Mayurji / CAM-Dropping-Neurons

When significant neurons of trained neural networks is dropped gradually, the trained model loses it ability to classify the label correctly.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Units Vs Neuron

Brain

Effects Of Units When Dropped

Units Dropped

Units-In-Deep-Neural-Networks

When top neurons of trained neural network of a specific class is dropped gradually, the trained model loses it ability to classify the label correctly. Interestingly, the model starts focusing on other regions of an Image with probability getting distributed among other classes.

Intutitive Understanding

Consider you have strong memories of particular events in your life like going to college or the time you met with an accident and it is stored as neuron among billions of neurons in your brain and now you vaguely remember those events because your neurons have vanished
slowly as the years have passed.

Similarly now compare those with neurons of your model which has learnt to identify a scene but as i drop some of those neurons, it starts loses ability to
understand its role and vaguely predicts with available neurons. Interest Concepts, Don't you think?

Key Insights

  1. Few Powerful Units are enough for the Model to predict the class.
  2. Class Activation Map highlights the focus of units in an image, as the units are dropped.
  3. Probability reduces for the Ground Truth label when few powerful units are dropped.
  4. Probability gets shared across best alike other labels when units are dropped.
  5. Obvious result, Predicted Label changes as units are dropped.

Installation

pip install -r requirements.txt

How to run IPywidget

Watch the video

Note

Top Most Neuron refers to neurons with large weight value and Bottom Most Neuron refers to weights with least value.

Packages

  • ipywidgets==7.5.1
  • opencv_python==4.2.0.34
  • numpy==1.19.1
  • torchvision==0.7.0
  • matplotlib==3.3.1
  • torch==1.6.0
  • ipython==7.18.1
  • Pillow==7.2.0

References

About

When significant neurons of trained neural networks is dropped gradually, the trained model loses it ability to classify the label correctly.

License:MIT License


Languages

Language:Jupyter Notebook 100.0%