Understanding Artificial Intelligence
When asking whether a computer is intelligent, a useful starting point is to ask what kind of human intellectual work the computer is being designed to perform, rather than whether it thinks like a person.
The Programming Question
Classical programming begins with rules written by humans. The computer receives those rules and data, follows the rules, and produces answers. Machine learning begins by asking a different question: can a computer examine data, discover useful rules, and apply those rules to information it has not seen before?
Two Routes from Input to Answer
In classical programming, a human supplies two things: a program containing rules and data for those rules to process. The computer follows the supplied rules and produces answers. The path from input to output is designed by the programmer before the computer runs the program.
Machine learning rearranges these inputs and outputs. Humans provide data together with the answers expected for that data. The computer uses this material to produce rules. Those learned rules can then be applied to new data, producing answers for cases that were not part of the original material.
From Examples to Predictions
Classifying messages
Build a system that gives one of two category answers for messages.
- 1Provide examples
A human gives the computer many messages as data and supplies the expected category answer for each message.
- 2Produce rules
The computer uses the paired messages and expected answers to learn rules.
- 3Receive new data
A message that was not part of the original material arrives.
- 4Generate an answer
The learned rules are applied to the new message to produce a category answer.
The system moves from human-provided examples to computer-produced rules and then uses those rules to answer a new case.
The important point is not that a human has written every decision in advance. The human has supplied material from which the computer produces rules. The rules are then used beyond the original examples, on new data.
The Historical Question
The shift toward machine learning connects to a historical question about what computers can do. In 1843, Ada Lovelace argued that Charles Babbage’s Analytical Engine could perform whatever people knew how to order it to perform, but could not originate anything. Alan Turing later discussed this idea as Lady Lovelace’s objection while considering whether general-purpose computers could learn and show originality.
Machine learning addresses the learning part of this question by changing where the rules come from. Instead of requiring people to explicitly order every rule, it asks whether a computer can examine data, discover useful rules, and use them on information it has not seen before. This does not remove the role of human-provided data and expected answers; it changes the route by which rules are produced.
What do you think happens?
Which change most directly represents the machine learning response to Lady Lovelace’s objection?
Reveal answer
The computer produces rules from data and expected answers.
Machine learning changes where the rules come from: the computer uses human-provided data and expected answers to produce rules that can later be applied to new data.
What Artificial Intelligence Includes
Artificial intelligence is the effort to automate intellectual tasks normally performed by humans.
This definition makes artificial intelligence a broad field rather than the name of one particular technique. When asking whether a computer is intelligent, a useful starting point is to ask what kind of human intellectual work the computer is being designed to perform, rather than whether it thinks like a person.
Rules and Fuzzy Problems
Symbolic AI depends on explicit rules handcrafted by programmers. This approach is suitable for problems whose logic can be described clearly. When the decisions and knowledge can be stated directly, explicit rules provide a direct way to describe the solution.
The difficulty changes when a problem is complex and fuzzy. For image classification, speech recognition, and language translation, it proved intractable to work out all the explicit rules needed to describe the solution. Machine learning arose as an approach for these kinds of problems.
Common Misunderstandings
Common mistake Treating artificial intelligence and machine learning as synonyms.
What it looks likeCalling every AI approach machine learning.
Why it failsArtificial intelligence is the broader field. It includes machine learning, symbolic AI, deep learning, and approaches that do not involve learning.
Do this insteadDescribe machine learning as one approach within artificial intelligence.
Common mistake Assuming machine learning means that humans provide no guidance.
What it looks likeSaying the computer creates rules without any human-provided material.
Why it failsThe machine learning model described here begins with data and expected answers supplied by humans.
Do this insteadRemember that humans provide the data and expected answers, while the computer produces rules from them.
Common mistake Assuming that every problem should be solved by explicit rules.
What it looks likeExpecting programmers to write all the rules for image classification, speech recognition, or language translation.
Why it failsFor complex and fuzzy problems, working out all the explicit rules needed to describe the solution proved intractable.
Do this insteadRecognize machine learning as an approach that arose for these complex, fuzzy problems.
Common mistake Defining artificial intelligence as a computer that thinks like a person.
What it looks likeJudging AI only by whether its internal process resembles human thinking.
Why it failsA useful starting point is to ask what human intellectual task the computer is designed to automate.
Do this insteadDefine AI in terms of automating intellectual tasks normally performed by humans.
Check Your Understanding
A team wants a computer to classify new messages into one of two categories. They can provide many messages and the expected category for each one, but they do not describe every classification rule themselves. Explain which programming approach fits this setup and trace the path from the supplied material to the answer for a new message.
Hints
- Identify whether humans provide explicit rules or data with expected answers.
- State what the computer produces from the supplied material.
- Explain what happens when a new message arrives.
Explain why symbolic AI is a natural fit for a problem whose logic can be described clearly, while machine learning may be useful for a complex, fuzzy problem such as image classification, speech recognition, or language translation.
Hints
- Start with the source of the rules in symbolic AI.
- Then consider why writing every rule becomes difficult for complex and fuzzy tasks.
- Place both approaches within the broader field of artificial intelligence.
Key Takeaways
- 1Classical programming uses human-written rules to process data and produce answers.
- 2Machine learning uses data and expected answers to produce rules that can be applied to new data.
- 3The machine learning idea connects to the historical question of whether computers can learn rather than merely follow explicitly ordered instructions.
- 4Artificial intelligence is the effort to automate intellectual tasks normally performed by humans.
- 5Symbolic AI uses explicit handcrafted rules, while machine learning arose for complex and fuzzy problems where describing every rule is difficult.
- 6Machine learning is one approach within the broader field of artificial intelligence.
At a glance
- Classical programming starts with rules written by humans; machine learning produces rules from data and expected answers.
- Learned rules can be applied to new data to generate answers for cases outside the original examples.
- Artificial intelligence is a broad effort to automate intellectual tasks normally performed by humans.
- Symbolic AI suits clearly described logic, while machine learning arose for complex and fuzzy problems.
- Machine learning is one approach within artificial intelligence, not a synonym for the entire field.