Module 11: Introduction to Machine Learning in Connectomics

🤖 What is Machine Learning?

Machine learning (ML) is a method of teaching computers to learn from data. In connectomics, ML is used to automatically label parts of the EM volume such as cell boundaries and synapses.

• Supervised vs. unsupervised learning
• Training data and ground truth
• Applications in connectomics

🧠 Neural Networks and Segmentation

We focus on convolutional neural networks (CNNs), which are particularly effective for image tasks. Segmentation models learn to assign a class label to each pixel or voxel.

• What is a CNN?
• Basic architecture: layers, activation functions
• Loss functions for segmentation

🛠️ Hands-On Training

Use Python and TensorFlow/Keras to define a small segmentation network and train it on provided EM slices.

• Loading data and preprocessing
• Defining the model
• Training loop and model evaluation

🎯 COMPASS Integration

• Knowledge: ML pipeline and architecture
• Skills: Coding, debugging, model tuning
• Character: Patience, problem-solving, experimental rigor
• Meta-Learning: Learning how algorithms improve with data

📚 References & Resources

• Januszewski et al., 2018. Flood-filling networks. Nature Methods.
• Goodfellow et al., Deep Learning, MIT Press.
• Colab: "Intro to CNNs for EM Segmentation"
• BossDB Cookbook: BossDB Dataset Classes for PyTorch

✅ Assessment

• Describe how CNNs are applied to EM segmentation
• Train and test a simple ML model on labeled data
• Evaluate accuracy and interpret confusion matrices