A Cognitive Architecture Based on Neuroscience for ... - Springer Link
A Cognitive Architecture Based on Neuroscience for the Control of Virtual 3D Human Creatures Felipe Rodr´ıguez1, Francisco Galvan1 , F´elix Ramos1 , Erick Castellanos1 , Gregorio Garc´ıa2, and Pablo Covarrubias3 1
Cinvestav Guadalajara, Av. Cient´ıfica 1145, Col. El Baj´ıo, Zapopan 45010, Jalisco, M´exico {lrodrigue,fgalvan,framos,ecastella}@gdl.cinvestav.mx 2 Instituto de Neurociencias, Francisco de Quevedo 180, Col. Arcos Vallarta, Guadalajara 44130, Jalisco, M´exico [email protected] 3 Universidad del Valle de M´exico, campus Zapopan, Perif´erico Poniente 7900, Col. Jardines del Colli, Zapopan 45010, Jalisco, M´exico [email protected]
Abstract. For the creation of a virtual 3D creature it is necessary an underlying structure that provides to it some desired capabilities. One of our main research objectives is creating a virtual 3D creature that resembles human behavior in an actual environment. In this paper, we propose a cognitive architecture inspired in the recent findings of the neuroscience which will represent the underlying structure for implementing virtual 3D creatures with human-like capabilities. Those virtual creatures will be useful to study human behavior in actual environments by means of simulations.
1
Introduction
Many virtual environments try to simulate our world, in this way virtual creatures that exist in them possess a set of capabilities similiar to those of humans. We propose a cognitive architecture conforming the underlying structure of a virtual 3D creature simulating human-like behavior. The paper is structured as follows: section 2 describes the motivations that leaded us to choose the neuroscience approach; section 3 states the set of desirable abilities; section 4 presents the whole proposed architecture; section 5 shows how each desirable ability works on the architecture; section 6 explains how the results of a neuroscientific experiment are related to our proposal; section 7 gives conclusions.
2
The Cognitive Architecture and the Neuroscience
Although there are a diversity of implemented cognitive processes models and simulations of brain structures (which are focused in specialized functions) [1], our main intention is to establish a fully designed cognitive architecture to build specialized processes based on this design. This paradigm avoids difficulties in Y.Y. Yao et al. (Eds.): BI 2010, LNAI 6334, pp. 328–335, 2010. c Springer-Verlag Berlin Heidelberg 2010
A Cognitive Architecture Based on Neuroscience
329
the integration phase presented in the unified theory of cognition [2]. The development of the architecture proposed in this paper is in accordance with the idea of a unified theory of cognition that Newell argued in [3]. We argue that since each of our components works according to a specific neuroscience theory, the resulting behavior must resemble those of humans. This idea is according to that of Newell description of how to build a unified theory of cognition grounded in our existing understanding of cognition. We claim that our approach is doable since the contemporary neuroscience research has explained more accurately some of the cognitive processes as well as its brain correlated structures [4], [5]. Stressing the benefits of using neuroscience for the construction of a cognitive architecture, we identify at least two important advantages. First, the computational cognitive model may help to integrate separated theories about cognitive processes, leading to unified explanations of cognition theory, covering the gap between isolated findings in neuroscience and unifying them in wider theories [6]. In addition, by using integrated computational models of cognitive processes, neuroscientists may achieve clearer explanations about those processes. On the other hand, related to artificial intelligence and to our main goal, with this approach we are searching for more realistic human behavior on virtual creatures (by means of a set of abilities). In fact, this approach allows the architecture to consider most of the capabilities and properties proposed in [7] and [8]. While our approach is based on neuroscience, there are other cognitive architectures grounded on different theories. Two of the most important architectures are ACT-R [9] and Soar [10]. Although, ACT-R is an architecture psychologically grounded, lately they have mapped some architecture’s modules to parts of the human brain, based on the functioning perspective [11]. But, from its conception, ACT-R was not thought in a neuroscience based design, therefore, that task has not been entirely transparent. Soar has lately been modified with missing capabilities they see granted in humans, such as reinforcement learning, appraisal detector, semantic memory, episodic memory, among others. Although, as they state, all of the new components have been built, integrated, and run individually with the traditional Soar components, there is not a single unified system that has all the components running at once [12]. Accordingly, in order to prevent limitations that have aroused during the development of similar projects our proposal is fully conceived on the neuroscience findings. Thus, the main objective of the paper is to present a cognitive software architecture based on the description of the brain provided by the neuroscience and complemented with the knowledge of the computer sciences useful to provide virtual creatures with human-like behavior. If the behavior is based on abilities, an important question should be what are the sufficient or desirable abilities to achieve a similar human behavior?
