Artificial Intelligence:  Machines which "Think"

• this raises the question of what we mean by "thinking"  Is doing calculations enough?  Playing chess?  Figuring out taxes?  These are not what we intuitively mean by "true intelligence"
• on the other hand, the term "computer" used to describe a job or profession for a human being who did computation for a living
• intelligence is not all or nothing, it is more or less.  At what point will a computer be as intelligent as we are?  Turing test?
• are chatterbots intelligent?  Try some and see what we think.

• both these approaches model part of how the human brain works
• each is useful for some tasks, not so useful for others
• rule-based AI follows a computer design by Jon VonNeumann, a brilliant mathematician and logical theorist, the inventor of game theory
• this design has one unit with a set of logical rules, another with a lot of memory.  It goes through the following steps:
• fetch an instruction from memory.
• fetch any data required by the instruction from memory.
• execute the instruction (process the data).
• store results in memory.
• fetch another instruction from memory
• this approach is sometimes called "recursive" because it means doing the same thing over and over, taking into account the results you got the last time
• a neural net is an interconnected assembly of simple processing elements, units or nodes, whose functionality is loosely based on the animal neuron. The processing ability of the network is stored in the inter-unit connection strengths, or weights, obtained by a process of adaptation to, or learning from, a set of training patterns.
• rule-based programs are good for tasks for which we know many specific rules, e.g., filling out taxes, making medical diagnoses from lists of symptoms
• neural net programs are good for tasks which we learn to do without knowing how we do them, e.g., recognizing handwriting

• Evolutionary algorithms are another approach to AI
• The basic idea here is to write programs which generate hundreds or thousands of new programs
• These programs compete with each other to accomplish goals set by the programmer
• The best programs survive, going on to build additional programs.
• Instead of just telling the computer what to do (rule-based AI) or letting it figure it out on its own (neural nets), evolutionary algorithms say "let the fittest software survive"

This mimics the evolutionary process in other realms.  Some scientist thinks that the brain itself functions through evolutionary processes.  The thoughts that do better survive, the others fall by the wayside

What will the future of AI be like?

• Many experts believe that future programs, will be more complex, combining rule-based, neural net, evolutionary and other perspectives in the same programs
• The Webmind software is one attempt to do this, combining combining insights from both the rule-based, neural net and evolutionary approaches, as well as on statistical or stochastic models
• Webmind is loosely modeled on the brain's structure, relationships between large groups of neurons, each of which is specialized
• Each of the nodes within Webmind follow rules, but the units interact with each other in ways which are analogous to how neurons in a neuronal net interact

• Here we have room for imagination.  There are many examples in science fiction, some of which are quite frightening.
• In the case of Webmind, the first thing we are asking it to do is to predict financial markets.  Was this a good choice?  Why did we those this assignment for Webmind?
• What would you like the AI of the future to do?
• What would you prefer that it NOT do?
• Should AI be stopped?  Are we building a monster which will destroy us?
• We will discuss these questions in our support groups over the break