Project Description
Abstract:
This project explores biomimetic training for vision models, inspired by the progressive development of the human visual system in early infancy. During the first months of life, human vision undergoes significant changes in acuity, color perception, spatial resolution, and noise tolerance. By emulating these developmental phases, this research investigates how vision models can be trained in stages that mirror biological maturation, potentially improving robustness and adaptability. Students will design biologically-inspired training regimens and evaluate visual model performance under various challenging conditions, such as adversarial perturbations, environmental noise, and optical turbulence.
Objectives:
- Develop Biomimetic Training Regimens – Design vision model training schedules that simulate developmental changes in visual fidelity, including resolution, color sensitivity, and noise levels.
- Optimize Learning Trajectories – Explore how progressive exposure to increasingly complex visual stimuli affects convergence, generalization, and robustness of vision models.
- Evaluate Robustness Under Degradations – Assess model performance under conditions such as noise, blur, turbulence, and adversarial attacks to determine the impact of biomimetic training.
- Compare Against Baselines – Benchmark against standard training approaches to identify whether biomimetic strategies confer advantages in generalization, efficiency, and robustness.
Methodology:
- Implement vision model training pipelines that simulate different stages of visual development, starting with low-resolution, high-noise inputs and gradually increasing fidelity.
- Experiment with various convolutional neural network (CNN) and vision transformer (ViT) architectures to evaluate adaptability to biomimetic schedules.
- Apply evaluation metrics such as classification accuracy, robustness to perturbations, calibration under uncertainty, and adversarial resilience.
- Introduce structured visual degradations (e.g., noise, blur, occlusion) during evaluation to assess the durability of learned representations.
Expected Outcomes:
This project aims to uncover how biologically-inspired visual learning trajectories can improve training efficiency and robustness in vision models. Outcomes may inform new training paradigms that are both computationally efficient and resilient to real-world noise and distortions. Insights from this research could contribute to more adaptable vision systems for use in autonomous agents, low-light imaging, and adversarial settings.
Please note that weekly meetings, even if brief, are expected to ensure consistent updates and contribute to a successful outcome
Project Description
Abstract:
This project explores biomimetic training for vision models, inspired by the progressive development of the human visual system in early infancy. During the first months of life, human vision undergoes significant changes in acuity, color perception, spatial resolution, and noise tolerance. By emulating these developmental phases, this research investigates how vision models can be trained in stages that mirror biological maturation, potentially improving robustness and adaptability. Students will design biologically-inspired training regimens and evaluate visual model performance under various challenging conditions, such as adversarial perturbations, environmental noise, and optical turbulence.
Objectives:
Methodology:
Expected Outcomes:
This project aims to uncover how biologically-inspired visual learning trajectories can improve training efficiency and robustness in vision models. Outcomes may inform new training paradigms that are both computationally efficient and resilient to real-world noise and distortions. Insights from this research could contribute to more adaptable vision systems for use in autonomous agents, low-light imaging, and adversarial settings.
Please note that weekly meetings, even if brief, are expected to ensure consistent updates and contribute to a successful outcome