Autonomous Developmental Cognitive Architecture Based On Self-Organizing Incremental Neural Network For Robots

Cognitive robots are a kind of advanced agent that can perceive,act,and develop cognition autonomously through learning like humans.In recent years,they have received widespread attention in the field of artificial intelligence.Having the ability of developing cognition enables robots to autonomously cope with complex and dynamic environments as well as adapt to non-specific and multi-tasking scenarios.This helps robots improve their level of autonomy and human-robot interaction to serve humans better.Cognitive developmental models should be able to learn,memorize new classes of objects in real time,and summarize learned knowledge autonomously to develop class-based object concepts.However,tranditional cognitive developmental models of cognitive robots have some problems such as single mode,weak generality,pre-defined structure,offline training,inability to detect and feedback false representations autonomously and large-scale continuous data streams.This thesis focuses on the cognitive development of the object concepts in human daily life for robots,which contributes to establishing a common knowledge base between robots and humans and promoting human-robot interaction.Inspired by human infant cognitive developmental theories and physiology researches of knowledge representations in brain,this thesis establishes cognitive development models based on incremental self-organizing neural network.These models can enable robots to increase the amount of knowledge through online and incremental learning as well as develop class-based object concepts through its self-organizing ability to improve robots'cognition.In order to improve the learning effect of the cognitive developmental model and solve the above problems of tranditional cognitive robots,this thesis conducts a more in-depth study of the cognitive developmental model of robots on object concepts.The main contributions of this thesis are given as follows1)Cognitive developmental model based on self-organizing incremental neural network.Inspired by the psychological theories of human infant cognitive development and the physiological theories of objects conceptual representations in brain,a self-organizing incremental neural network is utilized as the main framework of a general cognitive developmental model for robots to learn multimodal knowledge.This model with network cascade can not only simulate the correlation between different modalities in the brain,but also realize the development of different levels of knowledge from the concrete sample representations to the abstract symbol representations.2)A hierarchical cognitive developmental architecture based on audio-visual fusion.Conventional learning methods for robots are unimodal,not universal and should train a pre-defined model by large datasets in an offline way.Moreover,the fixed similarity threshold of tranditional SOINN mentioned in Chapter 2 is not suitable for open-ended learning.To track these problems,a Self-Organizing Hierarchical Cognitive Architecture based on Audio-Visual Fusion(SOHCN-AVF)for robots is peoposed to learn objects'visual and audio information and build the associative relationships between two modalities online.It contains audio and visual two pathways and each pathway has three layers of self-organizing incremental neural network.The sample lalyer learns and self-organizes visual features as well as names in an unsupervised way,in which a dynamically adjustable similarity threshold strategy is proposed to allow nodes to cluster autonomously based on data without human pre-defined thresholds.The symbol layer is a bridge for information transfer between high and low layer networks.It can not only encode the clustering results extracted from sample layers into a concise symbol representation and pass it to the correlation layer,but also decode the signal of the association layer for sample layers.The association layer binds the audio-visual symbols activated simultaneously to establish the associative relationship between two modalities.It includes a top-down response strategy that allows the robot to automatically recall the associative modalities,resolve conflicting associations,and adjust knowledge.Experimental results on two objects datasets and a real task show that our architecture is efficient to learn and associate object's view with its name in an online way.What's more,the robot can autonomously improve its cognitive level by utilizing its own experience without enquiring to human.3)A self-organizing developmental cognitive architecture with interactive reinforcement learning.The cognitive developmental model in Chapter 3 acquires knowledge through passive perception,which may generate incorrect representations inevitably and cannot correct them online without any feedback.To tackle this problem,a biologically-inspired hierarchical cognitive system called Self-Organizing Developmental Cognitive Architecture with Interactive Reinforcement Learning(SODCA-IRL)is proposed.This model normalizes the hierarchical cognitive developmental model in Chapter 3 to a framework of interactive reinforcement learning It can learn object concepts online and test learning effects by interacting with humans in a parallel and interlaced manner.In order to realize the integration between two algorithms,a novel exponential memory model,controlled by two forgetting factors,is designed for individual nodes to simulate the consolidation and forgetting process of human memory.In addition,an interactive reinforcement strategy is proposed to provide appropriate rewards and execute mistake correction.The feedback acts on the forgetting factors to reinforce or weaken the memory of nodes,which can consolidate correct knowledge,correct and forget wrong representations.Experimental results show that the proposed method can make effective use of the feedback from humans to improve the learning effectiveness significantly and reduce the model redundancy.4)A self-organizing and reflective cognitive developmental model with lifelong learning.The cognitive methods in Chapters 3 and 4 may be confronted with large storage and computation consumption when learning continuous data streams in lifelong learning.Moreover,the adjustment of similarity threshold for SOINN may be affected by the data input sequence.To track these problems,a self-organizing and reflecting cognitive network(SORCN)is proposed by integrating SOINN with a modified clustering by fast search and find of density peaks algorithm(CFS).This model can learn increamentally and reflect the learned knowledge periodically to realize robotic lifelong cognitive development.The modified CFS clustering algorithm is utilized in the reflection process.We improves the traditional CFS clustering algorithm into three steps:clustering,merging,and splitting.The clustering step adopts an autonomous center selection strategy,which enables CFS to adapt to SOINN's online learning method.The merging step contains a strategy based on intra-class topology construction,which can realize the incremental clustering of CFS and promote the adjustment of SOINN's similarity threshold.This also contributes to improving the generalization ability of nodes,thereby reducing the storage consumption of the network.The splitting step uses a strategy based on intra-class center reselection to adjust the clustering results.The model also includes a competitive learning strategy based on the reflection result,which can quickly find the matching node of the input according to clustering centers without traversing the entire network.This can significantly reduce the computational consumption when processing large-scale data.Experimental results demonstrate that SORCN can achieve better learning effectiveness and efficiency over several state-of-art algorithms.