Enterprise AI For Dummies. Zachary Jarvinen

your sources for bad data or hidden bias, such as race, gender, ideological differences, and the like. For example, a bank might remove gender and race from its loan-processing AI model, but U.S. ZIP codes can often serve as a proxy for race. Their inclusion could still lead to biased predictions that discriminate against historically underprivileged neighborhoods.

      6 Rigorously examine the criteria for identifying selection variables.

      7 Test machines for sources of bias and evidence of bias, and remedy any problems discovered.

Choosing a Model

      Although you might hear the term “artificial intelligence” bandied about as if it were a single thing, the reality is that it is an umbrella term for a vast discipline covering countless models or algorithms of varying complexity and rigor. Even within machine learning, dozens of methods can help you accomplish your goal, each used for a specific type of problem.

      Unsupervised learning

This method of ML recognizes patterns in a dataset and infers the structure or identifies correlations between data elements. You use unsupervised learning when you want to discover relationships, such as between account activity and fraud or an attack on the system. Table 3-4 lists AI project goals and the appropriate algorithms used for that task.

TABLE 3-4 Unsupervised Learning Algorithms

Goal	Algorithm
Organize data in clusters or trees, such as evaluating investments according to volatility or return	Hierarchical cluster analysis
Recommend a product or service based on the choices of similar customers	Recommendation engine
Optimize delivery routes by identifying proximate destinations	K-means clustering
Identify risk of heart disease based on heart sounds	Gaussian mixture

      Supervised learning

You use supervised learning when you want to classify new data based on known relationships in historical data, such as labeling incoming documents or screening job applications. Table 3-5 lists AI project goals and the appropriate algorithms used for that task.

TABLE 3-5 Supervised Learning Algorithms

Goal	Algorithm
Detect fraud in financial transactions	Random forest
Forecast for supply chain management	Regression
Forecast sales	Neural network
Underwrite loans	Decision tree

      Deep learning

This method of ML requires massive amounts of data, but typically with more accuracy and efficiency than other methods. You use deep learning when you want to solve complex problems such as image classification, natural-language processing, and speech recognition. Table 3-6 lists AI project goals and the appropriate algorithms used for that task.

TABLE 3-6 Deep Learning Algorithms

Goal	Algorithm
Train smarter chatbots, perform language translation	Recurrent neural network
Make medical diagnoses using computer vision	Convolutional neural network

      Reinforcement learning

      This method of ML learns a task through trial and error based on preferring actions that are rewarded and avoiding actions that are not. You use reinforcement learning when you need to find the optimum way of interacting with an environment, such as to automate stock trading or to teach a robot to perform a physical task.