AI: where it’s at and where it’s headed
Over the past few years, artificial intelligence has been a hot topic surrounded by excitement and uncertainty regarding how the rise of intelligent machines could change our lives. When most people think about AI, they tend to picture a supercomputer capable of performing any given task considerably better and faster than any human could. Having this kind of tool available could potentially mean that the activities done by human beings could be facilitated, and AI’s increased computing power could also contribute in making scientific and technological breakthroughs faster. However, this general-purpose artificial intelligence more powerful than a human brain is probably decades away from happening, as many AI experts believe that we will not achieve human-level machine intelligence until 2040 or 2050 (https://www.ted.com/talks/nick_bostrom_what_happens_when_our_computers_get_smarter_than_we_are/transcript#t-793982). It is important to mention that human-level intelligence in this context is defined as “the ability to perform almost any job at least as well as an adult human […] not just within some limited domain” (Ibid.).
Even though we are far from creating an artificial general intelligence (AGI), there have been many improvements in AI destined to perform specific tasks, which will be exemplified with some specific cases. The first example of an AI developed to do a specific task is MarI/O, a program that learned how to play and beat the first level of Super Mario World (http://hackaday.com/2015/06/14/neural-networks-and-mario/). This was done through machine learning involving artificial neural networks, which are inspired by the neural networks of animal brains. Having a program play a videogame and beat the first level is noteworthy since it had to learn everything from the basic controls to what it needed to do to beat the level, but its performance was arguably worse than that of a human.
Moving on to the next example, we have AlphaGo, which is a computer program designed to play the board game Go. Unlike MarI/O, AlphaGo was able to perform better than humans and even beat the greatest players of the game. In addition to this, it has introduced innovative strategies to the game as it made moves that were considered unorthodox but had great results (https://deepmind.com/blog/innovations-alphago/). This kind of knowledge is one of the benefits of AI, as it can introduce new points of view that humans are prone to overlook.
Another AI example which can add new insight to certain problems is IBM’s Watson, a question answering computer system. This system is capable of understanding spoken human language and determining the best possible answer using its databases. Watson was initially developed to play Jeopardy!, and it outperformed two of the best players of the game. Now, IBM offers Watson’s services to be used in many industries, such as the medical industry where it can assist doctors in diagnosing diseases in patients. In a test conducted in 2016 by the University of Carolina School of Medicine, Watson analyzed 1,000 cancer diagnoses, and in 99% of the cases it was able to suggest the same treatment plans as the ones suggested by oncologists. Also, in 30% of the cases it found treatment options missed by the oncologists (https://futurism.com/ibms-watson-ai-recommends-same-treatment-as-doctors-in-99-of-cancer-cases/).
These three examples show that task-specific AIs are possible and exist today, now the next step in the pursuit of an AGI would be to broaden the tasks which a single AI can do. However, this can prove to be very difficult as there are many tasks for which there may be many answers, with the correct one being subjective. In the case of MarI/O, AlphaGo, and Watson, the desired outcomes can be objectively classified: you win or you lose, you are correct or you are incorrect. For many other human activities, there may be several outcomes with no clear distinction of the best possible answer, which may also depend on an individual’s judgement and point of view. So, one of the biggest challenges in making an AGI is to determine how the program can understand the context of the task given and discern between the available options and their implications, and based on that choose the one with the best outcome.
2 comments on “AI: where it’s at and where it’s headed”
Comments are closed.
I enjoyed reading your post Rodrigo, and I find the topic Artificial Intelligence very interesting. We both agree that true cognitive computing will not be arriving for a while, but I would argue that current machine learning techniques can providing very insightful predictions and estimates by analysing large datasets to derive underlying patterns and yield high-accuracy predictions for decision makers, and the best part is as more dataset is analysed, the more it learns and it evolves over time. By using current level of artificial intelligence we can guide humans to make better decisions, and create immense business value when used correctly. This combination can be very powerful if used correctly.
You make a good point. As more and more data is generated, and every day at a faster rate, the accuracy of the AI-generated models will surely increase. To further improve the current AIs, a versatile machine learning model should be developed, so that a single AI system may be able to cover more tasks than the few that one can handle today, like playing Go.