In 1821, economist David Ricardo posed the machinery question with regard to the then ongoing industrial revolution, which concerns the “opinion entertained by the laboring class, that the employment of machinery is frequently detrimental to their interests”. Today in 2017, we are seeing the beginning of a new revolution that once again raises the bogeyman of worker displacement, and societal upheaval. As the byline of this post suggests, the future of society is perhaps destined to experience another upheaval reminiscent of the industrial age: namely, the advent and impact of Artificial Intelligence (AI).
When we think of AI in 2017, we still often think of popularized Hollywood versions of android robots like The Terminator, Data from Star Trek, or Ultron from the Avengers. While those generalized (Strong) forms of AI are not yet possible, real advances in AI in recent years have demonstrated amazing promise that has energized the AI community, corporations, government, and investors. The promise of AI today comes in several flavors characterized by the capability of their respective underlying architectures. In order to understand the threat, or promise, of AI in the coming years ahead, let’s discuss the details of some of those approaches, and see what conclusions that may lead to.
Today, all current forms of AI implementation are categorized by experts as Weak (or Narrow) AI. Intelligence similar to or exceeding human level, so called Strong (or General) AI, doesn’t yet exist and is unlikely to exist for decades to come.
Weak AI is skilled at only one specific task. While it may do that one task very well, sometimes beyond human ability, that’s all it can do. Due to that fact, it is not general intelligence, in the sense of approximating human intelligence. One example of Weak AI would be IBM’s Deep Blue, which beat chess grand master Garry Kasparov in 1996. Deep Blue, which was an early form of Weak AI, did not learn – it was essentially pre-programmed with a set of rules. More recently, Google DeepMind’s AlphaGo defeated Lee Sedol at the more complex game of Go in 2016. While still considered Weak AI, AlphaGo was trained on a large dataset of expert moves, through the added technique of Machine Learning (ML).
There are many current or developing Weak AI applications, including self-driving cars, voice-recognition systems, fraud detection in banks, clinical decision support systems for medical diagnosis, and many more. Depending on the application, these AI systems typically involve a combination of neural nets and pattern recognition, often in combination with training datasets and/or knowledge databases, which may or may not include semantic relationships. A key to these more modern Weak AI systems is the large training dataset, the larger the better. However, the defining characteristic of Weak AI remains, in that it only does one thing.
For a long time Weak AI applications that exceeded human ability at a task were a bit of a pipe dream. They needed to be trained on large datasets in order to exhibit the level of intelligence required to do a task correctly. Not only were these datasets hard to come by, they also needed immense processing power to digest. Recent advances in parallel processing technology such as Big Data frameworks, and GPU technology, have given a massive boost to the ability of ML to process large datasets. Along with these advances came new advances in AI, which are popularly termed Deep Learning (DL). These advances came about as a result of advances in the types and complexity of Neural Net algorithms, in conjunction with the aforementioned massively parallel processing capability. DL neural networks classify data and can make predictions based on ingesting a typically large training data set correlating data with labels. In some cases, DL is combined with Reinforcement Learning (RL), which introduces the added capability of the AI exploring possible reward mechanisms depending on what action it chooses based on its current state.
DL systems are often based on an architecture that can be reused for different purposes. For example, IBM’s Watson, which defeated the worlds top Jeopardy players in 2011, is based on the DeepQA architecture. The DeepQA architecture could be (and likely will be) put to other uses in the future.
While DL is essentially still Weak AI, it has created a lot of excitement in the AI community, since it appears to be another step towards Strong AI. It has also received some criticism, notably that it lacks theory to substantiate some of its claims. Also, there are cases noted where DL makes simple mistakes that could be due to limitations in the internal representations learned by these architectures. The DL industry is still in an experimental stage, however, in some cases the experiments are already paying off. Google’s purchase of DeepMind in 2014 for about $600M US was in part repaid by using DeepMind to introduce more efficient monitoring of power usage, which in turn is saving Google hundreds of millions of dollars over multiple years.
