چکیده مقاله
The present study aims to design and validate a scalable cognitive architecture for artificial intelligence platforms based on large language models Large language models, with billions of parameters and the ability to generate fluent and coherent text, have created a profound transformation in the field of natural language processing However, these models have fundamental limitations in terms of cognitive architecture structures that simulate cognitive processes such as memory, attention, reasoning, and learning : limited working memory processing short context windows , inability to perform symbolic and causal reasoning, lack of incremental learning requiring complete retraining for new data , very high computational cost for scaling, and lack of transparency and explainability This research employed a mixed method approach systematic review, algorithmic design, and computer simulation First, a systematic review of ۱۵۰ reputable articles identified the strengths and weaknesses of existing cognitive architectures Subsequently, a novel cognitive architecture was designed with six main components: dynamic working memory with a variable window, hierarchical semantic memory with incremental update capability, temporal episodic memory, hybrid reasoning engine including symbolic, causal, and probabilistic reasoning , multi level attention mechanism, and an explainer interface The proposed architecture was implemented on a cloud platform with ۶۴ processing nodes and tested using a standard dataset Findings showed that the proposed architecture, compared to the baseline model, achieved an average of ۳۵% improvement in causal reasoning accuracy, ۴۲% improvement in long term memory information retrieval after ۱۰۰۰ training steps , ۲۸% reduction in explainability error, and ۴۰% reduction in computational cost through variable window and incremental learning Additionally, the scalability of the architecture response time as a function of the number of parameters and processing nodes followed an approximately linear pattern coefficient ۱ ۲ The discussion emphasizes the necessity of transitioning from mere large language models to hybrid cognitive systems combining deep learning with symbolic and reasoning components It is recommended that a National Program for the Development of Cognitive Artificial Intelligence Systems be formulated and implemented with the participation of universities, knowledge based companies, and governmental institutions
کلیدواژهها
نویسندگان
Reza Rahnama
M S Graduate, Software Engineering Computer , Islamic Azad University, Borujerd Branch Field of work: Education Office Staff, Borujerd
شیوه ارجاع
Rahnama، Reza،1404،Designing a Scalable Cognitive Architecture for Large Language Model-Based Artificial Intelligence Platforms،دومین همایش بین المللی تحولات نوین در علوم تربیتی، روانشناسی و آموزش و پرورش،ارومیه
ارائهشده در
همایش ملی نهج البلاغه حکمرانی عمومی و خصوصی در مسیر ایران آینده15 بهمن 1404 · چالوس