Libraries for Explainable Artificial Intelligence (EXAI)
A. Helen Victoria, Ravi Shekhar Tiwari, Ayaan Khadir Ghulam · 2024
Artificial Intelligence (AI) was first coined by Sir John McCarthy in the 1960s, and many of the algorithms that are used today were invented by researchers several decades ago. But due to restrictions on storage and computing power, they were not able to be implemented in real-world applications. Around 2010, the implementation pace increased drastically because of the breakthroughs in computing as well as storage. Since then, AI has been implemented in almost every field. Healthcare is one of the few fields where AI is taking its baby steps. AI is dependent on the dataset though healthcare has existed for several centuries. But we have a dataset shortage when it comes to train the AI model, especially deep learning (DL) models. Somehow, researchers are able to solve the dataset shortage problem by augmenting the original dataset or by generating synthetic dataset – considering several parameters from the original dataset such as mean, dispersion, mode range, and a variety of parameters to train/test and deploy the model. Machine learning (ML) and DL are subsets of AI, which is dependent on quality data for training. In recent years, researchers have succeeded in training a state-of-the-art model that can predict fatigue, wear-tear, life cycle, and other numerous properties of materials. But with recent breakthroughs, one problem of bias is persistent, and it has made a catastrophic, i.e., biased model. Quality of model output is clearly dependent on the quality of the dataset, and if the dataset is biased, then the model will become biased about certain features that can lead to incorrect results. We have several epitomes of biased models; for instance, in the USA, black skin colored people were categorized as criminals, women were considered as homemakers, and men were considered as breadwinners by the NLP word-embedding model and several other models which became biased on certain features after training. So, there is a need to explain the model behavior with respect to the dataset by considering each specific feature as well as by considering the full dataset. So Explainable Artificial Intelligence (XAI) comes into existence which provides the reason why the model predicted this output. Healthcare is a very sensitive area where small error can cost numerous human lives and early detection of diseases can save numerous lives. ML is holding its ground but the reason why the respective ML model comes to certain decisions is still unknown; hence, it is known as black-box models. ML products and algorithms have been utilized in humongous amounts by our environment, and a single error or wrong prediction can cost numerous human lives. So, there is a need to explain why the model predicted this specific value of class so that the human supervisor can be satisfied as well as the chance of error can be minimized. Recently, XAI has explained the reason behind the model prediction satisfactorily by employing algorithms such as Feature Importance, Permutation of Features, SHAP Value, and Activation at each layer which are used by various libraries to visually represent the reason behind the prediction. Nowadays, there are a variety of frameworks and libraries to justify the prediction of the model such as What-If Tool by TensorFlow, ELI 5, Shapley, Lime, IBM 360 Explainability, Deep LIFT, Skater, etc.