THE FIRST CONSCIOUS MACHINES WILL PROBABLY BE ON WALL STREET
We must consider the possibility that intelligence, creativity and even consciousness are purely functions of the material world, with human beings as a peculiar kind of computer. In a world operating under this assumption, machines can theoretically have directed cognition, decision-making and consciousness. Even today we see supercomputers owned by financial institutions making trading decisions on behalf of the companies that own them. These are specialized machines that do something that produces similar results to cognition.
The fact that they are specialized thinking machines might lead one to believe that this precludes them from consciousness. I think the opposite is true; human beings are not generalized computing organisms. The machines in question, just like humans, are not general purpose beings, but highly selective imitation devices with an innate dedicated language system.
The financial industry is always on the bleeding edge of technological application. Always. Beyond ticker tape, the telecommunications revolution and mere computer algorithms, today’s Wall Street is the first and perhaps only industry putting artificial intelligence toward actual productive ends. Machine trading, which today mostly falls under high-frequency trading, (HFT) accounts for 73% of US equity trading volume, an increase from 25% of equity trading volume only five years prior.
HFT machines make dozens or even hundreds of trades within a second, which far outstrips the ability of a human or any group of humans to make such a decision. The Sharpe ratio, which is used in the financial world to indicate reward against risk, is much higher with HFT than with traditional strategies. Vast profits can be made by these trades being made before others, whether it is a few milliseconds before competing traders, or executing trades before others even realize it.
The supercomputers executing these strategies look for various signals, such as volume, volatility, changes in global interest rates and tiny economic fluctuations. But beyond quantitative data, news articles, tweet and other qualitative information accessed through the internet is taken into the calculations by use of natural language processing system that convert mined text into meaning for the machine.
A recent example of this artificially intelligent news analysis happened when Associated Press had its twitter account hacked, making a tweet on April 23, 2013, falsely asserting that there was an explosion at the White House. The S&P Index lost $136 billion in a matter of four minutes, though it was recovered just as quickly. Wallace Turbeville notes:
Most trading of securities and derivatives is accomplished using supercomputers wired directly into exchanges and other venues. They operate at trading speeds well below milliseconds so no human is involved. The trades are dictated by artificial intelligence software… pattern recognition software that infers motivations and other characteristics of other traders in the markets to pick which ones to exploit. Another element of the system is software that reads data, including Twitter traffic, for key word combinations so that the supercomputers can fly into action within, let’s say, one ten thousandth of a second of the appearance of the words. I am going to go out on a limb, here – I suspect that a tweet that comes from a “verified” Twitter account and includes “Obama”, “White House”, and “bombs” might qualify as a sell-triggering word combination.
Outside of quantitative finance, quasi-artificially intelligent programs known as expert systems reason independently make conclusions are used across various fields. These systems are often modeled on the cognitive processes of human beings. The cognitive model of expert systems in finance generally outperforms not only traditional equation-based based analytics, but unaided human actors. News analytics, an example of which was mentioned previously, is a concept in finance where natural language processing and artificial neural networks are used to allow machines to find relevant news sources, derive meaning from them and decide which data is the most meaningful.
Machines seem to outperform humans when classifying words or phrases. “…standard machine learning techniques do better than humans at classification.” Expert systems that are designed too similarly to human cognition can fail because of the same reasons that cause human decision makers fail. Roy S. Freedman lists the reasons as:
- This is the tendency not to stray from an initial judgment even when confronted with conflicting evidence. Humans are reluctant to revise their opinion in light of experience. In expert systems, this is seen in the difficulty to revise default assumptions in non-monotonic reasoning.
Inconsistency. Humans tend to violate properties of both exclusivity and transitivity of comparison: if a pair of alternatives is presented to a subject many times, successive presentations being well separated by other choices, a given subject does not necessarily choose the same alternative each time. In expert systems, this is seen in the representation of fuzzy and probabilistic reasoning.
- This refers to using only a portion of the information available. Human analysts make poor decisions when they must take into account a number of attributes simultaneously: decision-makers may be aware of many different factors, but it is seldom more than one or two that they consider at any one time. One effect is that experts are often influenced by irrelevant information.
- This refers to the improper use of probabilistic reasoning. Common errors include the failure to revise prior probabilities sufficiently based on new information and the discrepancy between subjective probability and objective probability.
- This refers to the focusing on how closely a hypothesis matches the most recent information to the exclusion of generally available information
Being “too close” to human is clearly detrimental to profit, even without consideration for speed of decision making and execution. This has some serious implications — if and when machines become self-aware, they will be self-aware in a way that is alien to our understanding of consciousness. On the other hand, human-like learning and reasoning is necessary, since reliance on any strict algorithm is also detrimental, in the face of market realities changing and competitors updating their strategies. “In order to stay ahead of the competition, firms must constantly alter their algorithms.” This has led to the use of machine learning in financial expert systems, where the systems can update themselves as they receive new information.
An innovation occurring in this area has occurred in the form of artificial neural networks. Artificial neural networks are computational systems that are inspired by the neural connections in the brains of animals. Different virtual neurons are connected via rules forming a network which information is fed through.
This process leads to an effective form of machine learning that does not require a human operator or supervisor. Artificial neural networks have an inherent capability to adapt the network parameters to the changes in the studied system. A neural network trained to a particular input data set corresponding to a particular environment can be easily retrained to a new environment to predict at the same leveling real time; that is to say, as soon as previously meaningless data comes in.
The hyper-specialization, learning capabilities and potential for autonomous existence seems to indicate that Wall Street will have the world’s first self-aware machines. It is interesting to note the developments in artificial intelligence financial systems have taken a turn towards theorized architecture of the human mind. Sanjiv R. Das, when developing the aforementioned news analytics systems, used multiple separate competing systems for its semantic classification process, which is reminiscent of parallel processing memes. To mitigate error, classifiers are first separately applied, and then a majority vote is taken across the classifiers to obtain the final category. This approach improves the signal to noise ratio of the classification algorithm.
If memes are indeed the foundation of consciousness, as Daniel C. Dennet and a considerable amount of research suggest, to back then the combination of learning and memory from artificial neural networks and competing cognition schemes from Das’ developments, then it seems likely that these systems could acquire conscious thought. Indeed, the fact that these systems are not general purpose does not at all preclude their ability to develop cognition similar to humans.
Human beings, like specialized intelligent machines, have categorization systems that are biased towards discriminating some objects and actions rather than others. These similarities indicate that cognition and consciousness like humans could exist in even specialized machines with learning capabilities. We are on perhaps the early stages of a world where diverse machines make conscious decisions concerning business, markets, and society in general, replacing humans in roles which once required a human thinker and executor. Instead of looking toward academics for the dawning of this age, we should keep an eye on the hotshots in Manhattan.