Program

tuesday, 3 december

14h00 Simon Thorpe Finding repeating patterns in spiking activity : The fundamental mechanism underlying Cognitive Development and Learning?
14h55 Jeffrey Krichmar A Neurobiological Schema Model for Contextual Awareness in Robotics
15h50 Pause
16h10 Yasuo Kuniyoshi Emergence and Development of Embodied Behavior/Cognition in Simulated
Human Fetus/Infant
17h05 Diogo Pata How do robots perform spatial encoding: insights from hippocampal mechanisms

wednesday 4 december
9h30 Coffee
9h45 Gregor Schoener Toward pervasively neural architectures of embodied and higher cognition
10h40 Emmanuel Dupoux
11h45 Nicolas Rougier Boostrapping cognition

A Neurobiological Schema Model for Contextual Awareness in Robotics
Jeffrey L. Krichmar and Tiffany J. Hwu
Department of Cognitive Sciences, Department of Computer Science
University of California, Irvine, Irvine, CA 92697-5100, USA

A robot operating in multiple settings must develop stable as well as flexible representations of the tasks and contexts associated with their environments. Taking inspiration from neurobiology, we apply a neural network model of schemas and memory encoding to train a human support robot to find and retrieve objects found in a typical household or school. We define schemas to be collections of objects bound together by a common context. Because the model develops schema representations for each room in its environment, the robot can rapidly learn retrieval tasks associated with familiar schemas and disambiguate the tasks by context without interfering with prior schemas. Our model suggests that this occurs through two processing streams of indexing and representation, in which the medial prefrontal cortex and hippocampus work together to index cortical activity. Additionally, our study shows how neuromodulation contributes to rapid encoding within consistent schemas. For robots designed to aid humans in day to day chores, context is important when determining how to carry out an assigned task. The present work demonstrates how ideas from memory models in the brain may improve robotic applications and issues in artificial intelligence, such as catastrophic forgetting and lifelong learning.

How do robots perform spatial encoding: insights from hippocampal mechanisms
Diogo Pata
Institute for Bioengineering of Catalonia

Do you remember that moment when you and your friends hiked up the hill? Animals continuously navigate their surroundings and still are capable of separating everyday life moments and remember them as unique experiences. Furthermore, when revisiting a familiar place, animals are able to detect environmental modifications. However, their robotic counterparts still encounter difficulties in setting up mnemonic boundaries and detect changes in their environment. In this talk, I will present physiological results and computational mechanisms based on the hippocampal circuitry allowing for episode separation, novelty detection, and visual discrimination. 

Boostrapping cognition

Nicolas Rougier

Labri Team INRIA Bordeaux

Brain research has made tremendous progress over the last few decades in nearly all areas of investigation and yet, the comprehension of cognition still eludes us because most high-level functions, such as decision making, results from the complex and dynamic interaction between several structures. On one hand we have a fragmented collection of computational models whose unification is out of reach and on the other hand, we have holistic approaches whose complexity obliterates any hope of comprehension. In order to explain how cognition can develop in most vertebrates, we posit for a general bootstrapping learning mechanism relying on basal ganglia teaching cortical circuits that is actually a late feature linked with the development of a specialized cortex or pallium that evolved in parallel in different taxa. Informed and guided by neuroscience and philosophy we therefore propose to define and to build a series of computational, embodied and intelligible models along both the phylogenetic and ontogenetic axes.

Toward pervasively neural architectures of embodied and higher cognition

Gregor Schöner
Institut für Neuroinformatik

Classical cognitive architectures built on symbol manipulation, enabling productivity and abstraction. The challenge was to link such architectures to the real world, both through sensory-motor grounding of cognition and by capturing background knowledge. Recent advances in deep learning suggest that this link could be provided by intelligent filters, who can be optimized for particular tasks or environment.

I will criticize that perspective and emphasize instead that cognition emerges from sensory-motor processes, both evolutionarily and developmentally. Embodied cognition unfolds in closed loop as an organism is situated in structured environments. Cognitive processes must, therefore, afford graded states, continuous metric content, continuous time, and, critically, stability. The strong embodiment hypothesis postulates that all cognition processes share these properties. The challenge is then to understand how embodied cognitive processes may achieve seemingly abstract and high-level competences such as reasoning, conceptual thinking, or language.
 

Finding repeating patterns in spiking activity : The fundamental mechanism underlying Cognitive Development and Learning?

Simon Thorpe

CerCo, Université Toulouse 3, CNRS

Recent experiments have shown that humans are able to detect repeating stimuli in streams of images presented at up to 120 frames per second (Thunell & Thorpe, 2019). And a similar phenomenon exists in the auditory domain. Such data presents a challenge for most models of learning, and is certainly not something that could be explained by the error backpropagation learning paradigms that are the basis of the current vogue for Deep Learning. However, we have found a simple algorithm based on Spike-Time Dependent Plasticity that can be implemented in inexpensive hardware, and could be used in a wide range of areas including robotics.