VIEWS: 9 PAGES: 5 POSTED ON: 10/25/2010
10 A cognitive architecture for robot brains 10.1 THE REQUIREMENTS FOR COGNITIVE ARCHITECTURES In artificial intelligence cognitive architectures are computer programs with an overall structure that tries to emulate and combine various cognitive functions in order to generate human-like cognition and general intelligence. An artificial cognitive system may try to emulate the human way of cognition, but it does not necessarily try to emulate the exact organization and structure of the human brain. These architectures may consist of separate program modules for the various cogni- tive functions; there may be individual modules for perception, memory, attention, reasoning, executive control and even consciousness. These architectures describe a certain organization and order of computation between these modules. They may be criticized for not introducing any real qualitative difference, as the executed computations are still those of the conventional computer. It can be argued that a computer program will not readily possess consciousness even if one of its program modules is labelled so. Thus, the author proposes that a true cognitive architecture would not be a computer program. It would be an embodied perceptive system that would be characterized by the effects that it seeks to produce. The foremost effect that a true cognitive architecture should produce would be the direct, lucid and enactive perception of the environment and the system itself. The system should perceive the world directly and apparently and not via some indirect sensory data. This effect should be accompanied by seamless sensorimotor coordination, the immediate readiness to execute actions on the entities of the environment. The lucid perception of the world should be accompanied by processing with meaning. The architecture should support the association of percepts with additional entities and, as a consequence, the creation of associative networks of meaning and models. This would also involve the ability to learn. A true cognitive architecture should be able to generate both reactive and deliberate responses. The latter would call for the ability to plan, imagine and judge Robot Brains: Circuits and Systems for Conscious Machines Pentti O. Haikonen © 2007 John Wiley & Sons, Ltd. ISBN: 978-0-470-06204-3 180 A COGNITIVE ARCHITECTURE FOR ROBOT BRAINS the actions. This, in turn, leads to the requirement of the flow of mental content, world models and abstractions that are detached from the direct sensory percepts. This kind of content could be described as ‘an inner world’ or ‘inner life’. Humans have the flow of inner speech. An advanced cognitive architecture should also be able to support natural language and inner speech. True cognitive architectures should be designed to support autonomous self-motivated systems such as robots. 10.2 THE HAIKONEN ARCHITECTURE FOR ROBOT BRAINS A cognitive architecture that integrates the circuits, subsystems and principles pre- sented in the previous chapters is outlined in the following. This architecture seeks to implement the above general requirements for true cognitive architectures. The cognitive architecture is designed for an embodied perceptive and interactive system, such as a cognitive robot, with physical body, sensors, effectors and system reactions. It is based on the cross-connected perception/response feed- back loops and the architecture contains a perception/response feedback loop module for each sensory modality. These modules work concurrently on their own, but they broadcast their percepts to each other and may also cooperate with some or all of the other modules. The architecture supports the processes of multimodal perception, prediction, distributed attention, emotional soundtrack, inner models, imagination, inner imagery, inner speech, learning, short-term and long-term memories, system reactions, machine emotions, good/bad value systems and motivation. The block diagram depicting the Haikonen cognitive architecture is presented in Figure 10.1, where seven perception/response feedback loop modules are depicted. This is not a fixed specification; other sensory modalities may be added as desired. These additional sensory modalities may even include some that humans do not have, such as the sensing of electric, magnetic and electromagnetic fields, etc. The structure of these modules follows the principles that are described earlier. Especially blocks of the ‘neuron groups’ depict the AH-STM/LTM style of circuitry of Figure 7.3, with additional circuits as discussed in previous chapters. The modules communicate with each other via the broadcast lines. The cognitive system is motivated by the modules 1 and 2. The module 1 evaluates the emotional good/bad significance of sensory and imaginary percepts and triggers suitable system reactions. This module also provides the emotional soundtrack, and facilitates emotional learning. Hard contact or mechanical damage sensors provide pain information. This module is important for the survival of the robot. Here the principles of Chapter 8, ‘Machine Emotions’, are used. The physical requirements of the system are detected by the module 2. For this purpose the module may include sensors for energy levels, motor drive levels, mechanical tension, balance, temperature, moisture, etc. These sensors may also provide pain information for the module 1. This module is also responsible for sensing mechanical and electrical soundness of the system. Detected anomalies evoke specific short-term and long-term goals and goal-oriented behaviour. THE HAIKONEN ARCHITECTURE FOR ROBOT BRAINS 181 MODULE 1 broadcast lines survival, pain & pleasure emotional syst. good/bad- sensors evaluation react. criteria MODULE 2 survival, self- feedback neuron basic percept WTA sensors neurons groups needs m/mm/n MODULE 3 self image, feedback neuron environment haptic percept WTA neurons groups sensing m/mm/n MODULE 4 environment feedback neuron perception, visual percept WTA neurons groups imagination m/mm/n MODULE 5 environment feedback neuron perception, auditory percept WTA neurons groups language m/mm/n MODULE 6 orientation, feedback neuron mental direction percept WTA neurons groups maps m/mm/n MODULE 7 motion, feedback sequence neuron action, kinestethic percept neurons assemblies language m/mm/n effectors Figure 10.1 The Haikonen cognitive architecture for robot brains The module 3 is responsible for the sensing of soft contact or touch information. This module can sense the shape and surface of objects. It can also create an inner model of the robot body by touching every reachable point and associating this with kinesthetic location information. The principles of section 5.6, ‘Haptic perception’, in Chapter 5 are used here. The