TL;DR
This paper introduces a novel computational model that simulates the development of optokinetic nystagmus (OKN) in a behaving organism. By integrating active efficient coding principles with detailed neural pathway modeling, it explains how both visual perception and eye movement control emerge and interact, offering a more biologically plausible explanation for OKN than previous approaches.
What Problem Does This Paper Address?
The paper tackles the challenge of understanding how the complex eye movement known as optokinetic nystagmus (OKN) develops in biological systems. OKN is crucial for stabilizing our vision when the environment moves relative to us. While previous models have attempted to explain OKN, they often relied on simplified representations of neural activity and didn't account for the crucial interplay between sensory processing and motor behavior development. Specifically, the authors aim to model the emergence of OKN, including its characteristic asymmetry in infants (monocular OKN or mOKN), and how this asymmetry resolves with the development of binocular vision and vergence control. They also investigate how developmental abnormalities, like strabismus, can lead to persistent OKN asymmetries.
Methods
The core of the model is the active efficient coding framework, an extension of Barlow's efficient coding hypothesis. This framework posits that not only do neural systems efficiently encode sensory data, but behavior itself evolves to shape sensory input for better encoding. The model simulates:
- Neural Pathway Development: It models the detailed connectivity from retinal input to the visual cortex and then to motor neurons controlling eye movements. This is a departure from previous models that used scalar representations of neural activity.
- Sensory Processing:
- Subcortical Pathway: Modeled as hardwired and functional from birth, it provides initial OKN control but is monocular, leading to initial mOKN asymmetry.
- Cortical Pathway: This pathway develops after birth. It uses sparse coding to learn receptive fields (basis vectors) from visual input, mimicking simple cells in the visual cortex. These learned receptive fields exhibit both disparity and motion tuning.
- Motor Control:
- OKN Control: Mediated by motor neurons in the nucleus of the optic tract (NOT), receiving input from both subcortical and cortical pathways.
- Vergence Control: Mediated by motor neurons receiving input solely from the cortical pathway.
- Learning Mechanisms:
- Cortical Sensory Neurons: Connections from retina to cortex are learned via sparse coding, where basis vectors are updated to minimize reconstruction error.
- Cortical to Motor Neurons (OKN): Learned via a Hebbian learning rule, with a regularizer that biases connections towards neurons with strong contralateral eye input, reflecting known ocular dominance effects.
- Cortical to Motor Neurons (Vergence): Learned using reinforcement learning (natural actor-critic algorithm).
- Behavioral Simulation: The model simulates eye movements (version and vergence) and how these movements influence the sensory input, creating a closed loop. A simplified re-centering reflex simulates the fast phase of OKN.
Datasets
The paper does not explicitly mention using pre-existing datasets. Instead, it utilizes a simulated environment developed for the iCub humanoid robot. This environment generates realistic image sequences based on simulated agent behavior and environmental motion. The visual input is derived from natural images, and the agent's visual field is modeled with foveal and peripheral regions, sampled at specific resolutions.
Key Insights and Ideas
- Joint Development is Key: The model strongly suggests that the development of sensory perception (e.g., motion and disparity tuning) and motor control (OKN and vergence) are not independent but co-develop and mutually influence each other.
- Behavior Shapes Perception: The active efficient coding framework highlights how an organism's behavior actively shapes the sensory input it receives, which in turn influences neural development.
- Cortical Pathway as the Driver of Symmetry: The transition from asymmetric infant mOKN to symmetric adult mOKN is primarily attributed to the development of the indirect binocular cortical pathway, which integrates information from both eyes.
- Strabismus's Impact: The model demonstrates how developmental disruptions in binocularity (simulated as strabismus) can lead to persistent mOKN asymmetry by interfering with the development of these crucial cortical connections.
- Subcortical Pathway as Scaffolding: The subcortical pathway, while functional at birth, appears to act as a developmental scaffold, guiding the learning of the more sophisticated cortical pathway, and becoming less influential later in life.
Key Results
- Qualitative and Quantitative Agreement: The model's behavior, particularly its OKN responses and asymmetries, qualitatively and quantitatively matches a range of experimental findings from the literature on both normal and strabismic development.
- mOKN Asymmetry Simulation: The model successfully reproduces the infant mOKN asymmetry (preferential response to temporal-nasal motion) and its resolution with normal binocular development.
- Strabismus-Induced Asymmetry: When simulating strabismus (fixed vergence), the model exhibits significant mOKN asymmetry, with reduced tracking ability for nasal-temporal stimuli. This asymmetry is quantified by the Nasal Bias Index (NBI) and Asymmetry Index (ASI), which align with experimental values.
- Ocular Dominance Histograms: The model's simulated cortical neurons show ocular dominance histograms that closely resemble those reported by Hubel and Wiesel for normal and squinting kittens, with a strong shift towards monocularity in the strabismic case.
- Effect of Disparity Statistics: The model predicts that the distribution of near vs. far disparities during development influences the degree of mOKN asymmetry, with balanced disparities leading to more symmetric OKN.
Limitations and Future Work
- Simplified Fast Phase: The model uses a simplified hard-wired re-centering reflex for the fast phase of OKN, rather than modeling its development. The authors suggest this could be a future extension.
- Full Field Stimulus: The current model primarily uses full-field stimuli. Extending it to small target tracking, similar to smooth pursuit eye movements, is proposed.
- Amblyopia: The authors suggest that the framework could be extended to study the effects of amblyopia.
- Learning Rule for OKN: While Hebbian learning is used for OKN, exploring reinforcement learning for this pathway is suggested.
- Realistic Environments: The current simulations use controlled environments. More complex environments with diverse depth contours and smaller objects could lead to more varied disparities and potentially different developmental outcomes.
- Temporal Dynamics: The model focuses on steady-state OKN characteristics and does not explicitly model transient responses.
Summary updated