Minds and Computers : An Introduction to the Philosophy of Artificial Intelligence
partially tokened, then the categories they represent can admit of
Download 1.05 Mb. Pdf ko'rish
|
document (2)
partially tokened, then the categories they represent can admit of imprecise borders and internal structure. Furthermore, if the content of representations is contextually modulated, then the extension of the category will be contextually sensitive. In other words, if representations are distributed and contextually modulated, then the categories they represent are such that there can be borderline cases of membership, the borders can shift contextually and there can be graded membership admitting of better and worse cases. To recap, advocates of distributed representation take the content conferring mechanism on representations to be essentially mediated by relations with other representations, the categories they represent to be contextually sensitive – allowing imprecise and shifting borders and internal structure – and the composition of mental representa- tion to be the complex, contextually modulated interaction of pat- terns of activation in a highly interconnected network. 18.4 COGNITIVE ARCHITECTURE So far in this chapter I’ve discussed two distinct views of mental rep- resentation and used this distinction as an entryway into understand- ing the competing symbolic and connectionist paradigms in artificial intelligence research. These di ffering views concerning mental representation are of central importance in distinguishing between the two paradigms but 185 they do not exhaust the di fferences between them. Connectionists also di ffer from their symbolic counterparts with respect to views concerning cognitive architecture. The term cognitive architecture refers to the structure and nature of the information processing systems of a cognitive agent. In other words, the term refers to the organisational and implementational features of the computational hardware which facilitates cognition. The symbolic tradition in artificial intelligence research sees the cognitive architecture of the human mind as a physical symbol system. Connectionists, on the other hand, view human cognitive architecture in terms of connectionist networks which facilitate parallel distributed processing. In previous chapters we’ve seen numerous examples of how we might implement cognitive functions with symbol systems. Connectionist net- works, as we will see in the following chapter, are particularly well suited to carrying out functions that are notoriously di fficult to implement in symbol systems architecture. To the extent that connectionist architecture is readily amenable to implementing functions which we take to be importantly constitutive of cognition and which prove problematic to implement with symbol systems, we have at least one reason for preferring a connectionist approach over a symbolic approach to artificial intelligence. The following chapter will be devoted to making clear the concepts which have so far only been mentioned with little in the way of expla- nation. After explaining these concepts and exemplifying the oper- ations of connectionist networks with numerous examples, we will then return to further discuss the relation between the symbolic and the connectionist paradigms. 186 C H A P T E R 1 9 ARTIFICIAL NEURAL NETWORKS The connectionist paradigm in artificial intelligence research rose to prominence in the last two decades of the twentieth century. Artificial neural networks were shown to be quite e fficacious in modelling certain cognitive phenomena that had been problematic to implement with symbolic computational architecture. The operations of artificial neural networks are designed to mimic the neural circuitry of the brain – they are often referred to as imple- menting ‘brain style’ processing. As such, it may aid your under- standing of this chapter to first revisit the discussion of the operations of neurons in Chapter 4. In this chapter we are going to develop a sound understanding of the operations of artificial neural networks and their utility in mod- elling cognitive functions. We’ll begin by describing the basic connec- tionist architecture and explaining how this di ffers from symbolic computational architecture. 19.1 CONNECTIONIST ARCHITECTURE Classical symbolic computational architecture – which we described at length in Chapters 7 to 9 and have seen many examples of since – admits of the following essential features. Firstly, there is only one processor in the architecture – a central pro- cessing unit (CPU) which processes program instructions. Secondly, the CPU carries out these instructions serially – one after the other. Thirdly, the CPU addresses and operates on localised register contents. Connectionist architecture, on the other hand, is crucially distinct with respect to each of these features. Connectionist networks are composed of a (typically large) number of simple processing units (nodes) which operate in parallel rather than serially. Content in con- nectionist networks is not local and addressable, but distributed across numerous nodes and encoded as a pattern of connections. 187 The basic elements of an artificial neural network are simple pro- cessing units which are designed to emulate the operations of indi- vidual neurons. These units are functionally organised in layers – there will be an input layer of nodes and an output layer of nodes. There will typically also be a ‘hidden’ layer of nodes – these are neither input nor output units but serve to mediate between these layers. As you have no doubt determined, nodes are connected to each other. Precisely how they are interconnected defines various architec- tural variations which needn’t concern us much here. In networks of interesting complexity, each node will be connected to a large number of other nodes – just as individual neurons are connected to large numbers of other neurons. The simplest type of connectionist archi- tecture (or the most complex depending on how you look at it) is such that every node is connected to every other node in the network. Information processing in artificial neural networks is achieved through the propagation of activation along the connections through the network. Each node in the network has a level of activation which is influenced by the activation it receives from other nodes which are connected to it. We’re going to make some simplifying assumptions here about activation. Firstly, we’re going to assume that at each time step, the activation of a node is entirely determined by the activation it receives along its incoming (a fferent) connections (rather than consider a more complicated function which also takes into account the antecedent level of activation of the node from the previous time step). Connections between nodes can be either excitatory or inhibitory and this is represented by assigning a weight – a positive or negative numerical value – to each connection. Excitatory connections – which are positively weighted – will excite (increase the level of activation of) the node they are connected to. Inhibitory connections – which are negatively weighted – will inhibit (decrease the level of activation of) the node to which they are connected. Each node in the network, you will recall, is a simple processing unit. These nodes implement two functions – an activation function and a transfer function. The activation function determines whether or not a node will fire based on its level of activation at that time step. We’re only going to consider the simplest of activation functions – a threshold function. Nodes with a threshold activation function will fire i ff their level of activation at that time step is above some threshold value assigned to 188 the node. If a node fires, it passes activation along each of its outgo- ing (e fferent) connections to other nodes, otherwise no activation propagates through that node. The transfer function determines how a node updates its level of activation based on the activation it receives along its a fferent con- nections. Again, we’re only going to consider the simplest of transfer functions – a weighted sum function. To determine the level of acti- vation of a node with a weighted sum transfer function, we simply take the sum of the values of the a fferent connection weights. 19.2 SIMPLE ARTIFICIAL NEURAL NETWORKS Let’s take a look at some basic examples to exemplify these oper- ations. To keep things simple, I’m going to use integers for connection weights and threshold values. Figure 19.1 depicts the simplest artifi- cial neural network that does something interesting. This network has two input nodes (A and B) and one output node (C). We’re interested in whether or not the output node will fire (although its e fferent connection is not afferent to any other node). The input nodes we can imagine as detectors of some kind. They are set to fire if some environmental condition is met – perhaps if a light is on or if a switch is in a particular position. The two connections in the network are both excitatory and equally weighted. If A fires it excites C and if B fires it excites C. The 189 Download 1.05 Mb. Do'stlaringiz bilan baham: |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling