https://github.com/jrieke/lstm-biology
Neural networks with LSTM to classify and predict biological cell movement
https://github.com/jrieke/lstm-biology
Last synced: 2 months ago
JSON representation
Neural networks with LSTM to classify and predict biological cell movement
- Host: GitHub
- URL: https://github.com/jrieke/lstm-biology
- Owner: jrieke
- License: mit
- Created: 2016-02-20T00:19:59.000Z (over 9 years ago)
- Default Branch: master
- Last Pushed: 2016-09-09T14:42:16.000Z (about 9 years ago)
- Last Synced: 2024-12-09T01:39:51.905Z (10 months ago)
- Language: Jupyter Notebook
- Homepage:
- Size: 10.2 MB
- Stars: 3
- Watchers: 3
- Forks: 4
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# lstm-biology
Code and report for **"Applying LSTM neural networks to biological cell movement"** (project at the [Biophysics Group, University of Erlangen-Nuremberg](http://lpmt.biomed.uni-erlangen.de/group)).
**Abstract:** Neural networks with Long Short-Term Memory (LSTM) were used on scientific time series
data. Each time series contains the positions of a biological cell while it moves through one
of three environments (a collagen network, a plastic surface, or a plastic surface coated with
fibronectin). The networks were used for two tasks: 1) Classifying the movement trajectories
based on the cell environment. Several networks of increasing complexity were trained on
parts of the trajectories, using softmax classification. The best networks achieved an
accuracy of ~95 % (on test data) and generalized well to longer trajectories. 2) Generating
new movement trajectories by predicting one step of a time series after another. For this
purpose, LSTM was combined with the idea of a mixture density network (MDN): It does not
predict the values of the next time step directly, but outputs the parameters of a mixture
distribution, from which they can be sampled. The generated trajectories replicated the
shape as well as the rough statistics of the original dataset.
**Requirements:** keras (v0.3.2), Theano (v0.8.0.dev0), numpy, matplotlib, jupyter (optional for computing statistics of generated trajectories: bayesloop, seaborn, scipy)