Artificial Intelligence in Practice
Page 5
Apple's vision for the future is powerful handheld devices that are capable of running their own machine learning on datasets gathered via their own array of sensors. This is clearly at odds with the vision of a future dominated by cloud computing and relatively low-powered terminals often championed by other tech companies.
This means running machine learning algorithms directly on their devices using powerful central processing unit or graphics processing unit chips built into their phones, watches or speakers.
One current example is the development of the Neural Engine inside the latest iPhone X models.2 This is a custom-designed chip specifically designed for carrying out the neural net calculations needed for deep learning. This makes it faster to process functions such as Face ID logins, features in the camera that help users take better pictures (or add silly effects), augmented reality and managing battery life.3
Running machine learning on devices can also be far faster than having to wait until data is returned from the cloud before insights can become actionable. However, it isn't without disadvantages. Only being trainable on data gathered from one device means those algorithms won't have the benefit of learning from the huge, crowdsourced datasets that cloud machine learning can plug into.
This ties in with Apple's focus on safeguarding users’ data. By ensuring that sensitive personal data doesn't have to leave the phone before it can be processed by machine learning, it hopes consumers will trust their data is safer with them.
Apple's proprietary AI ecosphere centers around its Core ML framework. Core ML allows developers to build machine learning algorithms into products, including deep learning, computer vision and natural language. It powers the smarts behind Siri, Apple's voice assistant, as well as the AI functionality in iPhone cameras and QuickType keyboards.4
Smarter Apps
A significant part of the credit for the iPhone's success has to go to the App Store. Smartphone users had been downloading apps to their phones for a while when it was launched in 2008, but Apple's streamlined store meant iPhone users could customize and add features to their phones more intuitively than was previously possible.
Cognisant of the way its app ecosystem keeps customers coming back to Apple year after year when their mobile contracts are up for renewal, it has pushed developers to integrate AI into their third-party apps. This tactic is aimed at continuing to provide compelling functionality that isn't available on other mobile platforms. To this end, Apple has provided developers with tools such as Create ML, which enable them to power apps with machine learning running on users’ devices.
A great example is an app called Homecourt, which was designed to assist with refereeing amateur basketball games. All a user needs to do is point the camera at a game in play, and the machine learning will tag the players in the game, logging when they pass and shoot, as well as recording their position on the court. All this is done through computer vision technology running on the device itself.5
Another app, known as Polyword, allows users to get the name of whatever object they are pointing their camera at in any one of 30 languages, using computer vision and machine learning.6
Other features allow it to take a critical look at the photographs you are shooting and suggest improvements in real time, and manage notifications to make it more likely important information is brought to your attention at the right time.
Natural Language Processing
When Apple launched Siri it became the first widely used AI-enabled and natural language processing (NLP)-powered assistant. Although it has been criticized for a lack of innovation compared to that seen in competitor AIs,7 recent updates brought in real-time, machine learning-driven translation between 40 pairs of languages.
Siri's NLP functions do send information into the cloud. However, user privacy is preserved by the fact that all identifying information is stripped from the voice command data before it leaves the user's device in an encrypted form.
Recent NLP research at Apple has focused on giving Siri users more accurate results when they search for information on businesses or points of interest in their local area.8 Researchers introduced location signals into the training data, giving Siri access to localized datasets, including place names and small businesses. In theory it will use the location data while it interprets spoken language to add to its understanding of what the user might mean. Alexa will have a better chance at guessing, for example, if someone who utters the words “I'm going to Kilkenny” intends to visit the town in Ireland or murder a man called Kenny.
Key Challenges, Learning Points And Takeaways
AI is very much at the heart of Apple's strategy, which is to build it into the fabric of its devices and supporting services.
Apple is prioritizing user privacy over an ability to pump all data into the cloud to train algorithms on bigger data sets.
It is also promoting the use of its proprietary machine learning platform Create ML to make apps that will only work on its devices, creating exclusivity within its own app ecosphere.
Notes
1https://www.theguardian.com/technology/2018/aug/02/apple-becomes-worlds-first-trillion-dollar-company
2Wired, Apple's Neural Engine Infuses the IPhone with AI Smarts: https://www.wired.com/story/apples-neural-engine-infuses-the-iphone-with-ai-smarts/
3CNBC, https://www.cnbc.com/2018/09/12/apple-upgrades-neural-engine-in-iphone-xsa12-bionic-chip.html
4Apple, Get Ready for Core ML 2: https://developer.apple.com/machine-learning/
5Wired, Apple's Plan to Bring Artificial Intelligence to Your Home: https:// www.wired.com/story/apples-plans-to-bring-artificial-intelligence-to- your-phone/
6Github, Polyword: https://github.com/Binb1/Polyword
7Wall Street Journal, “I'm Not Sure I Understand” – How Apple's Siri Lost Her Mojo: https://www.wsj.com/articles/apples-siri-once-an-original- now-struggles-to-be-heard-above-the-crowd-1496849095
8Apple, Finding Local Destinations with Siri's Regionally Specific Language Models for Speech Recognition: https://machinelearning.apple. com/2018/08/09/regionally-specific-language-models.html
5
Baidu: Machine Learning For Search Engines And Autonomous Cars
Baidu is a Chinese technology company that focuses on internet-related services and products. The company operates the most popular search engine in China, giving it access to the vast datasets generated by billions of search queries. On top of that, Baidu is also an app developer, runs an advertisement platform and is recognized and supported by the Chinese government for its work around developing self-driving vehicles. Its Project Apollo is one of the most mature autonomous driving programs in the world. In 2018, Baidu became the first Chinese artificial intelligence (AI) company to join the Partnership on AI established by Facebook, Amazon, Google, Microsoft and IBM to encourage ethical development of AI.1
How Does Baidu Use Artificial Intelligence?
