A predictive coding framework for a developmental agent: Speech motor skill acquisition and speech production

Social Reinforcement in Artificial Prelinguistic Development: A Study Using Intrinsically Motivated Exploration Architectures

“…Even though the new architecture has advantages over previous ones, future works should consider unstructured vocalizations that would allow to study canonical babbling that requires the production of supragottal consonants and more realistic speech perception. Shorter perception windows, using for example Mel Frequency Cepstral Coefficient as [8], must be considered. Finally, investigations must consider more realistic social scenarios attempting to cover other categories of maternal response and infants' vocalization directionality as defined in [16].…”

“…There exist some attempts to study artificial early vocal development as a mechanism to understand language emergence from an embodied developmental perspective [3][4][5][6][7][8]. However, most of the works aiming at studying artificial speech-based communication systems are rather focused on the natural language understanding problem.…”

“…Finally, in [6], the idea of an instructor was included, [4] had done some investigation regarding social emulation as a means of driving development, despite finding that social imitation was a crucial opportunity for development, [4] did not consider interactions neither realistic speech units. In [8], the authors made some progresses in the same line of research of [4], obtaining similar proportions of single vowels and sequence of vowels vocalizations as in [5]. [8] also studied the possibility of using Mel-frequency cepstral coefficients to model auditory signals instead of using the formant frequencies used here and in [4][5][6][7].…”

“…In [8], the authors made some progresses in the same line of research of [4], obtaining similar proportions of single vowels and sequence of vowels vocalizations as in [5]. [8] also studied the possibility of using Mel-frequency cepstral coefficients to model auditory signals instead of using the formant frequencies used here and in [4][5][6][7]. Finally, [7] considered a model that was also physically situated in a scenario where tool usage was being learned in parallel to vocal babbling, even though vocal learning was somehow more restricted compared to [4].…”

Learning, planning, and control in a monolithic neural event inference architecture

“…The predictive brain perspective and active inference principles have strongly influenced cognitive science over the last years (Bar, 2009;Butz and Kutter, 2017;Clark, 2016;Friston, 2009;Hohwy, 2013). Although predictive encodings have shown to yield promising results in artificial neural networks focusing on vision (Rao and Ballard, 1999), it remains highly challenging to realize these principles in scalable, temporal dynamic artificial neural network models, and particularly models that enable flexible, goal-directed planning (but see Najnin and Banerjee, 2017 for a recent, promising approach in phonological speech production). Moreover, it remains unclear how abstracted, hierarchical structures may be developed effectively (Botvinick and Weinstein, 2014;McClelland et al, 2010) -structures that are believed to be essential for enabling the generation of flexible, adaptive goal-directed behavior by means of hierarchical, model-based planning and reinforcement learning (Botvinick et al, 2009).…”