Editorial: Neural plasticity for rich and uncertain robotic information streams

Andrea Soltoggio; Frank van der Velde

摘要

Models of adaptation and neural plasticity are often demonstrated in robotic scenarios with heavily pre-processed and regulated information streams to provide learning algorithms with appropriate, well timed, and meaningful data to match the assumptions of learning rules. On the contrary, natural scenarios are often rich of raw, asynchronous, overlapping and uncertain inputs and outputs whose relationships and meaning are progressively acquired, disambiguated, and used for further learning. Therefore, recent research efforts focus on neural embodied systems that rely less on well timed and pre-processed inputs, but rather extract autonomously relationships and features in time and space. The bio-inspired focus does not seek the most effective machine learning method to solve those problems, it rather points toward a better understanding of problem solving mechanisms in neural systems, which can in turn also provide viable solutions to difficult problems. Realistic models of plasticity must account for delayed rewards (Soltoggio et al., 2013a ), noisy and ambiguous data (Soltoggio et al., 2013b ), and emerging and novel input features during online and value learning (Krichmar and R?hrbein, 2013 ). Those factors have indeed been an emerging focus of search (e.g., Sporns and Alexander, 2003 ; Lungarella and Sporns, 2006 ; Martius et al., 2013 ), with a growing number of studies that cannot be reviewed in this short editorial. Such approaches model the progressive acquisition of knowledge by neural systems through experience in environments that may be affected by ambiguities, uncertain signals, delays, or novel features (Pugh et al., 2014 ; Soltoggio, 2015 ). This Research Topic in Frontiers in Neurorobotics explored fundamental properties and dynamics of neural learning systems that are naturally immersed in a rich information flow. We are pleased with the contributions collected in this Research Topic, each of which addresses key topics in this emerging and important field of research. One overarching problem in this field is that of making sense of large amounts of data from sensory systems in order to recognize particular situations and perform basic tasks. Parisi and colleagues took a self-organizing neural approach to action recognition using human pose-motion features. The Growing When Required (GWR) networks manifest a high-level structural plasticity that regulates network complexity in relation to the task (Parisi et al., 2015 ). Such a bio-inspired approach recorded state-of-the-art performance on a dataset of full-body actions captured with a depth sensor, with competitive results in a public benchmark of domestic daily actions. Another source of large, noisy and uncertain data is found in robotic tactile sensors. Chou et al. ( 2015 ) deployed a specific robot called CARL-SJR with a full-body tactile sensory area. CARL-SJR encourages people to communicate with it through gentle touch, and provides feedback to users by displaying bright colors on its surface. The time-delayed and uncertain nature of the interactions poses challenges to the formation of correct associations between stimuli, rewards and actions. The approach devised by Chou et al. ( 2015 ) experiments with a strongly bio-inspired architecture of spiking neurons with neuromodulated plasticity. The model abstracts brain areas such as the primary somatosensory cortex, prefrontal cortex, striatum, and the insular cortex to process noisy data generated directly from CARL-SJR's tactile sensory area. The result is a robust learning mechanism that reliably forms correct associations and preferences for directions without heavily pre-processed inputs. Uncertainty and large amount of data are also found in collaborative multi-robot scenarios in which multiple robots work alongside humans. Galbraith and colleagues propose a motor babbling approach to learn a complex set of relations and interactions with the 11-degrees-of-freedom RoPro Calliope mobile robot (Galbraith et al., 2015 ). Motor babbling of its wheels and arm enabled the Calliope to learn how to relate visual and proprioceptive information to achieve hand-eye-body coordination. Motor control is a problem in which neural plasticity results in high level of adaptation, adjusting neural systems to operate in combination with specific bio-mechanical structures and morphologies. (Burms et al., 2015 ) demonstrated the utility of modulated Hebbian plasticity in embodied computation for compliant robotics. In such scenarios, control policies are generally unknown due to the partial offload of control policies to morphological computation. Modulated Hebbian plasticity was shown to lead to hybrid controllers that naturally integrate the computations that are performed by the robot's body into a neural network architecture. Those results demonstrate the potential of universal applicability of plasticity rules to complex control problems. A similar problem was tackled in Dasgupta et al. ( 2015

