The goal of Artificial Intelligence (AI), from its very start in the early 50’s, was to solve problems that we perceive as requiring intelligence to solve. A group of luminaries from Carnegie Mellon University (CMU, Newell and Simon), from MIT (McCarty and Misky) and from IBM (Samuel) gathered at a workshop at Dartmouth College in 1956 and laid the foundations for this new field, with a great sense of optimism for what could be achieved in the next few years.
The Starting Years
It’s rare that the start of a particular field of research can be identified so precisely, but, aside from the general goals and enthusiasm, no specific approach or methodology could be established and the field quickly splintered into different areas that seemed promising to researchers. These centered around problems that could have practical applications, such as planning and scheduling and pattern recognition, or which could be viewed as “challenges” to human intelligence, like playing chess. The general approach was to build “algorithms”. i.e. sets of instructions that can be programmed in a computer, based on mathematics and logic. Most solutions were limited to “toy problems” which could generally not be scaled up to real-world problems. Note that no consideration was given to the functioning of the brain, as the thinking was that all problems could be solved “in the abstract” independently of the underlying mechanisms that could “produce” thinking, such as the brain.
Knowledge Representation and Expert Systems
Interest in AI, as measured by government funding, began to dwindle in the 70’s (one of the “AI winters” as these periods have been somberly referred to by AI researchers), until a new important concept emerged that found practical applications: essentially the notion was that, if we could gather the knowledge of experts into sets of “rules”, we could navigate down a path of “if-then” options and get to the same answer an expert would give. This was essentially the same process physicians would follow in their mind, by asking you about your symptoms to come up with a diagnosis for your problem, or a car mechanic to know what to repair.
Since they were trying to mimic the mind processes of experts, these systems became known as “Expert Systems” or “Knowledge Based Systems”. The research challenge moved from algorithms that gave rigorous mathematical solutions and could be “proven” correct, to “heuristics” or “rules of thumb” for representing knowledge and navigating through it to find a solution. Edward Feigenbaum from Stanford University became known for his mantra the “The power is in the knowledge”, i.e more than in inference procedures.
“Real World” AI
I was working at NASA in an AI group in the late 80’s and 90’s and, proceeding according to the same philosophy, I led a project, in collaboration with MIT scientists, for an Expert System that could help astronauts conduct experiments in space, by giving them guidance, alerting them of problems or unusual results, etc. We called it “PI-in-a-box”, where PI stands for “Principal Investigator”. It flew in SLS-2 (Space Life Sciences) on the Space Shuttle in 1993. It was the first Expert System to help astronauts conduct science in space.
The reality is that over time all “real world” complex AI systems combine both mathematical and algorithmic aspects with knowledge representation and large amounts of data. This complexity has been made more and more possible by increased computational power and memory. Two examples of success achieved by these combined approaches were both IBM projects: “Deep Blue” that defeated a world chess master for the first time (Kasparov) in 1997 and “Watson” that defeated the two greatest Jeopardy champions (Rutter and Jennings) in an exhibition game show.
Another important concept that emerged from AI is the idea that systems could actually “learn” from interacting with the world, i.e. seeing “data”, as humans do, as opposed to being programmed to behave in completely predictable ways. These systems were originally mostly based on symbolic and statistical approaches and they were at times seen as “outside AI”, hence the separate name “Machine Learning” (ML). The need to distinguish the two fields was partly due to the reluctance of major government agencies to fund projects that produced solutions that might not be “proven” correct, or even optimal. From the point of view of NASA and the military, people could make mistakes, but not machines. To some extent, this is still the prevailing culture.
The Rise, Fall, and Re-Emergence of Connectionism
My original intent and curiosity in becoming involved in AI was to understand what “makes us” intelligent, or, basically, how are brain reasons. It was very surprising to me, that none of the approaches used by AI and ML had anything to do with how our brain actually functions. These fields could be viewed as offshoots of Mathematics and Computer Science. Neurophysiology was completely irrelevant, until work originally referred to as “artificial neural networks” began to emerge in parallel to AI. This work was based on a very simplified model of neurons and synapses, and the fact the signals to and from neurons could be regulated by changing the synaptic weights. Networks of simple artificial models of such units were shown to be able to “learn” input-output mappings of complex functions. For instance, simple images (e.g. characters) could be identified after training with a sample set of images. This was a fundamentally new paradigm: one did not need to provide the system with rules for finding the answer. One simply had to provide a set of examples. Unfortunately, this also meant that, just like for ML, one could not guarantee correctness, and, worse yet, one could not even “explain” how the answers were obtained.
This general approach had been known as “connectionism” and was viewed as separate and not as promising or reliable as AI… until about 2011, when a combination of new neural network architectures several layers deep (hence the new name “Deep Learning”) with the processing power of new specialized computer chips (Graphics Processing Units) demonstrated that the systems could outperform humans in several image recognition tests. So, ironically, what was considered “traditional AI” has been largely incorporated in software systems we use every day, with no particular AI label, while “Deep Learning” is now considered state-of-the-art AI! We will return to the future of AI in a later lesson.
Share with friends