Abstract |
In the current thesis we have studied the cognitive process of observational learning from a
computational modeling perspective. In this context we have employed data from neuroscientific
experiments, including higher-level imaging and single-cell recordings, in order to develop two
computational models of observational learning, inspired by the neurophysiology of human and
Macaque primates. To accomplish this we have devised a framework for designing computational
models based on neuroscientific findings, and used it in order to develop two novel implementations of
the cortical process in simulated agents. To facilitate learning during observation, both models are based
on the intuition that, during action execution and observation, the activated cortical networks in the two
primates overlap extensively. As a result, both agents treat perception as an active, cross-modal,
simulation of others’ actions and learn new motor skills without the active involvement of their body.
The first model maintains adequate consistency with the relevant brain areas and connectivity in
Macaques, and effectively provides insights about the cortical underpinnings of observational learning,
which can be summarized in three categories: (i) neuronal, i.e. how learning can be implemented at the
cellular level during observation, (ii) regional, by identifying the potential role of a certain region in
associating the motor representation with the visual image of the observer, (iii) system, how the
emergent pattern of activations observed during action observation and action execution is formed, and
what are the reasons for the lower activations during observation. In addition, due to the use of the
aforementioned modeling methodology, the agent is able to exhibit three important behavioral
functions: (i) observational learning in a similar manner as its biological counterparts, (ii) knowledge
generalization to different domains and knowledge integration on top of existing representations and
(iii) embodiment correspondence based on the overlapping pathway of activations.
The second model employs a phenomenological approach to design a motor control system that is
loosely based on the function of the regions that become active in humans during execution and
observation. For this reason we have developed novel implementations for each of the subsidiary motor
control processes, and integrated them in order to produce an agent able to learn only by observation.
The main contributions include: (i) a model that replicates the reward prediction properties of the
dopaminergic neurons in the Basal Ganglia, used to implement a variant of reinforcement learning, (ii) a
way to segregate the multidimensional control of the embodiment of the agent to basis functions using
a novel primitive model, (iii) a method to implement embodiment correspondence using associative
Page iv
networks, which enables an agent to develop and match symbolic representations of its own body and
the demonstrator’s, (iv) how higher-order motor control can be designed as an epiphenomenon of the
motor control system, i.e. as a subsidiary process built on top of basis motor functions and (v) how
learning can be implemented during observation using simple motor rules that can be derived only by
observation
|