Comment on page
Why do we need Explainablity?
Motivation for Expalianbility
To provide a complete explanation for any decision made by a black-box AI system can be a tedious job. The tedious job has led to a new line of research – Explainable AI science of interpreting what an AI system did.
People who use a machine learning model should trust the model with its predictions and decisions. This is highly subjective and varies from individual to individual.
Transparency refers to the requirement that the end-user can understand how a decision/prediction is made by the AI system. Essentially, this means that one should understand what a model predicts and if there is any bias in that decision. Correctional Offender Management Proﬁling for Alternative Sanctions (COMPAS), a widely used criminal risk assessment tool, found that its predictions were unreliable and racially biased. So, after gaining transparency, this insight revealed a bias in the model which was a crucial breakthrough for the company.
Not all models are 100 percent accurate and hence defining quality is important. It only makes sense for people to use a highly accurate model. To identify, log and articulate sources of error and uncertainty throughout the algorithm and its data sources helps understand the expected and worst-case implications and further aid mitigation procedures.
For decisions made by the model - who is accountable? The person who uses it or the person who built it? There should be a way to identify and assign the responsibility for a decision made by an AI system.
How do we ensure that algorithmic decisions do not create discriminatory or unjust impacts when comparing across different demographics (e.g. race, sex, etc).
To understand why Explainable AI is important, let us take an example of diabetic retinopathy - a diabetes complication that affects the eyes. It is caused by damage to the blood vessels of the light-sensitive tissue at the back of the eye (retina). Let's assume we use a Deep Learning model with Convolutional Neural Networks for the classification of the normal eye from the diabetic eye. We can easily make a model that does a fairly good job with a validation accuracy of 90%. Then, a few questions may arise - what did the model see in the image to classify it? Did the model look into the same diagnostic parts of the images as done by the doctors? Or did it do something else? This is a very important context wherein a person can lose eyesight if he/she is misdiagnosed (if our model says the eye is fine but the eye was damaged, someone is going to be in big trouble!). In such cases, explainability is meant to engender trust from a model.