机译：适应性模型和神经可塑性模型通常在机器人场景中进行演示，该场景具有大量经过预处理和调节的信息流，可为学习算法提供适当，适时且有意义的数据，以符合学习规则的假设。相反，自然场景通常充满原始，异步，重叠和不确定的输入和输出，它们的关系和含义逐渐被获取，消除歧义并用于进一步学习。因此，最近的研究工作集中在神经系统系统上，该系统较少依赖及时的和经过预处理的输入，而是自动提取时间和空间中的关系和特征。受生物启发的焦点并不是寻求解决这些问题的最有效的机器学习方法，而是指向对神经系统问题解决机制的更好理解，这反过来也可以为棘手的问题提供可行的解决方案。现实的可塑性模型必须考虑到延迟的奖励（Soltoggio等，2013a），嘈杂的数据（Soltoggio等，2013b）以及在线和价值学习期间出现的新兴输入特征（Krichmar和R？hrbein，2013））。这些因素确实已经成为搜索的新兴焦点（例如Sporns和Alexander，2003; Lungarella和Sporns，2006; Martius等，2013），并且越来越多的研究无法在本简短的社论中进行评论。这样的方法通过在不确定性，不确定信号，延迟或新颖特征可能影响的环境中的经验，通过神经系统对知识的逐步获取进行建模（Pugh等，2014; Soltoggio，2015）。《神经机器人学前沿》中的该研究主题探讨了自然浸入丰富信息流中的神经学习系统的基本特性和动力学。我们对本研究主题中所收集的贡献感到满意，每个主题都针对这一新兴的重要研究领域中的关键主题。在该领域中的一个首要问题是要从传感系统中获取大量数据，以识别特定情况并执行基本任务。 Parisi及其同事采用了一种自组织的神经方法来利用人类的姿势运动特征进行动作识别。需求时增长（GWR）网络表现出高层次的结构可塑性，可调节与任务相关的网络复杂性（Parisi等人，2015年）。这种受生物启发的方法在用深度传感器捕获的全身动作的数据集上记录了最先进的性能，并在国内日常动作的公开基准中具有竞争性结果。在机器人触觉传感器中发现了大量，嘈杂和不确定数据的另一个来源。周等。（2015年）部署了一种名为CARL-SJR的特定机器人，该机器人具有全身触觉感觉区域。 CARL-SJR鼓励人们通过温柔的触摸与之进行交流，并通过在其表面显示鲜艳的色彩来向用户提供反馈。交互作用的时间延迟和不确定性对刺激，奖励和行动之间正确的关联形成提出了挑战。周等人设计的方法。（2015年）进行了强烈的生物启发性的尖峰神经元具有神经调节可塑性结构的实验。该模型提取大脑区域，例如主要的体感皮层，前额叶皮层，纹状体和岛状皮层，以处理直接从CARL-SJR触觉感官区域生成的嘈杂数据。结果是一个强大的学习机制，可以可靠地形成正确的方向关联和偏好，而无需大量的预处理输入。在多机器人协作场景中还发现了不确定性和大量数据，在这种协作场景中，多个机器人与人类并肩工作。 Galbraith和他的同事提出了一种马达冒泡的方法，以学习与11自由度RoPro Calliope移动机器人之间的复杂关系和相互作用（Galbraith等，2015）。车轮和手臂的马达and动，使Calliope能够学习如何关联视觉和本体感受信息，以实现手眼与身体的协调。运动控制是一个问题，其中神经可塑性导致高度适应，将神经系统调整为与特定的生物机械结构和形态结合使用。 Burms等人（2015）证明了调制的Hebbian可塑性在顺应性机器人的体现计算中的效用。在这种情况下，由于部分控制策略将负载转移到形态计算中，因此通常未知控制策略。调制的Hebbian可塑性被证明可以导致混合控制器，该控制器自然地将机器人身体执行的计算集成到神经网络体系结构中。这些结果证明了可塑性规则普遍适用于复杂控制问题的潜力。 Dasgupta等人也解决了类似的问题。（2015年

Editorial: Neural plasticity for rich and uncertain robotic information streams

摘要

著录项

相似文献

相关主题

期刊订阅