Enterprise AI For Dummies. Zachary Jarvinen
a product to the customer. Through data mining and text mining, not only can you personalize the product to a specific customer, you can also discern trends across segments and use the information to inform product development.
Identifying the Enabling Technology
Just as constant as the challenges posed by competition throughout the ages is the role of innovation in addressing competitive pressure. Four millennia ago camels were domesticated, and a few centuries later ships were launched to enable long-distance trade.
In this new millennium, the continued pressure of competition has fueled advances in technology, particularly in the domain of artificial intelligence.
Like the camel and the ship, AI enables those in business to go farther and faster, to respond to global pressure to reduce cost, increase efficiency, and accelerate the development and delivery of products.
However, several enabling technologies had to reach maturity to create a foundation that would allow AI to realize the potential envisioned by the scientists at the 1956 Dartmouth Summer Research Project on Artificial Intelligence.
Processing
In a 1965 paper, Gordon Moore, the co-founder of Fairchild Semiconductor and CEO of Intel, observed that the number of transistors in a dense integrated circuit doubled about every year. In 1975, Moore revised his estimate going forward to doubling every two years.
The first single-chip central processing unit (CPU) was developed at Intel in 1970. In the intervening half-century, computing power has increased roughly according to Moore’s law. For example, in 1951, Christopher Strachey taught the Ferranti Mark 1 computer to play chess. Forty-six years later, the IBM Deep Blue computer beat world chess champion Garry Kasparov. Deep Blue was 10 million times faster than the Mark 1.
While the curve is starting to level out, 50 years of advances in processing power has established computing platforms capable of the massive, parallel-processing power required to develop natural-language processing (NLP), self-driving cars, advanced robotics, and other AI disciplines.
Algorithms
In the 1990s and beyond, work in AI expanded to include concepts from probability and decision theory and applied them to a broad range of disciplines.
Bayesian networks: A probabilistic graphical model that represents a set of variables and their conditional dependencies via a directed acyclic graph
Hidden Markov models: Statistical models used to capture hidden information from observable sequential symbols
Information theory: A mathematical study of the coding, storage, and communication of information in the form of sequences of symbols, impulses, and so on
Stochastic modeling: Estimates probability distributions of potential outcomes by allowing for random variation in one or more inputs over time
Classical optimization: Analytical methods that use differential calculus to identify an optimum solution
Neural networks: Systems that learn to perform tasks by considering examples without being programmed with task-specific rules
Evolutionary algorithms: Population-based optimization algorithms inspired by biological evolution, such as reproduction, mutation, recombination, and selection
Machine learning: Algorithms that analyze data to create models that make predictions, take decisions or identify context with significant accuracy, and improve as more targeted data is available
As the sophistication of the algorithms directed to the challenges of AI increased, so did the power of the solutions.
Data
The early days of life on Earth were dominated by single-celled organisms that sometimes organized into colonies. Then, back about 541 million years ago during the Cambrian era, most of the major animal phyla suddenly appeared in the fossil record. This is known as the Cambrian explosion.
It seems that the twenty-first century is experiencing its own Cambrian explosion of data. In the beginning, there was data. Pre-Cambrian data. It was pretty simple, mostly structured, and relevant to specific commercial applications such as accounting or inventory or payroll and the like. Data processing turned that data into information to answer questions, such as “What does that mean for me?”
Now, thanks to the Internet and other data-generating technologies, big data has arrived. Unfortunately, traditional data processing lacks the sophistication and power to answer all the questions that are hidden in the data. AI employs big-data analytics to turn big data into actionable information.
Volume
Variety
Velocity
Volume
Much more data is available now. In fact, the sheer volume of data being generated every minute is staggering:
On YouTube, 300 hours of video are uploaded.
On Facebook, 510,000 comments are posted, 293,000 statuses are updated, and 136,000 photos are uploaded.
On Twitter, 360,000 tweets are posted.
On Yelp, 26,380 reviews are posted.
On Instagram, 700,000 photos and videos are uploaded.
And all this is on just a few social media sites.
AI needs data, lots of data, to generate actionable recommendations. To develop text-to-speech capabilities, Microsoft burned through five years of continuous speech data. To create self-driving cars, Tesla accumulated 1.3 billion miles of driving data.
Variety
Many more types of data are available than ever before. Traditionally, companies focused their attention on the data created in their corporate systems. This was mainly structured data — data that follows the same structure for each record and fits neatly into relational databases or spreadsheets.
Today, valuable information is locked up in a broad array of external sources, such as social media, mobile devices, and, increasingly, Internet of Things (IoT) devices and sensors. This data is largely unstructured: It does not conform to set formats in the way that structured data does. This includes blog posts, images, videos, and podcasts. Unstructured data is inherently richer, more ambiguous, and fluid with a broad range of meanings and uses, so it is much more difficult to capture and analyze.
A big-data analytics tool works with structured and unstructured data to reveal patterns and trends that would be impossible to do using the previous generation of data tools. Of the three Vs of big data, variety is increasingly costly to manage, especially for unstructured data sources.
Velocity
Data is coming at us faster than ever. Texts, social media status updates, news feeds, podcasts, and videos are all being posted by the always-on, always-connected culture. Even cars and refrigerators and doorbells are data generators. The new Ford GT not only tops out at 216 miles per hour, it also has 50 IoT sensors and 28 microprocessors that can generate up to 100GB of data per hour.
And because it’s coming at us faster, it must be processed faster. A decade ago, it wasn’t uncommon to talk about batch processing data overnight. For a self-driving