3
Abilities for a Virtual 3D Human Creature
Currently, an area from neuroscience have been focused in the study of the executive functions. Those functions can be divided in “meta cognitive” and
330
F. Rodr´ıguez et al.
“emotional-motivational” executive functions, the first includes key abilities to achieve and pursue goals while the second is responsible for coordinating cognition and emotion [13]. In order to approximate human behavior in virtual creatures, the architecture provides some abilities that are based on this perspective. In addition, the perception process and motor response are mediated by the interaction of those executive functions (described below) as well as learning and memory processes. Here a description of the abilities: Perceptual function: (1) Perception: is the process of reception and interpretation of the different external and internal stimulus. This will give to the creature an internal representation of the world and itself. Subsequently, the creature will be able to evaluate the situation and try to shape the environment and itself in order to achieve its goals. Cognitive abilities: (2) Learning: stable changes in the mechanisms of behavior. (3) Memory: the ability to store, retain and recall knowledge. Emotional function: (4) Emotions: the ability to encode emotional stimulus and to influence a set of cognitive process with the emotional nuance extracted from the perceived stimulus. Metacognitive functions: (5) Planning: the ability to create a sequence of possible actions that will lead to a expected result. Within this ability is hide the ability to “imagine” and predict the results of application of actions. (6) Deliberation Process: represents the process of selecting one among a set of possible actions. (7) Cognitive flexibility: ability of spontaneously restructure the self knowledge in an adaptive way to change the environment demands. Motor function: (8) Motor action: the ability to control successfully the movements of each movable body part. As a result of the modification of its body, the environment changes. The architecture should support this set of processes. Next section depicts the design of the proposed architecture and a description of its constituent parts.
4
The Cognitive Architecture Design
The components of the architecture and its proposed function are directly related to brain components and processes [5]. We now describe the characteristics of each module and its role in the architecture. See figure 1. 1. Set of Sensory System: this module catches the environment status. Then, it sends the information to the Thalamus. Due the primitive nature of the olfactory system, this sense sends the information directly to the Olfactory Cortex, after a filter is applied by the Olfactory Bulb. 2. (α) Olfactory Bulb: this module is a first filter to olfactory information. Then, it sends to the association cortex through the Hippocampus. 3. Thalamus: this is the first processing phase for the data received from the sensors. It consists of four modules: (a) (β) Lateral Geniculate Nucleus: receives information from the vision sensor and sends the selected information to Visual Cortex.
A Cognitive Architecture Based on Neuroscience
331
(b) (γ) Medial Geniculate Nucleus: sends the auditory information selected to Auditory Cortex. (c) (δ) Ventrobasal : filters tactile sensory signals before sending them towards the Somatosensory Cortex. (d) () Ventral Posterior Medial Nucleus: taste information is put together here; only a selected amount of it is sent to Gustatory Cortex. Submodules of the Thalamus are all interconnected and share information. 4. Sensory Cortex : this set of modules are the ones in charge of giving an interpretation to the data received by the Set of Sensory System. (a) (ζ) Visual Cortex : incoming data is visually interpreted with knowledge provided by Hippocampus and sent to Association Cortex. (b) (η) Gustatory Cortex : taste information is interpreted, using the information provided by Hippocampus and sent to Association Cortex. (c) (θ) Somatosensory Cortex : somatic data is transformed using data provided by Hippocampus and sent to the Association Cortex. (d) (ι) Auditory Cortex : interpretation of auditory data is done using Hippocampus information. Information is sent to Association Cortex. (e) (κ) Olfactory Cortex : the olfactory data is interpreted using the information provided by the Hippocampus and sent to the Association Cortex. (f) (λ) Association Cortex : this module puts together current and past sensory interpretations and associations of the objects on the environment. It has connection with the Hippocampus to get past information and to return the deducted information, also, it has connections with the Olfactory, Gustatory, Visual, Somatosensory and Auditory Cortex. 5. Limbic system: two of its important functions are related to emotions and long-term memories. (a) (μ) Hippocampus: This module creates a context of all the information gathered. It manages the storage and recall of memories from cortex. At the signal of the Amygdala, store all information received in the recent past, present and future and creates a temporal relationship between those information. (b) (ν) Amygdala: here the information related to context and current state is received thanks to Sensory Cortex through the Hippocampus and Thalamus. This information is used to organize a set of emotional reactions. Those reactions take effect on the Thalamus to affect perception, the Hippocampus to instruct when to keep knowledge in the long term memory and affect context creation, and in the Orbitofrontal Cortex to modify the appraisal level of the information gathered. 6. Prefrontal cortex : coordinates the temporal organization of actions. (a) (ξ) Orbitofrontal Cortex : this module evaluates the affective information of perceived stimulus. It mainly receives information from the Amygdala and projects to the Ventromedial Prefrontal Cortex. (b) (o) Dorsolateral Prefrontal Cortex : it is related to the motor planning behavior, stores the current goal and integrates the information of long term memory and sensory input; the main objective is to create
