Connectionism and Cognitive Architecture Presented By: Abhishek Kumar Mall Rohit Gupta Abstract z Methodological questions about levels of explanation z Why connectionists and classical theories are incompatible? z Arguments for classical architecture z Lure of connectionism What is Connectionism? z Neural Networks or Neural Nets z Semantic Information z Structure Independent z Distributed pattern Classical Approach z String of Symbols z Explicit rules z Central controller z Digital Processing Theories of mind z Representationalist z Eliminativist Level of explanation z World has casual relation at many levels of analysis. z Any level at which states of system are taken to encode properties of world counts as cognitive level. Nature of the dispute z Assigning of semantic contents z Primitive relations z Architectural difference: – Combinatorial syntax and semantics for mental representations – Structure sensitivity of the process Complex Mental Representations z Role of labels in Connectionist theories z Connectionist network and graph structures z Distributed representations Structure Sensitive Operations z Learning z Reasoning Need for Symbolic Systems z Productivity z Systematicity z Compositionality z Inferential Coherence Productivity z Unbounded capacitites of a representational system z Connectionism cannot represent productivity z Only symbolic system can represent Systematicity of cognitive representations z Learning a part without learning the rest z Mastery of syntax of language z Thought is systematic z Ex: John loves the girl Compositionality of representations z Semantical relatedness and systematicity z Parallelism between syntactical structures and their entailments z Compositionality implies expressions have constituents Systematicity of inference z Exploiting the fact of combinatorial semantics and constituent structure z Ex: P&Q&R->P&Q->P z Connectionist can model the above example, also the other inference Summary z Argument from systematicity z Argument from compositionality z Argument from influential coherence The Lure of Connectionism z Rapidity of cognitive processes z Achieving large-capacity of pattern recognition z Sensitivity to damage and noise z “rule governed” and “exceptional” behavior Replies to the arguments z Objections are not intrinsic z Implemented on present computers Replies to the arguments( contd..) z Parallel computation and issue of speed z Resistance to noise and physical damage z Explicitness of rule Connectionism as Implementation z Virtual machine z Implemented in a certain kind of network z More of a theory of cognition than a theory of implementation Conclusion z Holding out for connectionism than classical view z Structured mental representation with an associationistic account z Treat connectionism as an implementational theory.