Lecture Notes From Artificial Intelligence Is The New Electricity By Andrew Ng

The guru of AI, Andrew Ng gave a lecture recently (published Feb 2017) to Stanford Graduate School of Business students.

Here are the important lecture notes and takeaways from this video.

Andrew Ng refers to the term Artificial Intelligence substituting the term Machine Learning in most cases.

AI is positioned today to have equally large transformation across industries as the invention of electricity had about 100 years ago.

Among industries, IT (search, advertising, e-commerce) has significantly transformed by AI, so is Fintech, supply chain and logistics is next, healthcare is seeing the very beginnings of this transformation journey, and transport (self-driving cars) will take some time.

AI is driving tremendous economic value across industries and its applicability is universal.

AI is maybe driving market cap in hundreds of billions of dollars already.

Most of AI applications of today that are creating this economic value are based on Supervised Learning class of algorithms.

Choosing AI problems to solve

Andrew lists supervised learning examples including Spam Detection, Image Recognition, Speech Recognition, Text to Speech, Translation, Radiology, and targetted online advertisements serving and click-through optimization (possibly the most lucrative applications of AI today with mature models available).

Despite the hype in AI today, it is relatively limited (Supervised Learning only) when compared with human cognition.

How to choose what problems to solve with AI?

Andrew’s Rule of Thumb: Anything a typical human can do with up to 1 second of thought, we can probably now or soon automate with AI.

The rule of thumb elaboration involves following questions.

• Is the problem even feasible? If humans cannot solve the problem then can it actually be done by machine? Examples may include stock market predictions.
• Is there lots of data available for the AI problem. To use for labeled training the AI models.
• Are there human (expert) insights available on the solutions for this problem? Can experts comment on a solution delivered by AI so that these insights can serve as a feedback loop for recalibrating the AI models?

Another reason for the rule of thumb for identifying AI problems to solve in relation to human performance for the same tasks.

The performance of AI improves rapidly until it reaches human performance, then the performance increase is not as rapid.

Andrew answers the question, why AI is achieving a lot of success over last 4–5 years when compared with earlier phases of AI. Read more about the AI Winters of 1970s, and setbacks in 1980s and 1990s.

He attributes the recent success of AI due to a combination of (a) availability of Big Data, (b) supercomputing power to process this Big Data over very large neural networks, (c) modern algorithms.

Supercomputers or high-performance computing (HPC) research is required for running very large neural networks.

The bleeding edge of AI research involves AI teams structured as a combination of ML experts and HPC experts working together.

Fundamentals of neural networks

Here is an example of a typical Machine Learning problem and a simple Neural Network algorithm to solve this problem.

The problem is to predict house price for a given house with a number of input features like area, number of rooms, zipcode, and wealth of the neighbourhood. The training set comprises of house prices for a number of houses across a range of these feature values. The machine learning solution involves fitting a linear function (a line) through the given training dataset points. This linear function can then be used by the Neural Network to predict the price.

According to Andrew the magic of Neural Networks is to figure out the derived features like school quality, walkability, and family size on its own without explicitly being programmed to do so.

Defensible AI business model

AI research is faily open and many leading AI researchers share their findings. So algorithms may not contribute towards a defensible business model.

There are two scarce or expensive resources for AI today. One is data. Large amounts of accurately labled, clean, structured data for training AI algorithms. Second is AI talent.

Talent for AI is actually the most scarce resource.

Here is a sense of scale data needs to be for leading AI research.

Speech recognition requires around 10 years of audio to train.

Computer vision (face recognition) requires 1 million images for most popular benchmarks of face recognition. Very largest academic papers publish based on 15 million images. Baidu face recognition system trains on 200 million images.

The virtuous circle of product lifecycle is where a product (a speech to text app) attracts several users, the users generate data (speech samples), this data is processed by ML algorithms to train better models, the product gets better and attracts more users, generating more data, and so on.

This in turn creates significant entry barriers for competition as your product has richer datasets and better ML algorithms.

The opposite of this virtuous cycle is funding going into research responding to “fear of evil AI” as this takes away talent from core AI research.

Job displacement because of AI is a real “fear” to be addressed by the industry.

Product management workflow

A good product management workflow for AI products involves product manager worrying about what user cares for.

The engineer or researcher brings her perspective on what is feasible with current technology. The intersection of these two perspectives is the space where we should build products.

Instead of PM providing wireframes for the AI product user interface, they can be tasked to provide the training dataset based on desired use cases for the product.

We are still in the early stages of organizing product development workflows when it concerns AI products.

Short term opportunities of AI

In the near term future Speech Recognition will take off rapidly. Speech recognition has passed the threshold of being 3X faster than typing on a smartphone in English and Mandarin languages.

Computer Vision is coming little bit later. Face recognition is taking off much more rapidly. Face recognition is being used for secured access to facilities and biometric identity verification for processing financial transactions.

Andrew Ng thinks that after seeing two winters, AI has entered the phase of eternal spring, just like we are in the eternal spring for technologies like Silicon or Transistors. There is clear roadmap for transforming several industries with AI applications.

Hope you find these notes a useful complement to the video lecture. If you want me to cover more such lectures, please like (heart), share, and comment below. Much appreciate!