332
F. Rodr´ıguez et al.
plans to achieve the current goal. It communicates with the Ventromedial Prefrontal Cortex when a plan is created and decisions must be made. When the next action is decided, the order is sent to the Basal Ganglia which in turn regulate (via feedback) the action previously decided. (c) (π) Ventromedial Prefrontal Cortex : receives perceived information from the Hippocampus, emotional appraisal information from Orbitofrontal Cortex and objective information from the Dorsolateral Prefrontal Cortex. With this information, chooses between possible actions to achieve the goal. When one action does not lead to the goal, the information is redirected to Dorsolateral Prefrontal Cortex to form a plan. 7. Motor System: here, the instructions given by the Prefrontal Cortex are translated into body movement attempts. (a) (ρ) Basal Ganglia: this module selects the possible muscles of the body to achieve the action sent by the Dorsolateral Prefrontal Cortex. (b) (σ) Motor Cortex : once an action is received from Basal Ganglia, this module makes the needed calculations to control body and, therefore, complete the action. Olfactory Bulb Lateral Geniculate Nucleus
Frontal Lobe
Medial Geniculate Nucleus Ventrobasal
Parietal Lobe
Ventral Posterior Medial Nucleus Visual Cortex Gustatory Cortex
Sensory System Vision
Limbic System
Somatosensory Cortex
Thalamus
Auditory Cortex Olfactory Cortex
Hearing Somatic Senses Taste Olfactory
Association Cortex
Temporal Lobe
Hippocampus
Occipital Lobe
Amygdala Orbitofrontal Cortex Dorsolatelar Prefrontal Cortex Ventromedial Prefrontal Cortex Basal Ganglia
Motor Cortex
Fig. 1. The Cognitive architecture design
Although we have explained the architecture design, it remains to describe how each ability arises from the interaction of some of those modules.
5
Abilities for Virtual 3D Creatures and the Architecture
Here we explain how each of those modules interact to allow the virtual 3D creature to show the desired abilities [5]. See diagrams in figure 2. 1. Perception: this ability will be granted by the following information cycle: (a) Environment information gathering would be done by the Sensory System (1). The olfactory sensor, after applying a filter to the data with the Olfactory Bulb, sends its data directly to the Olfactory Cortex (2). The rest of the sensors send its data to the Thalamus (3).
A Cognitive Architecture Based on Neuroscience
2.
3.
4.
5.
6.
6
333
(b) After the Thalamus finishes filtering the data, it is sent to the Sensory Cortex and to the Amygdala to affect emotional status (4). (c) Data interpretation and association is done in the Sensory Cortex. When done, it sends this information to the Hippocampus (5). (d) The Hippocampus helps to recall knowledge stored in memory to interpret the data received at the Sensory Cortex (6). It creates a context using memory and emotional information sent by the Amygdala (7). Also, all context and information deduced is stored in memory and sent to the Amygdala to update emotional state (8). Learning: for learning to exist in this architecture, there must be a discrepancy between actual and predicted rewards in the environment. That means that learning will occur if a stimulus is paired with an unexpected reward. Memory: when the Amygdala senses a high emotional level, a temporal window is opened (2). While this window is opened, all information previously passed (1), currently at and passed from that moment on to the Hippocampus (3), will be temporarily related and stored in the long term memory, located at the Association Cortex, for other modules to use (4). Emotions: the emotions would occur in the architecture as follows: (a) The Amygdala receives fast (1) and highly processed sensorial information from the Thalamus and Hippocampus (1.5) respectively. (b) The Amygdala sends processed information to the Thalamus, Orbitofrontal Cortex and Hippocampus to produce a emotional reaction (2). Planning and Decision Making: (a) At Dorsolateral Prefrontal Cortex a plan is built using current state sent by Sensory Cortex (1), the goal and data provided by Hippocampus (2). (b) The Orbitofrontal Cortex receives emotional information (1.5) and emotional appraisal level is set to the knowledge stored at Hippocampus (2). (c) The Ventromedial Prefrontal Cortex receives the plan and the appraisal level (3, 3.5). Using the raw knowledge and its appraisal level (4), the plan is trimmed and sent back to the Dorsolateral Prefrontal Cortex (5). (d) The plan could be refined, extended or returned to the Ventromedial Prefrontal Cortex. When the immediate next action is decided, the action is sent to the Basal Ganglia to be executed (6). Motor action: (a) Dorsolateral Prefrontal Cortex gives the following action to execute to the Basal Ganglia (1). (b) The Motor Cortex receives the information of the current state together with the set of muscles to move and the intended action (2). The action is passed to the muscles and the action is executed.
