BINP Research Progress Report

I’d like to introduce the research I’ve been conducting recently!
Simply put, the goal is to help restore vision for visually impaired individuals by implanting a chip into the brain.
To achieve that, I’ve been solving various problems related to the optimization of simulations used for those experiments.
Details are written below—feel free to read through if you’re curious!

---

Research Progress Report

Stimulation-Driven Optimization of Visual Cortical Prosthesis

Author: Yongtae Kim
Advisor: Prof. Seung Woo Lee
Date: August 27, 2025

1. Research Overview

The core of visual cortical prosthesis (VCP) based on cortical stimulation lies in the joint optimization of placement and stimulation strategy.
Previous electrical stimulation studies have shown that dynamic sequence stimulation can induce shape perception, complementing the limitations of the “pixel synthesis” assumption (Beauchamp et al., 2020).
Recently, with the development of high-channel (hundreds to thousands) electrodes/microcoils and the alignment of individual brain anatomy and retinal maps, automated placement optimization to match target phosphene distributions has been proposed (Van Hoof et al., 2025).

This study aims toward a micro-coil–based VCP, with three goals:
(i) alignment using human data (HCP 7T) and future Macaca fascicularis data,
(ii) extension of the vimplant pipeline, and
(iii) establishment of pre-implant and post-implant placement strategies via stimulation–perception simulations.
The related simulation ecosystem (virtual patient, differentiable phosphene simulator) is also actively referenced (Fine & Boynton, 2024).
Additionally, this study emphasizes the spatial locality and circuit activation potential of micro-magnetic stimulation.

2. Progress

2.1 Pipeline Reproduction and Organization

• Reproduced vimplant and wrote preprocessing guides.
• Built an environment based on FreeSurfer/Neuropythy alignment using HCP 7T retinotopy (181 subjects) pRF maps for V1 retinotopy alignment (Van Hoof et al., 2025).

2.2 Electrode/Coil Grid Generation Module

• Implemented code/electphos.py::create_grid_v2, enabling independent specification of (x, y) spacing in a 7×7×5 grid.
• Added options for changing channel number/pitch of electrode sets (“comb”), tested on subject 102816.

2.3 Subset Selection (Channel Selection)

• Formulated the problem as subset selection based on similarity to the target phosphene distribution.
• Implemented greedy forward selection (constraint: one channel per shank). Observed that for fixed n, the constraint has minimal effect; the total channel count n dominates the loss behavior.

2.4 Loss Function Definition

• Defined the loss as $$ \text{Loss} = a \cdot DC + b \cdot Y + c \cdot H + d \cdot V $$.
  • DC: Sørensen–Dice overlap between target and predicted phosphene maps
  • Y: Yield (ratio of effective intra-cortical channels)
  • H: Hellinger distance (distribution similarity)
  • V: Vessel-avoidance term, computed by aligning mid-to-large vessel atlas to fsaverage (Viviani, 2016)
• In subset selection problems, ignoring Y and V simplifies the loss to roughly $$ \text{Loss} ≈ DC − H $$.

2.5 Speed Issues and Alternative Search Methods

• Bayesian Optimization was slow; alternatives under prototyping include CMA-ES, PSO, and TuRBO-BO.
  • TuRBO: Local BO using multiple trust regions, effective in high dimensions (Eriksson et al., 2019)
  • CMA-ES: Standard robust method for continuous, non-convex, derivative-free optimization (Hansen, 2016)
  • PSO: Simple global heuristic with few parameters (Kennedy & Eberhart, 1995)

3. Rationale of the Current Model

• Efficacy of dynamic stimulation: Dynamic sequences allow accurate recognition of letter forms by both sighted and non-sighted participants (Beauchamp et al., 2020).
• Need for placement optimization: Large-scale automated placements based on individual anatomy/retinotopy have been proposed, with loss functions including vessel-avoidance terms (Van Hoof et al., 2025).
• Virtual patient and differentiable simulators: End-to-end simulation from pulse to neural to perceptual stages accelerates design-space exploration (Fine & Boynton, 2024).
• Locality of microcoil stimulation: Localized and sustained activations reported in V2 when stimulating V1, suggesting sub-millimeter spatial resolution potential (Lee & Fried, 2022).

4. Preliminary Results

• n-channel reduction experiments (n = 1–100): The total number of channels has a stronger impact on loss than constraints such as “one channel per shank.”
• Grid resolution sweeps (e.g., 7×7×5): Adjusting spacing allows control of cortical coverage and eccentricity (azimuth–elevation) range.

5. Next Plans

5.1 Application to Monkey (Macaca fascicularis) Data

• Extend vimplant upon availability of T1/T2 and retinotopy/task fMRI data to test cross-species generalization.

5.2 Incorporating Real Comb Specifications

• Collect parameters (spacing, length, wire diameter, turns, waveform, current range).
• Use Sim4Life/COMSOL to derive induced E-field and heat profiles, adding them as constraints to the loss.

5.3 Algorithm Benchmark Comparison

• Compare Bayes, TuRBO, CMA-ES, and PSO under identical initialization and budgets using:
  (i) final loss, (ii) wall-clock time, (iii) inter-iteration variance, (iv) local-minima avoidance ratio.
  TuRBO and CMA-ES prioritized for initial tests.

5.4 Multi-Comb (≥2) Optimization

• Analyze trade-offs between inter-array interference, vessel penalties, and cortical coverage (target: extend eccentricity coverage).

5.5 Dynamic Stimulation Strategy Learning

• After placement is fixed, design dynamic stimulation trajectories via Bayesian Optimization or reinforcement learning to maximize shape-perception performance.