Towards an Embodied Understanding of Neural Computation

Abstract

Information processing systems (IPSs) that behave in the real-world are constantly bombarded with noise from the environment. Although the real-world offers noise for free, this extrinsic noise source has a cost associated to it. The problem is related to the fact that the environmental intricacies (i.e. noisy distribution) in which the IPS samples from is everchanging. Consequently, the IPS is faced with the conundrum of maintaining stability in a dynamic environment, while at the same time, remaining flexible so as to match its intrinsic timescales with the timescales of the environment. Here, we propose conjoining three ingredients for solving the timescale incompatibility issue between IPSs and the environment. First, we propose evolving IPSs in an online fashion such that the system operates on-thefly. Second, we propose the implementation of online operations in robotic hardware – a methodological tool for allowing the IPS to provide feedback during its interaction with the environment. Finally, we propose learning from brain mechanisms as a source of inspiration for building more flexible and adaptive IPSs. In chapter one, we initiate our first attempt towards achieving flexibility in these systems. To this end, we study the interaction between short-term plasticity (STP) and spike-timingdependent plasticity (STDP), two important plasticity rules we’ve learned from the brain. As such, we construct a microcircuit motif of two units, and show in simulation, how each unit can discriminate the position of a moving stimulus. In chapter two, we study synaptic plasticity in the context of an online robotic domain. To do so, we increase network size to six units, and endow the circuit with STP as a candidate mechanism for microcircuit sensitivity to inputs. Here, we study motion discrimination using a Raspberry Pi microcontroller as the information processing unit. We also use a stationary camera to process images from the real-world. Finally, we attach two LED light sensors for providing feedback of how the system is behaving. Results show that the agent is capable of discriminating the direction of a moving stimulus. In the final chapter of the thesis, we move away from the static online robotic implementation, towards a more dynamic setting. In doing so, we develop a keyboard listener for online mobile robot control. Here, the motor trajectory of the robot is directly linked to network activity of 500 units. Furthermore, the agent is placed in an ecological context where it interacts with a human subject. During human-robot interaction, the motor trajectory of the robot is studied, enabling the human to make inferences about how neural computation is unfolding on-the-fly. The robot illustrates useful properties, one of which is high degree of flexibility and adaptation to ongoing input streams. Overall, we conjoin the three ingredients mentioned above as a framework for solving the timescale incompatibility issue between IPSs and the environment.