Neurofunctional Data Over the Architecture
To theoretically prove the system functionality we will make use of the delay match to sample task. We follow the brain activations that emerge during such task and then we compare with the information flow in the proposed architecture, see figure 3. During the training phase, a sample stimulus is presented to
334
F. Rodr´ıguez et al.
1,2,3 3.5
Fig. 2. Diagram of the different processes in the architecture
Test
Sensory System Vision
Thalamus
Limbic System
Medial Lateral Geniculate Geniculate Ventrobasal Nucleus Nucleus
b) Objective set c) A ligth on screen d) Image is captured e) Select the most relevant information f) No relevant information for this sense g) Information produces high excitation
Ventral Posterior Medial Neucleus
Sensory Cortex
Prefrontal Cortex
Orbitofrontal Dorsolateral Prefrontal Ventromedial Amygdala Hippocampus Visual Cortex Association Cortex Cortex Cortex Cortext
i) Information is stored in memory j) Perceptors are built based in grouped stimuli k) An evaluation of the current situation is done l) A meaning is built m) Stimuli appears on screen
Motor System Basal Ganglia
Motor Cortex
n) Hold information o) Makes a relationship between past and current stimuli p) A decision is made q) Pass information to Basal Ganglia r) Selects appropriate motor system s) Executes the response
Fig. 3. Functioning and module interconnection based on neuroscience paradigm
the subject (e.g., a green light) and then, a couple of stimuli to be compared are presented (e.g., a green and red lights together). The choice of the subject is awarded if match correctly with the sample stimulus. Once the subject identifies the relation between the sample and target stimuli, a delay is made between the stimuli presentation and the subject choice. Next, we describe the brain activation together with the architectural work flow. In a matching task there is activation of subcortical areas including sensory system along Thalamus (1); then activates Amygdala functioning that goes to orbitofrontal cortex and hippocampus (2); at the same time, activation then extends to Visual Cortex and follows a ventral path through association extrastriate areas of the occipital and limbic temporal sites near medial temporal circumvolution (3). At this point, activation spreads in a corticolimbic interaction involving the activation of ventromedial prefrontal cortex (4). Then, there are a feedback interaction from
A Cognitive Architecture Based on Neuroscience
335
prefrontal cortices to limbic areas. Following, interaction between Ventromedial and Dorsolateral Prefrontal Cortex produces a direct pass of information to the motor system (5).
7
Conclusions
We use the neuroscience approach and take the executive brain paradigm to orchestate a functionig view of the brain. According to this perspectives, we depict the design of the cognitive architecture, subsequently we show how each ability arises from the functional activity of the various modules and their coupling interactions. In order to explain the expected activations of the architecture, we use a neuroscience paradigm, which allows us to compare the data from the real experiment in humans with the mentioned activations. This is a work in progress where the next stage is the implementation of each process. The final objective is integrate them following the architecture design. Acknowledgments. This research is partially supported by CoECyT-Jal Project No. 2008-05-97094, whilst authors Felipe Rodr´ıguez, Francisco Galvan, Erick Castellanos are supported by CONACYT grant No. 229386, 219078, 219074 respectively.
References 1. Morton, J.B., Munakata, Y.: Active vs. latent representations: a neural network model of perseveration, dissociation, and decalage. Developmental Psychobiology 40, 255–265 (2002) 2. Langley, P.: Cognitive architectures and general intelligent systems. AI Mag. 27(2), 33–44 (2006) 3. Newell, A.: Unified theories of cognition. Harvard University Press, Boston (1990) 4. Carter, R.: Mapping the Mind. University of California Press, California (2000) 5. Kandel, E.: Principles of Neural Science, 4th edn. McGraw-Hill, New York (2000) 6. Thelen, E., Schner, G., Schneider, C., Smith, L.B.: The dynamics of embodiment: a field theory of infant perseverative reaching. Behavioral and Brain Sciences 24, 1–86 (2001) 7. Langley, P., Laird, J., Rogers, S.: Cognitive architectures: Research issues and challenges. Cognitive Systems Research 10(2), 141–160 (2009) 8. Sun, R.: Desiderata for cognitive architectures. Philosophical Psychology 17(3), 341–373 (2004) 9. Anderson, J.R.: Rules of the mind. Lawrence Erlbaum Associates, Mahwah (1993) 10. Laird, J.E., Newell, A., Rosenbloom, P.S.: SOAR: an architecture for general intelligence. Artif. Intell. 33(1) (1987) 11. Anderson, J.R., Bothell, D., Byrne, M.D., Douglass, S., Lebiere, C., Qin, Y.: An integrated theory of the mind. Psychological Review 111(4), 1036–1060 (2004) 12. Laird, J.E.: Extending the Soar Cognitive Architecture. In: Proceeding of the 2008 Conference on Artificial General Intelligence, vol. 171, pp. 224–235. IOS Press, Amsterdam (2008) 13. Ardila, A.: On the evolutionary origins of executive functions. Brain and Cognition 68(1), 92–99 (2008)
Cinvestav Guadalajara, Av. Cient´ıfica 1145, Col. El Baj´ıo, Zapopan 45010, Jalisco, M´exico {lrodrigue,fgalvan,framos,ecastella}@gdl.cinvestav.mx 2 Instituto de Neurociencias, Francisco de Quevedo 180, Col. Arcos Vallarta, Guadalajara 44130, Jalisco, M´exico [email protected] 3 Universidad del Valle de M´exico, campus Zapopan, Perif´erico Poniente 7900, Col. Jardines del Colli, Zapopan 45010, Jalisco, M´exico [email protected]
Abstract. For the creation of a virtual 3D creature it is necessary an underlying structure that provides to it some desired capabilities. One of our main research objectives is creating a virtual 3D creature that resembles human behavior in an actual environment. In this paper, we propose a cognitive architecture inspired in the recent findings of the neuroscience which will represent the underlying structure for implementing virtual 3D creatures with human-like capabilities. Those virtual creatures will be useful to study human behavior in actual environments by means of simulations.
