What Is AI Good For?

With all the hype surrounding artificial intelligence, it is important not to get too immersed in the technical and science fiction aspects. Steve Jobs said “You‘ve got to start with the customer experience and work backwards to the technology.” and this is true of artificial intelligence as well. Although a full discussion is totally out of the scope of a short blog post, I would like to provide a few perspectives.

Contextual interfaces

In my opinion, the right click in Windows 95 was a huge innovation. Prior to that, one had to look inside a huge array of options under a menu bar, or scan through a panel of small and often obfuscated icons. The right-click contextual menu showed a short list of tasks that were all relevant to the object that was currently selected and relieved you of this wasteful routine.

Although this may not strictly be classified as AI, the way that that it lessened the burden on the human brain was significant.

Similarly, one application of AI that would most certainly be very popular with users would be UI improvements that significantly reduced the need to scan through a list of options to find the relevant actions. In iOS 10, Apple has introduced AI that learns which emails should be sorted to which folders and intelligently provides a shortcut so that sorting emails is much quicker and easier.

Automated processing

Email spam filters learn what emails have a high probability of being span. From a user perspective, this is by all means artificial intelligence.

Although spam filters occasionally make mistakes, they help save our time and cognitive load by pre-filtering out stuff that is completely irrelevant to our work. Good spam filters also protect us from phishing attacks which can compromise whole corporate networks, and so it is no surprise that these are in high demand.

This is a very important market for AI.

Data Detectors

Apple had a patent for a very powerful technology commonly known as Data Detectors. This technology can detect addresses, event dates etc. inside text, and dramatically improves the user experience on smartphones where it is inconvenient to copy and paste.

The analysis of text, prediction of what a user might want to do with it, and providing a convenient and intuitive UI that enables the user to quickly get it done, can be a great timesaver.

Voice Recognition

It is well know that machine learning techniques have greatly improved voice recognition. Voice recognition has historically been valued by people who have difficulty typing. With mobile devices, voice recognition is convenient when you cannot use your hands.

Voice Interface

Graphical user interfaces are great for a stepwise approach for getting things done. However, since they operate by providing a list of options on a 2D screen, there is a limit to the breadth of commands that can be issued at any one time.

Command line interfaces and voice interfaces can get around this issue because they do not have to present a list of options. They are limited only by the ability of the user to memorise the available commands, and to issue them without referring to a menu. Hence voice interfaces are a convenient way to issue tasks quickly.

Summing up

My intention with this post was to show that there is much more to AI than a voice UI, and that from a practical perspective, the other applications have already proven to be very significant in terms of user benefit. Although voice UIs and predictive assistants like Google Now are interesting and futuristic, there is no reason why these applications will be the most useful and revolutionary.

Current advances in machine learning (Deep Learning) will build upon what we already have, and for smartphones with big screens, what we already have is a good graphical user interface.

AI, Voice UIs and predictive assistants should be evaluated based on their merits. How will they save us time and for what tasks? How will they help us when we cannot or it would be inconvenient to view our phone screens? How can they reduce our cognitive load?

Apple is pretty good at understanding what the user experience should be, and arguably, this will be just as important or maybe even more so than the underlying algorithms.

Thoughts on WWDC 2016

Here I want to jot down some of my key thoughts after viewing Apple’s WWDC 2016 keynote.

Core Apps as platforms

We saw a lot of the core apps being opened up to developers. We saw this for Siri, Maps, Messages and even the regular Phone app. Developers can now write code that directly extends the functionality of these core apps. This makes each app its own platform.

  1. This provides a path through which Apple Maps may become much better than Google Maps for many parts of the world. Third parties can innovate on how to provide better shop recommendations/information, transit information, rather then replicating core functionality.
  2. The same can be said of VoIP apps. I have never had a VoIP app that had nearly as nice a UI as the iOS default Phone app. Now VoIP apps can simply focus on providing good connection and voice quality.
  3. Ditto for Siri and Messages.
  4. This approach is only possible in some cases because Apple’s business model does not rely on advertising. For example, Google Maps could have trouble integrating information from Yelp, because this would conflict with their business model of profiting from the recommendations.

