---
tags:
- neuroscience
- spike-sorting
- electrophysiology
- mouse
- neuropixels
- scikit-learn
- spikeinterface
- modAL
license: mit
language:
- en
metrics:
- accuracy
pipeline_tag: tabular-classification
---



# 🧠 UnitRefine – Active Learning Classifier (Mice)

This repository supports an **active learning pipeline** for refining spike sorting classification using human-in-the-loop curation.
It includes both training data and a model trained on **mouse electrophysiological recordings**, annotated with labels such as  `sua` and `not-sua`.

---

## 📁 Files and Descriptions

| File              | Description                                                                 |
|-------------------|-----------------------------------------------------------------------------|
| `X_labeled.csv`   | Full feature matrix used for training the classifier  |
| `y_labeled.csv`   | Corresponding labels for `X_labeled.csv` (`sua as 1`, `not-sua as 0`)               |
| `X_new.csv`       | Features of newly queried units labeled by a human using SortingView        |
| `y_new.csv`       | Human-verified labels for the units in `X_new.csv`                          |
| `learner.skops`   | Trained `RandomForestClassifier` serialized using [`skops`](https://skops.readthedocs.io/) |

---

## 🔁 Active Learning Workflow

1. Query uncertain units using [modAL](https://modal-python.readthedocs.io/).
2. Present units to human curators via SortingView.
3. Incorporate corrected labels into the training set.
4. Update and re-train the model.
5. Push updates to the Hugging Face Hub.

---

## 🐭 Dataset Info

- **Species**: Mouse   
- **Features**: Quality metrics + Template metrics from Spikeinterface 
- **Label classes**: `not-sua`,  `sua`, `psua`

---

## 🚀 Model Use

To use the trained classifier:

```python
from skops.io import load

clf = load("learner.skops")
preds = clf.predict(X_test[feature_columns])