1
Introduction
Many virtual environments try to simulate our world, in this way virtual creatures that exist in them possess a set of capabilities similiar to those of humans. We propose a cognitive architecture conforming the underlying structure of a virtual 3D creature simulating human-like behavior. The paper is structured as follows: section 2 describes the motivations that leaded us to choose the neuroscience approach; section 3 states the set of desirable abilities; section 4 presents the whole proposed architecture; section 5 shows how each desirable ability works on the architecture; section 6 explains how the results of a neuroscientific experiment are related to our proposal; section 7 gives conclusions.
2
The Cognitive Architecture and the Neuroscience
Although there are a diversity of implemented cognitive processes models and simulations of brain structures (which are focused in specialized functions) [1], our main intention is to establish a fully designed cognitive architecture to build specialized processes based on this design. This paradigm avoids difficulties in Y.Y. Yao et al. (Eds.): BI 2010, LNAI 6334, pp. 328–335, 2010. c Springer-Verlag Berlin Heidelberg 2010
A Cognitive Architecture Based on Neuroscience
329
the integration phase presented in the unified theory of cognition [2]. The development of the architecture proposed in this paper is in accordance with the idea of a unified theory of cognition that Newell argued in [3]. We argue that since each of our components works according to a specific neuroscience theory, the resulting behavior must resemble those of humans. This idea is according to that of Newell description of how to build a unified theory of cognition grounded in our existing understanding of cognition. We claim that our approach is doable since the contemporary neuroscience research has explained more accurately some of the cognitive processes as well as its brain correlated structures [4], [5]. Stressing the benefits of using neuroscience for the construction of a cognitive architecture, we identify at least two important advantages. First, the computational cognitive model may help to integrate separated theories about cognitive processes, leading to unified explanations of cognition theory, covering the gap between isolated findings in neuroscience and unifying them in wider theories [6]. In addition, by using integrated computational models of cognitive processes, neuroscientists may achieve clearer explanations about those processes. On the other hand, related to artificial intelligence and to our main goal, with this approach we are searching for more realistic human behavior on virtual creatures (by means of a set of abilities). In fact, this approach allows the architecture to consider most of the capabilities and properties proposed in [7] and [8]. While our approach is based on neuroscience, there are other cognitive architectures grounded on different theories. Two of the most important architectures are ACT-R [9] and Soar [10]. Although, ACT-R is an architecture psychologically grounded, lately they have mapped some architecture’s modules to parts of the human brain, based on the functioning perspective [11]. But, from its conception, ACT-R was not thought in a neuroscience based design, therefore, that task has not been entirely transparent. Soar has lately been modified with missing capabilities they see granted in humans, such as reinforcement learning, appraisal detector, semantic memory, episodic memory, among others. Although, as they state, all of the new components have been built, integrated, and run individually with the traditional Soar components, there is not a single unified system that has all the components running at once [12]. Accordingly, in order to prevent limitations that have aroused during the development of similar projects our proposal is fully conceived on the neuroscience findings. Thus, the main objective of the paper is to present a cognitive software architecture based on the description of the brain provided by the neuroscience and complemented with the knowledge of the computer sciences useful to provide virtual creatures with human-like behavior. If the behavior is based on abilities, an important question should be what are the sufficient or desirable abilities to achieve a similar human behavior?