Differential Privacy

This is still a bold experiment. It has not yet been proved that this will allow sufficiently advanced artificial intelligence. In the following months, this will be put to the test. Differential privacy may prove to be just as useful as the lax privacy that companies like Google employ.

More importantly in my view though, is that differential privacy will allow Apple to get the most valuable data.

Privacy of health data is considered to be very important, especially genomic data. In genomic experiments using human-derived samples, great care is often demanded to defend the privacy of the donor. Google’s approach would probably be considered too relaxed to entrust such data, whereas Apple’s differential privacy may be sufficient. As a result, people might be very hesitant to give Google their DNA sequence information but not so for Apple (it might even be an FDA recommendation).

If this becomes the case, then Apple will have a huge advantage, not because it has better AI algorithms or more data, but because it has the most valuable data.

The same may occur with many other types of data. If this becomes that case, then Apple may gain preferential access to the more valuable and important data (that is not readily available by spying on your interactions with your phone). This will benefit Apple in the kind of conclusions that its AI will be able to make.

Google’s Justification

At Google I/O 2016, CEO Sundar Pichai showed a future filled with their artificial intelligence (AI).   It is all very interesting, but I do have some questions.

How much data does Google’s AI need?

Google’s AI is backed by enormous amounts of data about us. Data that is collected from photos that are publicly posted onto the Internet, and from photos that we upload onto the Google Cloud services from our mobile phones. Data from our messages on Gmail or events on Google Calendar. Data from the GPSs on our Android phones which tell Google where we are every hour of the day. Data from our browsers which tell Google (often without us knowing it), which website we have been visiting. No other company has access to similar amounts of private information.

However, what has not been answered is how much data Google’s AI actually needs.

Can effective AI be created without too much data?

A recent article by Steve Kovach on Apple’s next generation AI system is very interesting.

Siri brings in 1 billion queries per week from users to help it get better. But VocalIQ was able to learn with just a few thousand queries and still beat Siri.

This suggests that it is possible to construct a advanced AI system with magnitudes smaller data sets; data sets that do not have to be aggregates of private user information, but can simply be collated from a relatively small number of people who were paid for the work.

Of course we need to see the results to be sure. At the same time, I find it interesting that IBM Watson was able to win Jeopardy without tapping into huge data sets like those that Google uses.

Does an intelligent assistant mean you have to give up your privacy?

Apple tries hard not to see your private data. Apple believes that your private data belongs to you only, and that you should be the only one who holds the keys. Many people have questioned this approach, based on the assumption that widespread access to private information from millions of people on the server level is the only way to create a sufficiently good AI system.

Apple’s approach does not preclude the storage and analysis of personal data, as long as it happens in a way that Apple itself cannot see. One way to do this is to handle analysis on the smartphone. This is what the NSDataDetector class in the Mac/iOS API does. It’s actually pretty neat, and Apple has a patent on it. Similar but more advanced approaches could easily be implemented in iOS, given the performance of today’s CPUs.

The question is, is this approach sufficient? Will analysing your private data on your device always be much less powerful than analysing it on the server? Furthermore, will there be a significant benefit in collating the private data from strangers to analyse your own? If so, then Google’s approach (which sacrifices your privacy) will remain significantly superior. If not, then Apple’s approach will suffice. That is, you will not necessarily have to give up your privacy to benefit from intelligent assistants.

Does Google need the data for other purposes?

Let us assume that there existed a technology that allowed you to create an effective intelligent assistant, but that did not require that you give up your personal data. Would Google still collect your personal data?

The answer to this question is quite obviously YES. Google ultimately needs your private information for ad targeting purposes.

Could Google be using the big data/AI argument to justify the collection of huge amounts of private data for ad targeting purposes? I think, very possibly YES.