Flexible sequence learning in a SOM model of the mirror system - eScholarship
Serge Thill, Josef Behr, Tom Ziemke · Proceedings of the Annual Meeting of the Cognitive Science Society · 2012
Flexible sequence learning in a SOM model of the mirror system Serge Thill ([email protected]) Interaction Lab, School of Humanities and Informatics University of Sk¨ovde 54128 Sk¨ovde, Sweden Josef Behr ([email protected]) Institut f¨ur Kognitionswissenschaft, Neurokybernetik University of Osnabr¨uck 49069 Osnabr¨uck, Germany Tom Ziemke ([email protected]) Interaction Lab, School of Humanities and Informatics University of Sk¨ovde 54128 Sk¨ovde, Sweden Sequencing via ordinal nodes and conditions of satisfaction Abstract We present initial work on a biologically and cognitively in- spired model that may allow embodied agents to autonomously learn sequences of action primitives (forming an overall be- haviour). Specifically, we combine a flexible model of se- quence generation with a model of parietal mirror neuron ac- tivity. The main purpose is to illustrate that the approach is viable. Although further work is needed to improve the re- sults sketched out here, the concept is sound and relevant both to efforts in modelling mirror neuron activity and enabling ar- tificial embodied agents to autonomously learn sequences of action primitives. Keywords: Behavioural sequence learning; Ordinal node model; Self-organising maps; Mirror neurons Introduction We are concerned with the problem of generating sequences of action primitives which are flexible with respect to the pre- cise time it takes to execute the different components (primi- tives) of the same sequence at different times. A thorough dis- cussion of the issue is given, for instance, by Sandamirskaya & Sch¨oner (2010). In a nutshell, part of the problem is that one cannot simply chain together the different primi- tives through, for example, simple Hebbian learning. Rather, mechanisms must exist for keeping track of the current loca- tion in the sequence, including ways of verifying that the cur- rent action has successfully completed or failed to complete. Sandamirskaya & Sch¨oner (2010) describe a general frame- work which can address these issues and we briefly sketch the main points in the next section. Overall, the aim of the work in the present paper is to com- bine said framework with a model of parietal mirror neuron activity (Thill et al., 2011) and to illustrate that such an ap- proach is, in principle, viable. Importantly, since the mirror neuron model used here autonomously organises itself, the work proposed here may be relevant and helpful in designing artificial embodied agents that should autonomously learn se- quences of actions and use them to predict actions of others. The gist of the framework by Sandamirskaya & Sch¨oner (2010) is the existence of ordinal nodes which essentially count through the sequence. These nodes are implemented via coupled dynamical systems (see Methods), designed so that only one node can be active at a time. Upon comple- tion of the element of the sequence represented by the ac- tive node, activation is passed onto the next node in the se- quence. In their work (e.g. Sandamirskaya & Sch¨oner, 2010; Sandamirskaya et al., 2011), the action primitives forming the sequence exist in the sensorimotor representation of an em- bodied agent, implemented using techniques from Dynamic Field Theory (Sch¨oner, 2009; Spencer et al., 2009). This has the advantage that the sensorimotor representations of these primitives are stable (since they are essentially stable fixed- point attractors), which makes it particularly simple to link specific locations in the dynamic fields representing the sen- sorimotor space of the agent to specific ordinal nodes. Part of the challenge of the work presented in the present paper is to illustrate that the ordinal node system could also be at- tached to a representation with more noise and less stability than dynamic fields. The decision that a given action primitive has completed is implemented a separate system (also exploiting dynamic fields) that checks for a Condition of Satisfaction (CoS). One of the open challenges here is the question of how to best learn the CoS for specific primitives (including identifying that the primitive has, for whatever reason, failed). It is not the purpose of the present work to address the open issues regarding the CoS - rather, we focus on combining the ordinal node model with a model of mirror neuron activity discussed in the next section. Mirror system sequences One example of sequencing in biology is given by the hy- pothesised functioning of the mirror system. Without enter- ing the debate on what higher-level cognitive abilities mirror