Highlights on Applied AI Book

May 22, 2020

G ot to know about this book after watching one of the VLAB sessions at Stanford back in 2018. The book on Amazon. The gist of the book is AI is still in early stages, will be hard to introduce AI to organization, but can be done if we follow the right approach. Very cool projects out there that showcase Machine Learning.

Applied Artificial Intelligence: A Handbook For Business Leaders by Mariya Yao, Adelyn Zhou and Marlene Jia

What business leaders need to know

• We’re very far from having machines that can learn the most basic things about the world in the way humans and animals can,” said Yann LeCun, head of AI at Facebook. Page 12
• Descriptive statistics describes or visualizes the basic features of the data being studied. Page 14
• Inferential statistics is used to draw conclusions that apply to more than just the data being studied. Page 14
• Supervised learning occurs when the computer is given labeled training data, Page 15
• Supervised learning is commonly used for classification, where inputs are divided into discrete and unordered output categories, and for regression, where inputs are used to predict or estimate outputs that are numeric values. If you are trying to predict whether an image is of a cat or a dog, this is a classification problem with discrete classes. If you are trying to predict the numerical price of a stock or some other asset, this can be framed as a regression problem with continuous outputs. Unsupervised learning occurs when computers are given unstructured rather than labeled data, i.e. no input-output pairs, and asked to discover inherent structures and patterns that lie within the data. One common application of unsupervised learning is clustering, where input data is divided into different groups based on a measure of “similarity.” For example, you may want to cluster your LinkedIn or Facebook friends into social groups based on how connected they are to each other. Page 16
• Supervised learning is commonly used for classification, where inputs are divided into discrete and unordered output categories, and for regression, where inputs are used to predict or estimate outputs that are numeric values. Page 16
• While the research and applications are in its early days, many experts see probabilistic programming as an alternative approach in areas where deep learning performs poorly, Page 18
• Check out the MIT Probabilistic Computing Project7 for recommended readings and tutorials. Page 19
• We designed the Machine Intelligence Continuum (MIC) to present the different types of machine intelligence based on the complexity of their capabilities. Page 20
• Nearly every company has a few executive HiPPOs. While you can probably manage a handful of naysayers, your company is unlikely to be competitive in AI if you’re up against a HiPPO army or an extremely powerful C-Suite HiPPO. Page 40
• Ng proposes that organizations establish a new role, the Chief AI Officer (CAIO), to govern and champion the role of AI in enterprises. Page 45
• Enterprises early in the maturity cycle for big data and analytics may need to wait until a basic data and analytics infrastructure is in place before chasing AI. Page 47
• Team should be also surrounded by advisors from non-technical departments, such as HR or Finance, who can inform you of the areas in most urgent need of automation and the areas with the highest ROI. Page 49
• Pick a smaller, sure-win project to demonstrate possibilities. While returns may be limited, an early success will give you confidence when you request that the project scope be expanded. Page 50

How to develop an enterprise AI strategy

• Don’t Call It Artificial Intelligence When pitching your project, emphasize the value that new technology can deliver instead of the technical details of implementation. Page 50
• The head of a prominent Silicon Valley AI lab recently confided to us that American universities only graduate about 100 competent researchers and engineers in this field each year! Page 53
• ML engineers build machine learning solutions to solve business and customer problems. These specialized engineers deploy models, manage infrastructure, and run operations related to machine learning projects. Page 54
• For cutting-edge AI research positions, advanced mathematical intuition is a prerequisite in order to design and develop novel methodologies. Curiosity Training ML algorithms requires a constant sense of curiosity. The model builder needs to take in abstract information and make sense of it through continuous experimentation. This person will need to enjoy constantly learning new information and taking on new challenges. Page 56
• Look for pragmatic applicants who recognize that a “good enough” model that meets product deadlines is better than a model that sits in development awaiting “just a few more tweaks.” Page 57
• You can also host competitions on Kaggle or similar platforms. Provide a problem, a dataset, and a prize purse to attract competitors. Page 59
• Google and Facebook attracted university professors and AI research pioneers such as Geoffrey Hinton, Fei-Fei Li, and Yann LeCun with plum. Page 59
• In a recent survey of data scientists, 57 percent reported that data cleaning consumed 60 percent of their time and was the least enjoyable part of their job. Page 61
• AWS Rekognition or Apple’s Core ML. Page 69
• Goals like “Increase revenue by X dollars” or “Cut costs by Y percent” are almost always too high-level to be useful, since a confluence of intermediate factors drive your end result. Choosing a more specific metric helps you to qualify your AI strategy and also check that your decisions align and advance your business. 81
• Facebook’s true north metric for platform growth is the number of members who connect with ten friends in seven days. 81
• Slack focuses on teams that have exchanged at least 2,000 messages. 81
• The foundation of artificial intelligence is data. 88
• Precision tells us the model’s ability to correctly classify the instances that we care about in a dataset. 89
• Measuring the precision of a spam filter would tell you the number of messages that the filter correctly identifies as being spam (true positive) out of the total number of messages that it classified as spam, both actual spam messages (true positive) and legitimate messages that were mistakenly labeled (false positive). 89
• Underfitting occurs when your model is too simple to capture the complexities of your underlying data. 91
• Overfitting occurs when your model does not generalize well outside of your training data. 91
• Machine learning models are never “done,” in the sense that they will need continuous monitoring, iteration, and retraining to maintain required levels of performance over time. 95
• When IBM Watson collaborated with Toyota to create advertising copy for the new Mirai car model, the first versions of algorithmically generated texts were incoherent. After a couple months of tuning, the AI system learned to write thousands of useful ads. 95
• “The Hidden Technical Debt In Machine Learning Systems,” Google engineers warn that writing the code for a model constitutes only a small fraction of the engineering required for producing a machine learning system. 96
• Companies such as Google, Facebook, and AirBnB have created internal Machine Learning as a Service (MLaaS) platforms to enable their engineering teams to build, deploy, and operate machine learning. 98
• Uber developed the Michelangelo MLaaS platform to address these issues. Similar internal platforms at other companies include Google’s TFX and Facebook’s FBLearner Flow. 98
• Models are typically retrained every few weeks or months, or when there is a substantial change in external conditions that fundamentally changes the model trajectory. 100
• As a rule of thumb, half of your time should be spent on measurements and maintenance rather than on model creation. 100
• learning platforms like Facebook’s FBLearner, Uber’s Michelangelo, Google’s TFX, and Twitter’s Cortex to enable their engineers. 101
• Andrej Karpathy, the Director of AI at Tesla, offers a similar vision, describing a future in which “a large portion of programmers of tomorrow do not maintain complex software repositories, write intricate programs, or analyze their running times. They collect, clean manipulate, label, analyze and visualize data that feeds neural networks.” 114
• Examples of such assistants include Kite for Python and Codota for Java. 116
• The White House reported that other OECD countries spent, on average, about 0.6 percent of their GDP on active labor market policies in 2014; in contrast, the United States spent 0.1 percent of its GDP, which was less than half of what was spent thirty years ago. 125
• For up-to-date content and case studies on how AI is being used in specific industries, visit our website at appliedaibook.com. 128