Though we spend our time discussing how to implement AI to improve businesses, we still get asked “What is AI?”. The quick answer is: a machine demonstrating intelligence. Then what’s intelligence? According to Wikipedia, it is the ability to convert data to information and to apply it. That is a pretty broad definition so this is not a very helpful answer. That definition makes pretty much any software intelligent which is wrong. AI is a sub-field of computer science not another name of it.
A better layman’s description would be “a machine that can demonstrate human like intelligence in a very specific or broad field”. While this is more understandable and prevents us from categorizing all software as AI, it is extremely subjective. What does human-like mean? First, humans have a huge variation of skills, I am certainly no Einstein. Second, as soon as machines become as good or better than humans at a certain task, accomplishing that task no longer earns the human-like title any more.
For example, is a state-of-the-art transcription system, which has writing and listening skills beyond human levels, an AI system? Well, to a scribe from Middle Ages, such a machine would have looked like a piece of magic. But to a modern man, it seems to be doing a pretty straight-forward task.
However, these are not unsolvable issues. First, we can solve the variation of human skills problem, by assuming top human capabilities today represent a good yard-stick for comparing AI against. Second, to solve for our bias for recognizing machine abilities as non-intelligence, I think we can focus on whether a system is learning or not. Undoubtedly, regardless of whether machines are capable or learning or not, humans will always see learning as a fundamental part of human intelligence. After all, scientific progress, pillar of our modern achievements, is based on the foundation that there are things we do not know but that we can learn them through the scientific method.
So finally, we have a compact, yet powerful description. An AI system is a learning system and we can measure its effectiveness by comparing it against the best performing humans in the same field.
This solves for some other questions about AI as well. For example, is solving simple problems also the realm of AI? The issue of simple vs complex is a difficult one to get into as simple and complex are points in the spectrum of complexity. Very simple mathematical constructs like Turing machines have been proven to be capable of solving immensely complex problems. But when we consider that learning is required for AI, we already exclude simple problems outside the scope of AI as simple to solve problems would not require us to build learning systems.
A criticism of this approach would be that it doesn’t include rule-based or completely human programmed systems as AI while history of AI up until 2010s was dominated by such efforts. However, learning systems have proven much more versatile than human programmed systems and most of current AI research is focused on learning so considering the modern AI field, this is not a bad definition of scope for AI.
History of AI
History of AI has been marked by euphoria of optimism based on high expectations and lack of funding when those expectations failed to be met. Though replicating the mind has been an area of interest since the antiquity and scientists like Turing made significant theoretical contributions to the field, AI research backed by significant funding only started around 1956 after the Dartmouth Summer Research Project on Artificial Intelligence. Winning at the game checkers, solving word problems in algebra and proving logical theorems were some AI breakthroughs from the period. Though significant such breakthroughs were achieved in 1950s and 1960s, the initial promises of AI research remained unfulfilled in 1970s leading to a scarcity of funding called the AI winter.
For example, neural networks that form the foundation of most of today’s AI approaches were invented in 1957. Though initially such networks were believed to be capable of human level functions, those beliefs were soon shattered. In 1969, AI researchers Marvin Minsky and Seymour Papert published Perceptrons where they proved that neural networks with only a single layer were incapable of learning even an XOR function, a basic logic function. Given constraints on processing power in 1970s, it was not feasible to build neural networks with enough layers to mimic even slightly advanced intelligent functions.
Since 2000s, machine learning, a subbranch of AI, has received significant interest and AI surpassed best of humanity in a number of games such as Jeopardy, Go and some video games.
Currently AI is increasingly becoming a topic of general interest and debate. Anyone from tech CEOs to pundits are joining the debate. A recent topic of interest was the spat between Elon Musk and Mark Zuckerberg. While Elon is firmly in favor of more AI regulation and Mark is vehemently opposing negative perceptions of AI. Stephen Hawking also supports a cautious approach to AI in numerous articles, interviews and panels, fearing machines will supersede humanity.
Future of AI
Though we hate pundits and punditry, we have a few data backed observations about the future of AI.
First, dramatic advances in AI tend to follow periods of little progress. Until now hardware has been the limiting factor in AI. In 1970s, scientists could not build complex neural networks as processing power was too expensive. In 1980s, rules-based AI started gaining popularity as we still did not have sufficient processing power for purely data based techniques. We can say that we really started using models from 1950s only in 2000s. Since computational power tends to grow with Moore’s law, to progress faster than that, scientists need to come up with more effective learning techniques and famous AI researchers like Geoff Hinton are hard at work on that field.
If past is any indication, even if new techniques are discovered, it may be a long time until we have enough computational power to use them. After all, human brain is estimated to have more computational power than the best super computers and we don’t exactly know the computational capacity of the human brain yet. And given the efficiency of the evolution process, we are unlikely to come up with a more efficient structure than our brain.
Secondly, even smaller advances in AI can unlock economic miracles. Self-driving cars are already on the roads and if we can improve their effectiveness by a few percentage points, they could add hours to many commuters lives. Even though such breakthroughs are merely incremental technical advances, they will require significant investment, engineering, testing, regulatory approvals. However once unleashed they will generate billions in value and keep interest in AI research active.
AI applications areas
Though natural language processing or machine vision are solving similar problems (i.e. interpreting raw data to produce insights), techniques used in both are specialized to produce the best outcomes in each area. These are the major areas of AI application:
- Natural language processing: Processing natural language to understand it completely with all its nuances and subtleties. We explained why this is such a hard problem.
- Machine vision: Recognizing images. This has a myriad of applications from vehicle registration plate recognition to facial recognition.
- Machine learning: When not all input data is images or language, machine learning solutions deal with numerous different input data to create insights.
So how can AI help you?
Regardless of your job, AI is there to save you time and improve your effectiveness. For example, an HR professional can leverage AI systems to improve hiring, churn management and other areas. Who wouldn’t want to get a great, automatic CV filter that also uses data from the web? Who wouldn’t want their performance management systems augmented with objective data?
HR is just an example, there are AI solutions for every professional. Take a look at our use case list and filter it by the business area you are interested in, such as marketing, sales or customer service. We have use cases for 10+ business functions so we believe you will find what you are looking for.