Training Weak AI
Training Weak AI applications to become experts in their field involves supervision and feedback, human or otherwise. Naturally, there is some interest on the part of AI companies in minimizing the amount (cost) of training for Weak AI machines, and there is also an interest in making sure that the training is sufficient and accurate. The sufficiency and accuracy requirement particularly holds true when machine errors could cause catastrophic results, such as a failure of a self-driving vehicle, or a medical monitoring system. Learning methods abound, yet they are generally organized under the following categories, which relate to the amount and type of supervision required for the learning:
- Supervised – the training data presented to the ML algorithm is fully labeled, often by humans. That means: all examples are presented with a classification that the machine is meant to learn. Labeled training data can be expensive to produce, and thus it may be limited in quantity.
- Semi-supervised – in addition to the labeled examples from the supervised learning, the ML algorithm is allowed to additionally take unlabeled data into account. In this approach, there are typically far more unlabeled examples than labeled examples. Semi-supervised systems often outperform their supervised counterparts using the same labeled examples. The reason for this improvement is that more unlabeled data enables the system to model the inherent structure of the data more accurately.
- Unsupervised – the ML algorithm is not provided any labeled training examples at all. There is no objective evaluation of the accuracy of the structure that is output by the relevant algorithm, and the resulting relations/categorizations may not be easy to map to relations needed for a particular knowledge base. The benefit is that you avoid the overhead (cost) of labeling training data.
- Weakly-supervised – (aka Bootstrapping) uses even fewer labeled examples than semi-supervised learning. It starts with a few training examples, and trains a ML classifier, using thought-to-be positive examples as yielded by this classifier for retraining. As the set of training examples grows, the classifier improves, provided that not too many negative examples are misclassified as positive, which could lead to low precision and semantic drift.
- Distant supervision – refers to using an outside source (database) as a reference in place of labels for training data. For example, Freebase which is a large semantic database, can provide distant supervision for relation extraction. The inference is that any sentence in the training data that contains a pair of entities that participate in a known Freebase relation is likely to express that relation in some way as it relates to the training data.
Many Weak AI systems, although trained w/ massive datasets on powerful parallel processing computers, can actually run on small computers in the field e.g. self-driving cars, drones. In other words, it’s the training that requires all the processing power. This makes AI more like a smartphone than a piece of expensive manufacturing equipment i.e. the value is in the intellectual property, not the physical object, which can be mass produced at low cost.
Cyc is an AI project that deserves some special mention, as it was begun in 1984 in an attempt to assemble a comprehensive ontology and knowledge base (KB) of everyday common sense knowledge, with the original goal of enabling AI applications to perform human-like reasoning. In 2016, Cyc was made available to AI researchers under a research-purposes license, and it has also been commercialized under the banner of Cycorp.
An example of semantic knowledge represented in the Cyc KB is “Every tree is a plant” and “Plants die eventually”. When asked whether trees die, the inference engine can draw the obvious conclusion and answer the question correctly.
From the Cyc site:
“At the present time, the Cyc KB contains over five hundred thousand terms, including about seventeen thousand types of relations, and about seven million assertions relating these terms. New assertions are continually added to the KB through a combination of automated and manual means. Many more concepts can be expressed functionally, thereby enabling the automatic creation of millions of non-atomic terms, such as (LiquidFn Nitrogen) being used to describe liquid nitrogen. Additionally, Cyc adds a vast number of assertions to the KB by itself as a result of the inferencing process.”
As a semantic KB, Cyc can also be used as a Distant Supervision partner for training DL machines. In this way, it can impart commonsense everyday human knowledge to the learning process.
AI and The Industrial Revolution
Despite the fears expressed by David Ricardo in the early 1800’s, among others, that “substitution of machinery for human labor…may render the population redundant”, the overall effect of mechanization turned out to be job creation on an unprecedented scale throughout the late 19th century and well into the 20th century. Note that the jobs were created on an unprecedented scale in the latter half of the 19th century, while the job displacements began in the first half of the 19th century. This gap, known as Engels’ pause, is a byproduct of the two stage process that occurs between the invention of technology that enriches capitalists at the expense of displaced workers, followed by reinvestment back into the economy by capitalists, which in turn benefits the workers – but not necessarily the workers who were originally displaced, since the pause can take decades.
But what parallels can we draw between the Industrial Revolution and the advent of AI?
A widely cited study by Carl Benedikt Frey and Michael Osborne of Oxford University, published in 2013, found that 47% of jobs in America were at high risk of being “substituted by computer capital” soon.
In 2015, Bank of America Merrill Lynch predicted that by 2025 the “annual creative disruption impact” from AI could amount to $14 trillion-33 trillion, including a $9 trillion reduction in employment costs thanks to AI-enabled automation of knowledge work; cost reductions of $8 trillion in manufacturing and health care; and $2 trillion in efficiency gains from the deployment of self-driving cars and drones.
Ironically, at the same time, the Wall Street Journal reports that “In the next 10 to 15 years, we expect U.S. employers to demand more labor than will be available, which will, in turn, constrain overall economic growth”.
Taken together, this implies a shift in the job market, with some jobs being displaced by technology while others will be under resourced. It will be of utmost importance to identify which industries to train people for, and establish programs to train new graduates, as well as retrain existing workers. For further information on the potential impact of Weak AI on the US economy see the following reports to the Executive Office of the President: Preparing for the Future of Artificial Intelligence (October 2016), and Artificial Intelligence, Automation, and the Economy (December 2016).
Is AI a Danger?
Weak AI has rounded a corner, with the elimination of some previous inhibitors to progress. We’ve seen:
- great leaps in parallel processing power
- significant improvements in neural nets, although not of the order that can produce Strong AI
- availability of large training datasets
- development of some knowledge bases that can act as distant training partners
Weak AI poses a threat to job security for some sectors, but based on our previous experience with technology innovation, in the long term it will probably result in greater prosperity overall.
However, some caution that Strong AI, when it’s invented, will entail great risk to humanity. Elon Musk says “With artificial intelligence, we are summoning the demon”, while Stephen Hawking warns that “The development of full artificial intelligence could spell the end of the human race.” Not all scientists take this dire point of view. Peter Norvig, Director of Research at Google, and co-author (with Stuart Russell) of Artificial Intelligence: A Modern Approach, is more optimistic. In a recent interview in Dec. 2016, he says “Five years ago, I was interviewed by a reporter that said ‘how come AI is such a failure?’ Today, they say ‘how come AI is going to take over the world and kill all humans or steal all our jobs?'”
Then there is the topic of emergent AI, which is Strong AI that we didn’t consciously invent at all. It just emerged on its own, much like biological intelligence (our own) emerged through evolution. Currently, this concept gets short shrift from leading AI thinkers. Former head of Research at Google, Alfred Spector, once commented in 2013, “[AI] just happens on its own? I’m too practical – we have to make it happen”.
AI and Society
So will Weak AI be economically traumatic? Probably, in some sectors. Witness the effect of companies like Uber on the incumbent taxi system. Now imagine it with AI drivers i.e. no human drivers. Will we be able to ultimately predict the effect of AI on society? I think we’ll be able to predict on a sector by sector basis what will happen as new Weak AI capabilities appear on the horizon, although we may not get much advance warning as the incubation period of these capabilities could be quite short.
On the other hand, I think our ability to predict the impact of Strong AI is quite limited in comparison, since Strong AI could potentially alter the quality and quantity of available brain material i.e. we could become exponentially smarter and have the equivalent of millions of super-geniuses on the planet. It’s pretty hard to imagine what those brains might come up with.
Glenn Reid is the CEO of RJB Technology Inc., a Canadian firm with Branch Offices in Makati, Philippines. Contact us today for more information about our company, or to discuss your custom software development needs.