Peeking inside the Black Box: Comprehensible Neural Networks, for Cancer Diagnosis
Team: 1005
School: Los Alamos High
Area of Science: Sparse Coding, Cancer Diagnosis
Proposal:
Rationale
When a pathologist makes a diagnosis of cancer from images of tissue, the doctor can explain the basis of the diagnosis. Many (deep) machine learning models have been created to provide fully automated detection of tumors. However, these models appear as black boxes, so are regulated FDA, for one reason: dense neural networks are difficult to explain and rarely can we determine what features they are basing a decision on. Worse, success on testing data isn’t a guarantee of future reliability. Adversarial training methods have proven that many seemingly reliable deep neural networks internally rely on irrelevant features in the training data, leading to potentially unsafe diagnostic results. Sparse neural networks have been proven to be more robust to adversarial attacks in the past. I hope to find that sparse coding neural networks are also more explainable than deep learning models. I will compare the performance of deep and sparse networks on tumor pathology slides.
Research Question
Will a sparse coding neural network classify better or worse than a deep neural network?
Will a sparse coding neural network be more explainable?
Are sparse coding neural networks robust to adversarial pathology examples?
Hypothesis
A sparse coding neural network will be more explainable, because fewer neurons participate in a decision and because they tend to be more robust to adversarial examples, than deep neural networks.
Expected Outcomes
Deep learning models will be tricked by adversarial examples, while the sparse coding model stays intact, and is explainable.
Methodology
Write models in Python, TensorFlow, Keras, and Numpy.
Write multiple neural networks; one which uses sparse coding, another which uses deep learning.
Data Analysis
I will compare the predictions of my models with the real-world annotations of modern pathologists.