Continual Learning by Task Likelihood Analysis
Introduction
Task Conditional Neural Networks (TCNN) leverage the probabilistic neural networks to estimate the probability density of the training samples. Then produce the task likelihood during the test state to fire the task-specific neurons correspondin to the test sampels. TCNN can detect and learn the new tasks fully-automatically without informing the changes to the model.
The schematic diagram of TCNN learning process.
Dependencies
The code is tested on
- Python 3.6.8
- Tensorflow 1.13.1
- Keras 2.2.4
Reproduce
To run the experiment, run the '''shell split_mnist.py''' files via: 'python split_mnist.py'
The result should be similar to the figure below:
