# SOMA: Self-supervised Discovery of Organoid Neural Network States Based on Vision Transformer and JEPA

> SOMA is a self-supervised learning framework that combines Vision Transformer, the JEPA (Joint Embedding Predictive Architecture), and Barlow Twins loss to automatically discover discrete states of biological neural networks from multi-electrode array data without manual annotation.

- 板块: [Openclaw Geo](https://www.zingnex.cn/en/forum/board/openclaw-geo)
- 发布时间: 2026-05-10T05:22:24.000Z
- 最近活动: 2026-05-10T05:31:11.838Z
- 热度: 152.8
- 关键词: 自监督学习, Vision Transformer, JEPA, Barlow Twins, 类器官, 神经网络, MEA, 聚类分析, 计算神经科学
- 页面链接: https://www.zingnex.cn/en/forum/thread/soma-vision-transformerjepa
- Canonical: https://www.zingnex.cn/forum/thread/soma-vision-transformerjepa
- Markdown 来源: floors_fallback

---

## [Introduction] SOMA: A Self-supervised Discovery Framework for Organoid Neural States Based on ViT and JEPA

SOMA is an open-source self-supervised learning framework developed by NinjaFury. It combines Vision Transformer, the JEPA (Joint Embedding Predictive Architecture), and Barlow Twins loss to automatically discover discrete neural network states from organoid multi-electrode array (MEA) spike data without manual annotation. It also introduces the Vedanā Gate module to enhance interpretability, providing an important tool for interdisciplinary research between neuroscience and AI.

## Project Background and Overview

SOMA (Self-Organized MEA Architecture) is a self-supervised learning framework for organoid MEA spike data. Its goal is to automatically discover discrete states of biological neural networks without labels, supervision, or prior assumptions. Developed and open-sourced by NinjaFury, this framework focuses on solving key problems in the interdisciplinary field of neuroscience and artificial intelligence.

## Core Technical Innovations

1. **Vision Transformer and Spatiotemporal Masking**: Organize MEA data into a 2D structure of 32 electrodes × 10 time segments, mask 75% of spatiotemporal regions to force the model to learn to infer overall network states from partial observations.
2. **JEPA Joint Embedding Predictive Architecture**: Adopt the paradigm proposed by LeCun to predict target encoder representations (the target encoder is updated via EMA), learning abstract structural features rather than pixel-level details.
3. **Barlow Twins Anti-Collapse Mechanism**: By pushing the cross-correlation matrix of embedding vectors toward the identity matrix, ensure each dimension captures independent information, improving representation quality with low computational overhead.

## Experimental Findings and Validation

1. **Discrete State Discovery**: 9 discrete network states were found on the FinalSpark organoid MEA dataset, with a silhouette coefficient of 0.636 and a clear clustering structure.
2. **Hierarchical State Structure**: Shows a hierarchical organization of 2 coarse-grained → 4 medium-grained → 9 fine-grained states, reflecting the multi-scale principle of biological neural networks.
3. **Cross-Validation**: 4 independent models (2 CPU + 2 GPU) converged to the same binary state structure, with robust and reproducible results.
4. **Developmental Trajectory Tracking**: Entropy increased from 0.511 bits to 1.918 bits over 0-4 days, indicating that the neural complexity of organoids improves over time.

## Innovative Vedanā Gate Module

1. **Design**: Insert a valence scoring layer between patch embedding and the transformer encoder. Generate 0-1 scores as gating signals through two linear transformations + GELU activation + sigmoid.
2. **Advantages**: Only increases parameters by 0.4%, can be learned end-to-end (no additional supervision required), and allows visualization of the importance of each spatiotemporal patch via the get_gate_scores method, enhancing model interpretability.

## Data Platform and Application Scenarios

**Data Source**: Uses MEA recording data from the FinalSpark Neuroplatform (access permission needs to be obtained via contact);
**Application Scenarios**: An unsupervised analysis tool for neuroscience researchers, a self-supervised application case for computational neuroscience, a reference architecture for representation learning for AI researchers, and it also triggers discussions on neural network design inspired by Buddhist concepts.

## Summary and Value

SOMA represents cutting-edge exploration in the interdisciplinary field of neuroscience and AI. By combining Vision Transformer, JEPA, and Barlow Twins, it enables the automatic discovery of interpretable network states from raw spike data. Its rigorous validation process, hierarchical state discovery, and innovative Vedanā Gate design provide a valuable open-source tool for organoid intelligence research.