[Video Transcript] Designing and Building AI Products with MIT xPRO
The different weeks of the course will cover examples of AI working as a pure play data technology, how AI can work with humans, or how AI can work with physical products. Many AI products only require digital data. AI works a bit as a black box and does not interact very much with any other activity in the firm. AI, in this case, is sort of an independent technology very much like a light with a switch. An example of this is a diagnostic tool based on images, for example when you detect cancer from images and all you do is tell people whether they are at risk or not based on a picture taken from a camera in a hallway. You may have to do many things differently if you detect positive cases just from the pictures. You may have to call a doctor or send the person for a more precise test.
The next three weeks will focus on this pure technology play. In the following two, we will review what is the heart of machine learning, and in the following one, we will present a complete digital health diagnostic tool and those will be the three weeks focused on technology. These are very critical technology fundamentals that are part of the development process of most of today's AI products.
Ready to Learn More? Apply to Enroll in the Designing and Building AI Products Course with MIT xPRO
Professor Duane Boning will present an overview of artificial intelligence techniques and will mostly revolve around machine learning. That's because most of the success of AI comes from algorithms that exhibit intelligent behavior by processing a lot of data and are being able to learn sort of "a desired behavior," a sort of intelligence you identified in stage one. There are many types of machine learning algorithms, and there are also many taxonomies of machine learning. The simplest form of machine learning to understand is supervised learning.
Here, the AI algorithms use as input data that has some label and try to generalize such labeling to new data that's unlabeled. For example, one can have a set of 10,000 images of a cat and 10,000 images without a cat and the machine learning algorithm has to learn when a new image has or does not have a cat. Another category of algorithms that you'll also be exposed are unsupervised learning. They try to identify clusters automatically and generate their own labels. One may have the history of all purchases in a retail chain and the machine learning may automatically find the most representative set of 10 customer segments such that their purchases are similar.
There are also algorithms that combine supervised and unsupervised learning, the so-called semi-supervised learning algorithms. Another type of machine learning beyond supervised and unsupervised and semi-supervised and a very old one is called reinforcement learning. An example of reinforcement learning is a robot that knows nothing about its environment and has to learn how to navigate through trial-and-error like sort of your Roomba vacuum cleaner which, by the way, was designed at the AI Lab in the 90s by Colin Angle one of Rodney Brook's students.
Related Topic: Browse all Online Courses at MIT xPRO
Another example of reinforcement learning is video games. An algorithm starts to play a game and has to learn how to excel at it. The only help provided to the algorithm is a reward based on how long it can play the game or based on how many points it achieves. There are thousands of reinforcement learning proposals out there. Perhaps the most surprising about humans is that they combine all types of learning. A baby, simply by listening to the environment, can learn many things, in fact a whole language. As a proof of concept, scientists at MIT Center for Brains, Minds, and Machines, or CBMM, have been able to demonstrate how given a set of vocabulary words and video sequences containing these words, artificial intelligence inspired by psychophysics is able to learn the corresponding grammar and vice versa almost like magic. This means that given a grammar in a video sequence, they can also extend the vocabulary.
If you combine the CBMM example with a baby conjecture, it seems plausible that we'll be able to employ algorithms that will learn new languages on their own simply by observing video sequences very much like humans do in a very sophisticated reinforcement learning strategy employed by humans. This is perhaps one of the biggest proofs of where AI is headed. You simply enter a video sequence and out comes a program that parses language. Once this is achieved, then the artificial intelligence can start making its own descriptions of what is going on in other video sequences. In fact, there are several tests around human learning that robots have not been able to achieve yet but maybe for not too long. For example, the coffee test invented by Apple's cofounder Steve Wozniak, that is to be able to enter a house, wander around like Roomba, learn where things are to make coffee and then learn how the machine works and finally produce a coffee.
We will review the use of AI in human-computer interaction. This can take on many forms such as vision, voice, brain scans. Adoption of new HCI forms can be very quick. HCI stands for human-computer interface. For example, Apple pioneered the voice revolution in 2011 with introduction of Siri in its iPhone 4S and in less than 10 years, all phones basically had voice recognition capabilities, and smart speakers like Google Home or Alexa are almost every home. The kind of IP that you will need will depend strongly on the HCI form you pick and on what expectations you have for the outcome. For example, if you need face recognition for a faster kiosk experience, you probably can use one of the pay-as-you-go services of the large cloud providers and you don't need to develop any own IP.
On the other hand, if you want to develop a new toothbrush based on face recognition, you may want to file your own patent which is actually what, in fact, Oral-B did. You can use the Oral-B app that looks at how you're brushing your teeth and alerts you if you have missed any of them. So the topic of human-computer interface, or HCI, will be covered by Professor Stefanie Mueller and Andy Lippman.
Stefanie Mueller will give a brief overview of the field and then talk about some really exciting research that's happening at MIT where basically you're interfacing with the 3D environment. Think about your video games where the environment is changing, think about the same thing happening. So it's just fascinating. You are in a sports stadium and then all of a sudden it changes from being a tennis court to a basketball ring, and there's many changes. All of a sudden your car changes colors to some fascinating, mind-blowing applications of AI.
And Andy Lippman will tell you a lot about the media. He will do a deep dive on video, video editing, and many things that can happen related to the media industries and that will also broaden your perspective on what's possible with AI.
After looking at HCI, Professor Tom Malone will present a different use of AI, one where humans and computers combine efforts to act in ways that seem intelligent. He will introduce the concept of superminds.
Finally, another type of IP choice and AI choice that we will look at last but not least is how AI can impact manufacturing and robotics. Bruce Lawler will introduce the latest thinking on how AI can be incorporated in manufacturing processes. Here, the IP may be related to the specific manufacturing process. For example, you may develop a new computer-based inspection system. IP may also be related to new robots. Here, the IP may be related more to the mechanical design of the parts than with a specific algorithm.[End of Video Transcript]
Also Related: AI and Machine Learning Courses