Dopo un po’ di (felice, causa overdose tesistica non ancora smaltita del tutto) assenza dagli strumenti software di text editing, questa mattina (nonostante qui, tecnicamente, sarebbe Bank Holiday) sono tornato a prendere in mano uno di questi per mettere per iscritto un po’ di pensieri vari. La scusa era quella di preparare un abstract per un workshop del progetto al quale sono correntemente associato (al secolo VALUE), in programma per il prossimo 15 settembre in quel di Potsdam.
Sorvoliamo sul fatto che il limite di 200 parole e’ stato piu’ o meno bellamente (per quanto non volutamente) ignorato ed il computo totale recita circa 450. D’altronde, come direbbe l’Alberto Tomba dei bei tempi, chi mi conosce lo sa. Ad essere sintetico proprio non ce la faccio. Provo comunque a tagliare corto e chiudo il post copia/incollando il succitato abstract, nel caso in cui a qualcuno interessasse darci un’occhiata.
Towards a cognitive architecture for studying the effect of spontaneous eye movements during the memorising/rehearsal of complex instruction sequences
Fabio Ruini, Anthony Morse, Angelo Cangelosi
Centre for Robotics and Neural Systems, University of Plymouth, UK
The aim of the current work is to develop a cognitive architecture for humanoid robots capable of memorising sequences of instructions, taking into account the phenomena investigated in Apel et al. (in press).
The robotic platform used as reference is the iCub robot (Sandini et al., 2007), with preliminary investigations carried out using a computer-simulated version of it (Tikhanoff et al., 2008).
In the experimental setup, the robot is put in front of a table where a 3-by-3 grid (every cell being marked with a number) and several objects are deployed. The robot listens to a sequence of instructions in the form “move object x to grid cell y”, memorises the series, and eventually replays it.
The design of the cognitive architecture takes inspiration from the ERA architecture (Morse et al., 2010). Its topology relies on a network of four interacting two-dimensional Self-Organising Maps (Kohonen, 1990). Each (pre-trained) SOM encodes different elements of the perceptual space: a) the neck-eyes posture (4 DoF), b) the perceived colours (having the left eye fovea as reference), c) the names of the existing objects, and d) those of the various grid cells.
The robot goes through a sequence of learning stages to develop the desired capability. The first step consists in motor babbling, with the robot looking around the environment and creating bi-directional associations (through Hebbian learning) between the colours and the postures SOMs. During the second stage, the robot learns the names of the objects and of the grid cells in front of it. Creating links between the SOMs encoding for neck-eyes postures and object/grid cell names (also being helped by the connections established during the previous stage), it becomes possible for the robot to look at a specific object/cell when its name is heard. During the final stage, consisting in the exposure to the entire instruction sequence, a competitive dynamic is established within the object/cell names SOMs. When the robot listens to an instruction component (either an object name or a cell), its attention is directed towards it and this triggers the BMU in the corresponding self-organising map to increase its activation level. This activation decreases over time (if not regularly reinforced or in case alternative units get activated), thus leaving traces of recency in robot’s memory. When a following instruction component is heard, other than repeating this process a unidirectional inter-SOM connection is created going from the previous to the current instruction component, thus storing the “directionality” of the sequence. Once the entire process is done, the rehearsal of the instruction sequence is done by “reading” the activation levels from the object/cell names SOMs, and following the existing connections.