3
Abilities for a Virtual 3D Human Creature
Currently, an area from neuroscience have been focused in the study of the executive functions. Those functions can be divided in “meta cognitive” and
330
F. Rodr´ıguez et al.
“emotional-motivational” executive functions, the first includes key abilities to achieve and pursue goals while the second is responsible for coordinating cognition and emotion [13]. In order to approximate human behavior in virtual creatures, the architecture provides some abilities that are based on this perspective. In addition, the perception process and motor response are mediated by the interaction of those executive functions (described below) as well as learning and memory processes. Here a description of the abilities: Perceptual function: (1) Perception: is the process of reception and interpretation of the different external and internal stimulus. This will give to the creature an internal representation of the world and itself. Subsequently, the creature will be able to evaluate the situation and try to shape the environment and itself in order to achieve its goals. Cognitive abilities: (2) Learning: stable changes in the mechanisms of behavior. (3) Memory: the ability to store, retain and recall knowledge. Emotional function: (4) Emotions: the ability to encode emotional stimulus and to influence a set of cognitive process with the emotional nuance extracted from the perceived stimulus. Metacognitive functions: (5) Planning: the ability to create a sequence of possible actions that will lead to a expected result. Within this ability is hide the ability to “imagine” and predict the results of application of actions. (6) Deliberation Process: represents the process of selecting one among a set of possible actions. (7) Cognitive flexibility: ability of spontaneously restructure the self knowledge in an adaptive way to change the environment demands. Motor function: (8) Motor action: the ability to control successfully the movements of each movable body part. As a result of the modification of its body, the environment changes. The architecture should support this set of processes. Next section depicts the design of the proposed architecture and a description of its constituent parts.
4
The Cognitive Architecture Design
The components of the architecture and its proposed function are directly related to brain components and processes [5]. We now describe the characteristics of each module and its role in the architecture. See figure 1. 1. Set of Sensory System: this module catches the environment status. Then, it sends the information to the Thalamus. Due the primitive nature of the olfactory system, this sense sends the information directly to the Olfactory Cortex, after a filter is applied by the Olfactory Bulb. 2. (α) Olfactory Bulb: this module is a first filter to olfactory information. Then, it sends to the association cortex through the Hippocampus. 3. Thalamus: this is the first processing phase for the data received from the sensors. It consists of four modules: (a) (β) Lateral Geniculate Nucleus: receives information from the vision sensor and sends the selected information to Visual Cortex.
A Cognitive Architecture Based on Neuroscience
331
(b) (γ) Medial Geniculate Nucleus: sends the auditory information selected to Auditory Cortex. (c) (δ) Ventrobasal : filters tactile sensory signals before sending them towards the Somatosensory Cortex. (d) () Ventral Posterior Medial Nucleus: taste information is put together here; only a selected amount of it is sent to Gustatory Cortex. Submodules of the Thalamus are all interconnected and share information. 4. Sensory Cortex : this set of modules are the ones in charge of giving an interpretation to the data received by the Set of Sensory System. (a) (ζ) Visual Cortex : incoming data is visually interpreted with knowledge provided by Hippocampus and sent to Association Cortex. (b) (η) Gustatory Cortex : taste information is interpreted, using the information provided by Hippocampus and sent to Association Cortex. (c) (θ) Somatosensory Cortex : somatic data is transformed using data provided by Hippocampus and sent to the Association Cortex. (d) (ι) Auditory Cortex : interpretation of auditory data is done using Hippocampus information. Information is sent to Association Cortex. (e) (κ) Olfactory Cortex : the olfactory data is interpreted using the information provided by the Hippocampus and sent to the Association Cortex. (f) (λ) Association Cortex : this module puts together current and past sensory interpretations and associations of the objects on the environment. It has connection with the Hippocampus to get past information and to return the deducted information, also, it has connections with the Olfactory, Gustatory, Visual, Somatosensory and Auditory Cortex. 5. Limbic system: two of its important functions are related to emotions and long-term memories. (a) (μ) Hippocampus: This module creates a context of all the information gathered. It manages the storage and recall of memories from cortex. At the signal of the Amygdala, store all information received in the recent past, present and future and creates a temporal relationship between those information. (b) (ν) Amygdala: here the information related to context and current state is received thanks to Sensory Cortex through the Hippocampus and Thalamus. This information is used to organize a set of emotional reactions. Those reactions take effect on the Thalamus to affect perception, the Hippocampus to instruct when to keep knowledge in the long term memory and affect context creation, and in the Orbitofrontal Cortex to modify the appraisal level of the information gathered. 6. Prefrontal cortex : coordinates the temporal organization of actions. (a) (ξ) Orbitofrontal Cortex : this module evaluates the affective information of perceived stimulus. It mainly receives information from the Amygdala and projects to the Ventromedial Prefrontal Cortex. (b) (o) Dorsolateral Prefrontal Cortex : it is related to the motor planning behavior, stores the current goal and integrates the information of long term memory and sensory input; the main objective is to create
