THE BRIGHT FUTURE OF ARTIFICIAL INTELLIGENCE
The height of evolution on Earth in terms of intelligence is us humans – for the moment. The SHIFT 2018 Future of Intelligence track contemplated the future of humanity in a world where artificial intelligence takes the stage to rival the human brain.
CEO and founder of Curious AI, Dr Harri Valpola is one of those experts in the field of AI who see the future of intelligence as bright and encourage us to look ahead with optimism. With over 25 years of industry experience, Valpola steers Curious AI towards creating an artificial intelligence inspired by the human brain and believes that AI can work to the benefit of humanity, as long as we guide the development in the right direction.
Although AI is one of the trendy topics of our age, AI development first began over 60 years ago, in 1956. After an artificial intelligence proved capable of solving equations and learning to beat human players in a game of checkers, AI garnered much interest and development progressed rapidly.
Despite the original hype and visions of the future, however, the industry later came to a standstill, and in the mid-1970s, an “AI winter” started – funding was reduced and research slowed down tremendously. Nor were the first generation of AIs very intelligent by today’s standards: all data was manually entered and all possible outcomes were hardcoded into the AI, which could then make decisions based on the data from the list of choices it was given.
Later, with the advancing development of computers and increasing computational power, AI surfaced again in the public sector. From the late 1990s to the early 2000s, AI development saw a second acceleration, right around the time when Valpola was starting in the industry. With the advent of machine learning, by the turn of the millennium AI had the power to learn from massive amounts of information. This lead to multiple different approaches to AI development using different methods of pattern recognition and data categorization to reach a conclusion.
Machine learning has also made the use of AI possible in multiple different fields. Translators, self-driving cars and image recognition all use AI to reach a decision, to give a few examples. Machine learning is still missing a key ingredient, however, when it comes to human-level problem solving: humans can adjust their decisions and actions to accommodate the surrounding world – something that machines working with a fixed ruleset cannot.
Current-day AIs use systems that are even more complex and further developed. Neural networks can predict upcoming events based on their data, but they still cannot choose the best approach to a problem based on knowing the desired result alone. Instead, they need to iterate through different options and evolve the solution based on the results of these tests.
The missing human component may be found by emulating the human brain. Daniel Kahneman’s book Thinking, fast and slow builds on the widely accepted dual-process theory to explain the processes of human thinking: System 1 is fast, automatic and unconscious, while System 2 is slow, deliberate and conscious. With both of these systems programmed into an AI, the next generation could reach a completely new level of intelligence.
Most current AI systems only use System 1, allowing them to reach decisions quickly. Deep learning systems become extremely fast as they continue to process more and more data, as is commonly seen in image recognition. However, as they lack contextual knowledge about the data they are processing, they cannot make creative solutions.
A few AIs have also incorporated some forms of System 2. AlphaGo is an AI developed for the Chinese game of Go. It has Go’s rules coded into it, and with its System 1 and “near-System 2”, it can analyze the current game situation, decide on subsequent moves, and predict the eventual winner. AlphaGo was developed by first showing the AI games between human players to calculate a median for human players. Then, by playing against a human and by simulating games against itself, the AI was able to learn and become even better, acquiring the “intuition” that only the best human players have.
In the real world, though, instead of a strictly controlled game environment, the number of rules that govern decisions and outcomes is massive. Curious AI’s ambitious goal is to be able to use all that data. Instead of attempting to eclipse human intelligence with an artificial one, Curious AI is building an AI co-worker. By using a combination of planning and learned models, their AI can learn cause and effect and predict what a given action might cause. This information can then be delivered to a human, who in turn can use it to make a better decision.
In an industry application, these AI co-workers could digitize and interpret physical documents, significantly reducing time spent on data entry by humans. With a large network of AI co-workers who also collaborate with each other, human decision making can be augmented and the need to perform mundane tasks reduced, allowing us to focus on more important tasks. Or as Valpola sums it up, “by building more human-like AI, we are allowing humans to be more human.”