module 4 is responsible for the sensing of visual information. This module also facilitates visual introspection, imagination and the use of visual symbols and visual models of objects. The principles of Section 5.7, ‘Visual perception’, are used here. The module 5 is responsible for the sensing of auditory information and auditory introspection. This module also facilitates the use of linguistic symbols, words and inner speech. The principles of Section 5.8, ‘Auditory perception’, and Chapter 9, ‘Natural Language in Robot Brains’, are used here. 182 A COGNITIVE ARCHITECTURE FOR ROBOT BRAINS The module 6 provides ‘absolute’ and relative direction information about the robot’s orientation. This information is utilized in the generation of temporary and long-term inner maps of surroundings. The principles of Section 5.9, ‘Direction sensing’ in Chapter 5 are used here. The module 7 is responsible for the learning and generation of motion and motor actions. The operation of this module is closely related to the haptic, visual and auditory sensory modules. The principles of Chapter 6, ‘Motor Actions for Robots’, are used here. In this kind of cross-connected architecture each module may be trying to broad- cast its percepts to all the other modules. However, it is not feasible that every module would receive and accept broadcasts from every other module at the same time, as each of these broadcasts might try to evoke completely different and potentially clashing percepts and responses in the receiving modules. Therefore an internal attention mechanism is needed. This mechanism should select the most pertinent broadcasts at each moment that the modules should accept and receive. The inner attention should also work the other way around; a module that processes an important stream of percepts should be able to request a response from other modules. Both these attention control situations may be realized by signal intensity modulation and threshold control by the means that have been already discussed. The response request may be realized by the principle of Figure 5.5. Emotional significance may be used in attention control by the principles of Figure 8.1. Does this architecture conform to the requirements for true cognitive architec- tures? The foremost effect that the architecture should produce is the direct, lucid and enactive perception of the environment and the system itself. How would this architecture produce the internal ‘lucid appearance’ that objects and entities are located out there? For humans things are out there because they intuitively know what it would take in motor action terms to reach out for them. Humans also notice that when they turn, the relative directions towards the things change, but the things remain stationary and do not move along their movements. The enabling fact behind the possibility to perceive things to be out there is that the real origination point of the sensation, the sensor, is not taken as the origination point of the sensed stimuli. Thus the retina is not taken as the location of images and the eardrums are not taken as the origination points of the sounds. This architecture reproduces these conditions. In this architecture sensory information is represented in a way that allows the association of a sensation with an external point of origin, one that is coherent with the results of explorative motor actions, such as turning the head. This is enabled by the associative coupling between the sensory modalities and the motor modalities. How would this architecture process meaning? This architecture is based on the perception/response feedback loops that inherently operate with meaning. The intrinsic meanings of the signals are grounded to the feature detectors, but each signal and group of signals may convey additional associated meanings and may operate as a symbol. These meanings are learned. The perception/response feedback loops also allow match and mismatch detection between predicted or desired conditions and actual conditions. This facilitates the ON HARDWARE REQUIREMENTS 183 adjustment of behaviour to suit the situation and the system’s needs. In the introspective mode the perception/response feedback loops disengage from the sensory perception and generate percepts of imagined events. In this way the system may try actions with- out actually executing them as described in Chapter 7, ‘Machine Cognition’. 10.3 ON HARDWARE REQUIREMENTS The Haikonen architecture utilizes the associative neuron group as the basic building block and the perception/response feedback loop as the next-level assembly. At this moment these are not freely available as integrated circuits and the question is: ‘If one were to design such chips then how many neurons should be integrated?’ The human brain has some 1011 neurons and some 1014 synapses. On the other hand, the honey bee has only about 960 000 neurons (Giurfa, 2003) and yet has impressive cognitive and behavioural capacities. The required capacity may also be estimated from the sensory requirements. Auditory perception might do with around 100 feature signals and visual perception might generate some 1000 feature signals. In a neuron group each signal would correspond to a neuron and the number of synapses for each neuron would depend on the number of cross-couplings. If each neuron were coupled to every other neuron, then the number of synapses for each neuron would be the total number of neurons. In small systems this might be the case; in large systems all neurons are not connected to every other neuron. Thus, a small 100 neuron group might have 10 000 synapses and a 1000 neuron group could have one million synapses. A complete system would, of course, utilize several neuron groups. Systems with minimal sensory capacity and clever preprocessing could do with a rather small number of neurons and synapses. However, the integration of a large number of synapses should not be a problem now that humans are able to produce low-cost gigabyte memories (around 1010 one bit memory locations). Another problem with the hardware realization is the large number of intercon- necting wires. Within an integrated neuron group the wiring reduces into a simple geometry and should be manageable. The wiring between chips, however, becomes easily impractical due to the immensely large number of parallel lines. This can be solved by serial communication between chips. A simple protocol that allows the cross-connection of chips in a transparent way can be easily devised. The serial communication can be very fast and will not deteriorate the overall speed perfor- mance of the system. However, the serial communication will turn the system into a temporally sampling system.
Pages to are hidden for
"Robot Brains 10"Please download to view full document