332
F. Rodr´ıguez et al.
plans to achieve the current goal. It communicates with the Ventromedial Prefrontal Cortex when a plan is created and decisions must be made. When the next action is decided, the order is sent to the Basal Ganglia which in turn regulate (via feedback) the action previously decided. (c) (π) Ventromedial Prefrontal Cortex : receives perceived information from the Hippocampus, emotional appraisal information from Orbitofrontal Cortex and objective information from the Dorsolateral Prefrontal Cortex. With this information, chooses between possible actions to achieve the goal. When one action does not lead to the goal, the information is redirected to Dorsolateral Prefrontal Cortex to form a plan. 7. Motor System: here, the instructions given by the Prefrontal Cortex are translated into body movement attempts. (a) (ρ) Basal Ganglia: this module selects the possible muscles of the body to achieve the action sent by the Dorsolateral Prefrontal Cortex. (b) (σ) Motor Cortex : once an action is received from Basal Ganglia, this module makes the needed calculations to control body and, therefore, complete the action. Olfactory Bulb Lateral Geniculate Nucleus
Frontal Lobe
Medial Geniculate Nucleus Ventrobasal
Parietal Lobe
Ventral Posterior Medial Nucleus Visual Cortex Gustatory Cortex
Sensory System Vision
Limbic System
Somatosensory Cortex
Thalamus
Auditory Cortex Olfactory Cortex
Hearing Somatic Senses Taste Olfactory
Association Cortex
Temporal Lobe
Hippocampus
Occipital Lobe
Amygdala Orbitofrontal Cortex Dorsolatelar Prefrontal Cortex Ventromedial Prefrontal Cortex Basal Ganglia
Motor Cortex
Fig. 1. The Cognitive architecture design
Although we have explained the architecture design, it remains to describe how each ability arises from the interaction of some of those modules.
5
Abilities for Virtual 3D Creatures and the Architecture
Here we explain how each of those modules interact to allow the virtual 3D creature to show the desired abilities [5]. See diagrams in figure 2. 1. Perception: this ability will be granted by the following information cycle: (a) Environment information gathering would be done by the Sensory System (1). The olfactory sensor, after applying a filter to the data with the Olfactory Bulb, sends its data directly to the Olfactory Cortex (2). The rest of the sensors send its data to the Thalamus (3).
A Cognitive Architecture Based on Neuroscience
2.
3.
4.
5.
6.
6
333
(b) After the Thalamus finishes filtering the data, it is sent to the Sensory Cortex and to the Amygdala to affect emotional status (4). (c) Data interpretation and association is done in the Sensory Cortex. When done, it sends this information to the Hippocampus (5). (d) The Hippocampus helps to recall knowledge stored in memory to interpret the data received at the Sensory Cortex (6). It creates a context using memory and emotional information sent by the Amygdala (7). Also, all context and information deduced is stored in memory and sent to the Amygdala to update emotional state (8). Learning: for learning to exist in this architecture, there must be a discrepancy between actual and predicted rewards in the environment. That means that learning will occur if a stimulus is paired with an unexpected reward. Memory: when the Amygdala senses a high emotional level, a temporal window is opened (2). While this window is opened, all information previously passed (1), currently at and passed from that moment on to the Hippocampus (3), will be temporarily related and stored in the long term memory, located at the Association Cortex, for other modules to use (4). Emotions: the emotions would occur in the architecture as follows: (a) The Amygdala receives fast (1) and highly processed sensorial information from the Thalamus and Hippocampus (1.5) respectively. (b) The Amygdala sends processed information to the Thalamus, Orbitofrontal Cortex and Hippocampus to produce a emotional reaction (2). Planning and Decision Making: (a) At Dorsolateral Prefrontal Cortex a plan is built using current state sent by Sensory Cortex (1), the goal and data provided by Hippocampus (2). (b) The Orbitofrontal Cortex receives emotional information (1.5) and emotional appraisal level is set to the knowledge stored at Hippocampus (2). (c) The Ventromedial Prefrontal Cortex receives the plan and the appraisal level (3, 3.5). Using the raw knowledge and its appraisal level (4), the plan is trimmed and sent back to the Dorsolateral Prefrontal Cortex (5). (d) The plan could be refined, extended or returned to the Ventromedial Prefrontal Cortex. When the immediate next action is decided, the action is sent to the Basal Ganglia to be executed (6). Motor action: (a) Dorsolateral Prefrontal Cortex gives the following action to execute to the Basal Ganglia (1). (b) The Motor Cortex receives the information of the current state together with the set of muscles to move and the intended action (2). The action is passed to the muscles and the action is executed.
