human arm motor patterns. "As learning advances, different spatial patterns of PB neuron firings are obtained for different motor patterns. When the learning process has been nearly completed, a single learned motor pattern is shown to QRIO. Then, the firing pattern of the PB neurons converges into a particular learnt spatial pattern to mimic human arm motor patterns that are visually input to the behavior module. As a result, QRIO begins generating the corresponding motor patterns (Figure 3). Here, the PB neurons work to recognize sensor input time series patterns and generate corresponding motor patterns. This is also based on the above-described mechanism in which words are recognized and corresponding behavioral patterns are generated. The nerve cells capable of such simultaneous signal expression of both cognition and generation are known as mirror neurons, and have been found in electrophysiological experiments in the monkey and elsewhere." The same experimental study has revealed other interesting things, says Tani. "If a single neuronal circuit in QRIO is allowed to learn multiple motor patterns, the robot becomes able to generate novel motor patterns. This process can be considered to arise from a distortion of the internal structure of memory due to the hustling and jostling of the numerous different motor patterns packed in the memory. During interactive motions of QRIO and a human, the robot sometimes begins generating totally new motor patterns as if it possesses its own spontaneity. It can be considered, however, that this apparent spontaneity is not provided from outside the robot, but comes from potential distortions inside the memory. At the moment that QRIO unexpectedly begins generating new movement patters, rather than simply repeating what it had learnt, I realized the existence of a subjective entity beyond the machine." Tani's robot work is also directed to the exploration of major issues in brain science, including mind and consciousness. "When we are performing an ordinary action, our behavioral pattern proceeds automatically while we are virtually unconscious of it. It can be thought, however, that when something unusual occurs, we become conscious of what it is. In the QRIO experiments, any discrepancy between information from the sensor and prediction based on memory alters the firing pattern of PB neurons to induce another motion. I think that we become 'conscious' of the boundary between ourselves and the outer world when a discrepancy between the reality of the outer world and its subjective image has occurred." 
|
|||
 | |||
|
|||
[ Top ] |
We are constantly influencing our environment through our behavior, and are also influenced by the same environment. How do we humans acquire meaning and ideas in the world from our repeated experiences of these patterns of interaction? Additionally, humans are able to create new actions, meaning and ideas by uniquely combining experienced interactive patterns. What are the mechanisms in the brain that enable this process? The objective of the research activities of the Laboratory of Behavior and Dynamic Cognition, headed by Jun Tani, is to answer these questions. To this end, the members of the team are working to elucidate the processes of learning, cognition and behavior of a robot incorporating a neuronal circuit model based on the human brain mechanism. Their work may provide the right answer to the question, why we have consciousness. 