As well as a search function, Baidu also offers image searching, maps, videos, news and translation services to its users. AI has been deployed across all of these functions to more accurately return results that users will find useful.
This is helped by the fact there are more than 800 million internet users in China – more than twice the total population of the United States,2 which means that AI algorithms have a far bigger pool of data to work with.
Baidu collectively labels all of its AI operations as Baidu Brain. The platform, currently on version 3, offers access to 110 AI technologies, including natural language processing, image recognition, facial recognition and automated labelling of video data. It also includes the EasyDL tool, which allows development of deep learning systems without the need for any coding.
At a conference hosted by Baidu in Beijing in July 2018, a doctor with no programming skills used the platform and was able to develop a deep learning tool capable of identifying 40 types of parasitic worm, which is now being clinically tested.3
Self-Driving Cars
Baidu was chosen above its Chinese rivals as the nation's “champion” for the development of fully autonomous vehicles. The company is aiming to have self-driving cars on the roads of Beijing by 2019,4 and begin mass production by 2021.5
To get there, it launched Project Apollo, which involves partnerships with several high-profile car makers, including Ford and Hyundai.6
AI is crucial to autonomous driving. The cars are equipped with sensors connected to machine learning algorithms in the cloud and running locally in the vehicle to enable them to “see” conditions and hazards on the road.
Baidu's cars also use high-resolution 3D mapping data, collected from satellite imagery as well as by camera-equipped cars that have built detailed image databases covering China's road system.7
Ford's cars will take part in the first tests scheduled for early next year, when they will be fitted with Baidu's Virtual Driver System. This system is capable of giving the cars “level 4” autonomous driving ability, based on the Society of Auto Engineers (SAE) level system.
The SAE has defined five levels of autonomy, ranging from level 0 (no automation) to level 5, which is “full automation” – meaning the car can drive itself anywhere a human would be capable of driving it. Level 4 – which Baidu intends to road test next year – requires a car to be able to carry out all of the driver functions, with no need for the human driver to pay attention.8
As well as cars, the open source Apollo Virtual Driver System can be fitted to trucks, giving them the ability to operate autonomously within geofenced areas of open highway.
Mobile Artificial Intelligence
Baidu has partnered with Huawei to build an open AI platform for mobile development. The aim is to provide mobile users with “AI that knows you better”, making it more convenient than ever to use the functions and services we're used to accessing through our phones.9
It will let developers write code that will run machine learning tasks on the neural processing unit circuitry built into Huawei's phones. They will be able to take advantage of the voice and image recognition abilities of machine learning, as well as its suitability for building augmented reality apps. The move puts it into competition with Apple and Samsung, which are both developing their own mobile AI frameworks internally.
Real-Time Translation
Baidu has also developed a handheld device capable of generating deep learning translations between English, Mandarin Chinese and Japanese.10 It is currently aimed at the tourist market, assisting users to navigate their way around foreign cities and carry out tasks such as ordering food in restaurants and using public transport. It uses deep learning natural language processing algorithms, and translation is carried out in the cloud.11
Key Challenges, Learning Points And Takeaways
The huge population of its native country – around half of whom are online – has helped Baidu collect a vast dataset of consumer profiles and behaviors. This is used to streamline services, as well as sell to advertisers to allow them to more accurately target their campaigns.
Baidu offers AI services to businesses to enable them to develop and release their own AI-powered applications under its Baidu Brain framework.
Although Baidu was slow to catch on to mobile, it is making up for lost ground now through a strategic partnership with China's largest smartphone manufacturer, Huawei, to put AI inside phones.
Baidu has China's, and possibly the world's, most advanced autonomous vehicle program, with cars powered by its Apollo technology expected to bring level 4 autonomy to the roads soon.
