blazecolby/PyTorch-LIME

PyTorch - L.I.M.E ( Feature attribution for deep learning models )

Introduction

This repository walks through an example of LIME(Local Interpretable Model-Agnostic Explanations). Orginial LIME paper - https://arxiv.org/abs/1602.04938

LIME provides a means to explain any black-box classifier or regressor.

Models can be difficult to analyze on a global level but may be possible to analyze for a specific instance.

Desirable characterstics of an explainable model:

• Interpretable
• Local Fidelity
• Model-Agnostic
• Global Perspective

Resources

Utilizes:

• https://github.com/marcotcr/lime LIME model
• Pytorch
• ImageNet Pretrained model
• PIL