Towards Biological Plausibility of Vocal Learning Models: a Short Review

Silvia Pagliarini, Arthur Leblois, Xavier HinautHAL 2018

TL;DR

This paper reviews computational models of vocal learning, focusing on how they can be made more biologically plausible. The authors highlight the common components of these models—motor control, sensory systems, and learning architectures—and discuss how different approaches to these components can lead to more realistic simulations of how humans and animals learn to vocalize.

What Problem Does This Paper Address?

Learning to vocalize, whether it's a baby learning to speak or a songbird mastering its song, is a complex process that involves intricate coordination between motor commands and sensory feedback. While computational models have made strides in simulating aspects of this learning, many fall short of capturing the full biological reality. Specifically, there's a need to bridge the gap between simplified models and the nuanced neuro-anatomical and developmental constraints observed in biological systems. This paper addresses the challenge of building vocal learning models that are not only functional but also biologically plausible, drawing inspiration from how brains actually learn and produce sounds.

Methods

The paper provides a review of existing sensorimotor models for vocal learning, categorizing them based on their core components:

  • Motor Control: This involves mapping motor commands to sensory outputs. Models differ in how they represent the vocal apparatus (e.g., using articulatory parameters like tongue height or simplified models like sound sources and resonators) and how they generate sound.
  • Sensory Space: This represents the perceived sound, often characterized by features like formants (resonant frequencies of the vocal tract) or raw audio waveforms.
  • Learning Architecture: This is the mechanism by which the model learns to associate motor commands with desired sensory outcomes. The paper discusses various architectures, including:
  • Inverse Models: These aim to generate the correct motor command for a given perceptual goal.
  • Forward Models: These predict the sensory outcome of a given motor command.
  • Reinforcement Learning: Models that learn through reward signals, often driven by an internal goal.
  • Neural Networks: Various types of neural networks, including recurrent neural networks (RNNs) and simpler feedforward networks, are employed to connect different spaces.
  • Learning Rules: The paper touches upon different learning rules used to update model parameters, such as Hebbian learning, STDP (Spike-Timing-Dependent Plasticity), and reinforcement learning algorithms.

The authors emphasize the importance of considering neuro-anatomical and developmental constraints, although the review itself focuses on the functional components of the models.

Datasets

Not discussed in the paper. The paper reviews existing models and does not present new experimental results or datasets.

Key Insights and Ideas

  • Common Building Blocks: Despite the diversity in models, there's a recurring set of components: a motor space, a sensory space, a perceptual space, and a learning architecture that connects them. Understanding these commonalities is crucial for comparing different approaches.
  • The Role of Internal Models: The concept of "internal models" (both inverse and forward) is central to explaining how the brain can link sensory feedback to motor commands for adaptive learning. This is a key biological mechanism that many computational models aim to replicate.
  • Biological Plausibility is Key: The authors argue that for models to truly explain vocal learning, they must incorporate biological constraints, such as developmental stages and neuro-anatomical structures. This is where many current models fall short.
  • Exploration is Necessary: Motor exploration, whether random or goal-directed, is a fundamental aspect of learning. Models need to account for how agents explore their motor capabilities to discover effective vocalizations.
  • Challenges in Comparison: The diversity in how sensory and perceptual spaces are defined, and the varying complexity of motor control and learning architectures, makes direct comparison between models difficult.

Key Results

This paper is a review, not an empirical study, so it does not present quantitative results in terms of performance metrics like accuracy or learning speed. Instead, its key "results" are the insights derived from synthesizing the literature:

  • A comprehensive summary of sensorimotor models for vocal learning, presented in a table that details their components (motor space, motor control, sensory space, perceptual space, learning architecture).
  • An analysis of how different models implement motor control, sensory representation, and learning architectures, highlighting commonalities and divergences.
  • An identification of the challenges in creating biologically plausible vocal learning models, particularly concerning the integration of neuro-anatomical and developmental aspects.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that it's difficult to compare models due to the varied definitions of their components and the lack of a universal framework.
  • The review does not delve deeply into the developmental aspects of vocal learning, which is a crucial area for biological plausibility.
  • Some models simplify or omit certain components (e.g., motor control), leading to a direct coincidence between motor and sensory spaces, which is biologically unrealistic.
  • Future Work:
    • The authors suggest the need for a common model that can address multiple questions and facilitate the comparison of different experimental approaches.
  • There's a call to incorporate more general biological mechanisms, including developmental aspects, into future models.
  • The paper implies that future research should focus on bridging the gap between existing models and the complex biological realities of vocal learning.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{pagliarini2018towards,
  author = {Silvia Pagliarini and Arthur Leblois and Xavier Hinaut},
  title = {{Towards Biological Plausibility of Vocal Learning Models: a Short Review}},
  booktitle = {HAL},
  year = {2018},
}
Figure