Neurofunctional Data Over the Architecture
To theoretically prove the system functionality we will make use of the delay match to sample task. We follow the brain activations that emerge during such task and then we compare with the information flow in the proposed architecture, see figure 3. During the training phase, a sample stimulus is presented to
334
F. Rodr´ıguez et al.
1,2,3 3.5
Fig. 2. Diagram of the different processes in the architecture
Test
Sensory System Vision
Thalamus
Limbic System
Medial Lateral Geniculate Geniculate Ventrobasal Nucleus Nucleus
b) Objective set c) A ligth on screen d) Image is captured e) Select the most relevant information f) No relevant information for this sense g) Information produces high excitation
Ventral Posterior Medial Neucleus
Sensory Cortex
Prefrontal Cortex
Orbitofrontal Dorsolateral Prefrontal Ventromedial Amygdala Hippocampus Visual Cortex Association Cortex Cortex Cortex Cortext
i) Information is stored in memory j) Perceptors are built based in grouped stimuli k) An evaluation of the current situation is done l) A meaning is built m) Stimuli appears on screen
Motor System Basal Ganglia
Motor Cortex
n) Hold information o) Makes a relationship between past and current stimuli p) A decision is made q) Pass information to Basal Ganglia r) Selects appropriate motor system s) Executes the response
Fig. 3. Functioning and module interconnection based on neuroscience paradigm
the subject (e.g., a green light) and then, a couple of stimuli to be compared are presented (e.g., a green and red lights together). The choice of the subject is awarded if match correctly with the sample stimulus. Once the subject identifies the relation between the sample and target stimuli, a delay is made between the stimuli presentation and the subject choice. Next, we describe the brain activation together with the architectural work flow. In a matching task there is activation of subcortical areas including sensory system along Thalamus (1); then activates Amygdala functioning that goes to orbitofrontal cortex and hippocampus (2); at the same time, activation then extends to Visual Cortex and follows a ventral path through association extrastriate areas of the occipital and limbic temporal sites near medial temporal circumvolution (3). At this point, activation spreads in a corticolimbic interaction involving the activation of ventromedial prefrontal cortex (4). Then, there are a feedback interaction from
A Cognitive Architecture Based on Neuroscience
335
prefrontal cortices to limbic areas. Following, interaction between Ventromedial and Dorsolateral Prefrontal Cortex produces a direct pass of information to the motor system (5).
7
Conclusions
We use the neuroscience approach and take the executive brain paradigm to orchestate a functionig view of the brain. According to this perspectives, we depict the design of the cognitive architecture, subsequently we show how each ability arises from the functional activity of the various modules and their coupling interactions. In order to explain the expected activations of the architecture, we use a neuroscience paradigm, which allows us to compare the data from the real experiment in humans with the mentioned activations. This is a work in progress where the next stage is the implementation of each process. The final objective is integrate them following the architecture design. Acknowledgments. This research is partially supported by CoECyT-Jal Project No. 2008-05-97094, whilst authors Felipe Rodr´ıguez, Francisco Galvan, Erick Castellanos are supported by CONACYT grant No. 229386, 219078, 219074 respectively.
References 1. Morton, J.B., Munakata, Y.: Active vs. latent representations: a neural network model of perseveration, dissociation, and decalage. Developmental Psychobiology 40, 255–265 (2002) 2. Langley, P.: Cognitive architectures and general intelligent systems. AI Mag. 27(2), 33–44 (2006) 3. Newell, A.: Unified theories of cognition. Harvard University Press, Boston (1990) 4. Carter, R.: Mapping the Mind. University of California Press, California (2000) 5. Kandel, E.: Principles of Neural Science, 4th edn. McGraw-Hill, New York (2000) 6. Thelen, E., Schner, G., Schneider, C., Smith, L.B.: The dynamics of embodiment: a field theory of infant perseverative reaching. Behavioral and Brain Sciences 24, 1–86 (2001) 7. Langley, P., Laird, J., Rogers, S.: Cognitive architectures: Research issues and challenges. Cognitive Systems Research 10(2), 141–160 (2009) 8. Sun, R.: Desiderata for cognitive architectures. Philosophical Psychology 17(3), 341–373 (2004) 9. Anderson, J.R.: Rules of the mind. Lawrence Erlbaum Associates, Mahwah (1993) 10. Laird, J.E., Newell, A., Rosenbloom, P.S.: SOAR: an architecture for general intelligence. Artif. Intell. 33(1) (1987) 11. Anderson, J.R., Bothell, D., Byrne, M.D., Douglass, S., Lebiere, C., Qin, Y.: An integrated theory of the mind. Psychological Review 111(4), 1036–1060 (2004) 12. Laird, J.E.: Extending the Soar Cognitive Architecture. In: Proceeding of the 2008 Conference on Artificial General Intelligence, vol. 171, pp. 224–235. IOS Press, Amsterdam (2008) 13. Ardila, A.: On the evolutionary origins of executive functions. Brain and Cognition 68(1), 92–99 (2008)
