• No results found

In conclusion, GWT states that there is a global workspace formed by a dominant coalition of contexts. This core is dynamic, in a sense that it is not localized in a specific region of the brain, but it moves around according to the activation of the small networks in the brain, as if it was a serialization of the parallel activity in the brain. This idea leads

us to our definition of Machine Consciousness, which will be presented in section 3.5.

3.4

State of the Art in Applications of the GWT

3.4.1

Stan Franklin and the Learning Intelligent Distribution

Agent (LIDA)

Stan Franklin’s research group, from the University of Memphis, was the first to de- velop a machine consciousness algorithm inspired in the Global Workspace Theory (Baars, 1988). This algorithm was applied with success to solve different problems such as the creation of a virtual personal secretary (C-Mattie), an American Navy tasks planner (IDA - Intelligent Distribution Agent) and an intelligent tutor system to train Canadian astro- nauts on how to operate a mechanical arm for the International Space Station (Bogner,

1999;Negatu,2006; Dubois, 2007).

Based in his previous experience in developing systems with consciousness, Franklin evolved the LIDA architecture (Learning Intelligent Distribution Agent) (Baars & Franklin,

2009; Franklin et al., 2014a), as both a conceptual and computational model grounded

mainly in the GWT (Baars, 1988).

The LIDA model and its architecture are based in a cognitive cycle that can be di- vided in three phases: perception, interpretation and action. An agent’s life can thus be seen as a continuous sequence of such cognitive cycles. The architecture uses many com- putational mechanisms known in the literature, such as the Copycat Architecture (Hof-

stadter & Mitchell, 1994), Sparse Distributed Memory (Kanerva, 1988), Schema Mecha-

nism (Drescher, 1991) and Behavior Net (Maes, 1989). The LIDA model is presented in

3.4. State of the Art in Applications of the GWT 64

Figure 3.7: The LIDA model

Franklin’s LIDA makes heavy use of the concept of Codelets. Codelets, followingHof-

stadter & Mitchell(1994), are small pieces of non-blocking code, each of them executing

a well defined and simple task. The idea of a codelet is of a piece of code which ideally shall be executed continuously and cyclically, time after time, being responsible for the behavior of a system’s independent component running in parallel. A codelet can be seen

as a nervous system basic functional unit, like the cortical column Mountcastle (1978)

first described. In computational systems, codelets are scalable nodes in a network. Coalitions are groups of one or more codelets. Coalitions are formed by codelets that sum their skills in order to perform more complex tasks, which they would not be able to perform by themselves in an isolated manner.

In the perception phase, sensory stimuli are grabbed from the internal and external environment and stored in the Sensory Memory. These stimuli work as cues for the

3.4. State of the Art in Applications of the GWT 65 Model. This mechanism works as follows. When one stimulus is identified as relevant, one node in the Perceptual Associative Memory receives a greater activation. This node passes on this activation to other related nodes, called links. When this process stabilizes, the node group that received sufficient activation is referenced as a Percept and moved to the Current Situational Model. During this stabilization phase, the Percept is seen as a group of ontology elements relevant to the stimulus. It is written as a binary vector called a cue, that is later used to search for an autobiographical memory, a declarative memory, and a transient episodic memory. The transient episodic memory has information about events that the agent remembers. The declarative memory has information learned during lifetime, including some pieces of information which were once in the transient memory. The search return new cues, that are in turn used for a new search. This process repeats itself until a new information is returned.

During the interpretation phase, the percept created and stabilized in the perception phase is copied to the long term working memory, in a workspace secondary partition, where it will join older percepts. Attention Codelets analyze long term working memory content, searching for interesting elements. Coalitions of attention codelets and percepts are formed whenever it is possible. The coalition with greatest activation wins and its content is moved to the Global Workspace to be broadcasted.

In the action phase, the conscious broadcast goes to each agent’s subsystem, including the Procedural Memory. The Procedural Memory is a collection of organized schemes in the form of a scheme net. When the conscious broadcast provides information that com- bines the context of one or more schemes, the procedural memory will suggest schemes that should be copied to the action selection model. The action selection model, imple- mented as a Behaviour Net, will chose which action is most appropriated for the context in place. The action is sent to the Sensory-Motor Memory, that will execute the action in the environment.

In da Silva (2009), our research group investigated the LIDA model. This was our

first contact with the GWT. The work suggested that this mechanism is in fact capable of promoting an executive summary of the perception, and the automatization of new

3.5. Our Definition of Machine Consciousness 66

Related documents