Notes
1CNN, Silicon Valley is working with China to ease fears about AI: https://amp.cnn.com/cnn/2018/10/17/tech/baidu-artificial-intelligence -china/index.html
2Forbes, China Now Boasts More Than 800 Million Internet Users And 98% Of Them Are Mobile: https://www.forbes.com/sites/ niallmccarthy/2018/08/23/china-now-boasts-more-than-800-million- internet-users-and-98-of-them-are-mobile-infographic/#21c9e8807092
3Tech Republic, Baidu no-code EasyDL tool could democratize AI for small businesses, bridge talent gap: https://www.techrepublic. com/article/baidu-no-code-easydl-tool-could-democratize-ai-for-small -businesses-bridge-talent-gap/#ftag=RSS56d97e7
4Reuters, https://www.reuters.com/article/autos-selfdriving-baidu/chinas -baidu-gets-green-light-for-self-driving-vehicle-tests-in-beijing- idUSL3N1R51A5
5Tech Crunch, Baidu plans to mass produce Level 4 self-driving cars with BAIC by 2021: https://techcrunch.com/2017/10/13/baidu-plans-to-mass-produce-level-4-self-driving-cars-with-baic-by-2021/
6Ford, Ford and Baidu Announce Joint Autonomous Vehicle Testing: https://media.ford.com/content/fordmedia/fna/us/en/news/2018/10/ 31/ford-and-baidu-announce-joint-autonomous-vehicle-testing.html
7Bloomberg, Wanted in China: Detailed Maps for 30 Million Self-Driving Cars: https://www.bloomberg.com/news/articles/2018-08-22/wanted-in-china-detailed-maps-for-30-million-self-driving-cars
8SAE, Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles: https://www.sae.org/ standards/content/j3016_201806/
9Huawei, Huawei and Baidu Sign Strategic Agreement to Lead the New Era of Mobile AI: https://www.huawei.com/en/press-events/news/ 2017/12/huawei-baidu-strategic-agreement-mobileai
10Digital Trends: https://www.digitaltrends.com/cool-tech/baidu-machine -translator/
11MIT Technology Review, Baidu Shows Off Its Instant Pocket Translator: https://www.technologyreview.com/s/610623/baidu-shows-off-its-instant-pocket-translator/
6
Facebook: Using Artificial Intelligence To Improve Social Media Services
Facebook is a US-based, multinational social media and social networking company. It has been a part of the fabric of modern-day life for over a decade now. Around 2.2 billion1 people use the Facebook social media platform to keep up to date with friends and family, arrange their social lives, find local businesses and, of course, share pictures of their pets with the world.
Every time any one of us uses Facebook we generate data – what we are doing, where we are, who we are with. Before social media, we didn't have anywhere to upload the 136,000 images per minute we currently add to Facebook, let alone the 510,000 comments and 293,000 status updates.2
All that data is great training fodder for artificial intelligence (AI), too – and the company has launched a number of tools and projects that put machine learning in the service of its users.
How Does Facebook Use Artificial Intelligence?
Facebook uses its AI engine, FBLearner Flow, to personalize users’ news feeds and homepages, putting information (and advertising) in front of them, which it believes they will find useful or of interest.3
It uses machine learning to analyze and segment the platform's billions of users, generally according to information the users provide themselves – where they live, work, who they are friends with, where they travel to, what they search for online and what their signals (such as “likes” and “shares”) suggest about them.
Monitoring Content
As well as filling news feeds with interesting updates and stories, the machine learning algorithms also work to filter out content such as violence or nudity, which it doesn't allow users to post on the service.
One focus in this area has been a crackdown on distributors of “fake news” – either politically motivated or when done by fraudsters hoping to make money. Machine learning algorithms are used in conjunction with manual and automated fact-checking services.4 When stories are flagged, either by machines or humans, as fake, their spread across Facebook's network can be tracked, and steps taken to prevent users from being harmed. This could include deleting the material or flagging it as likely to be false.
Facial Recognition
One area of AI research where Facebook is head and shoulders above the competition is facial recognition technology, which is hardly surprising, considering how many pictures of people's faces it has on its servers.
The technology is
called Deep Face, and it's what springs into action when you upload a photograph and Facebook starts to suggest who it thinks is in the picture. It uses neural networks to parse 68 data points from every face it analyzes, measuring facial features, colouring and proportion.
It was fed over 4 million facial images to train it how to recognize individual facial elements, and understand how facial characteristics give each human a unique look. When another facial image it analyzes matches or closely fits a unique pattern it has already recorded, then it knows there is a higher probability that it has two pictures of the same person.
As well as conveniently tagging people in your photographs, Facebook has also used the technology to help users to keep track of where photos of themselves are cropping up on the site, and also to create audio descriptions of photographs to help the visually impaired.5
Facebook says that its facial recognition algorithms have a success rate of 97.35% when used with publicly available test datasets – very close to human-level accuracy.6
Understanding Text
AI is also used by Facebook to wring insights from the half a million text comments posted to the site every minute. Its aim here is to use contextual analysis to get a deeper understanding of what we're trying to say, and offer information or services that we might find useful without us having to ask for them. An example that Facebook gives is that machine learning algorithms “listening in” on a conversation between friends about a journey they have to make might automatically throw up links to ride-hailing services operating in the locality.7