Imagine a world in which AI systems are capable of creating AI.
With artificial intelligence (AI) technology becoming more advanced, processes are becoming more and more automated. Hence, robots and other pieces of AI technology are doing the tasks that people once had to do. “The next wave of economic dislocations won’t come from overseas. It will come from the relentless pace of automation that makes a lot of good, middle-class jobs obsolete,” Barack Obama said in his farewell address, noting the current upward trend in advancements in automation. New efforts are being put in place to make AI more automated and efficient, as shown by companies like Google which is working improve AI technology through a project known as the the Google Brain project.
Working with AI technology encompasses certain elements that are not found in conventional programming. For instance, AI programmers deal with the having to use and process large amounts of data to train their model.
Ankita Mitra, a MVHS Junior who has experience with AI programming, shared some thoughts about her experience with AI programming.
“I have done AI in projects that have to do with entertainment and predictive models. Normally in my experience, whenever you trying to get AI to work better and better, the time it takes is really large when you have huge data sets,” Mitra explains.
With new projects like the Google Brain project, programmers that are specialized in AI are able to work collaboratively to improve AI. One of the aims of the Google Brain Project is to develop AI software that can make AI software.
While limiting the current need for AI expertise, this technology could potentially improve the accuracy of the AI technology and the scope of problems that can be addressed using AI technology.
Programmers can use AI technology that creates other AI technology to their advantage. For instance, Google Brain project programmers were able to program a neural network to improve its own encryption methods. A neural network is an system that does computing in a way that is similar to the computing that is carried out by the brain using neurons. The Google Brain Project programmers started with three neural networks and named them Alice, Bob and Eve. Alice’s job was to create a message for Bob to read, and Eve’s job was to eavesdrop on the message that Alice was sending to Bob. Alice and Bob were given a key that could be used to encrypt and decrypt the message while Bob was not. Through multiple runs, Alice worked to improve the method of encryption she used and Bob was able to adjust accordingly. Hence, at the end of all the runs, Bob was only able to guess the fraction of the message that could be guessed with pure luck, showing the efficiency of the system.
However, there are some cases in which AI technology that creates AI technology might not work because of the computer’s inability to change its approach when one approach does not work.
Mitra explains, “The only problem is the way the AI thinks can be very narrow. Like one piece of AI technology is only going to think in a certain specific way, rather than a programmer, who would look at different aspects.”
One of the questions that comes with the development of AI that can create AI is how a robot programming AI compares to a person who is programming AI. The role of the programmer in AI is to guide predictive models to maximize accuracy and find which predictive model is the best, but in the case of AI that can make other AI the AI itself would create the best predictive model itself without the help of a programmer. The accuracy of the AI in choosing the best model would indefinitely be better than the accuracy of a human choosing the best model which is currently how AI works.
“Letting it work completely on its own may be a problem,” Mitra explained, “but if you have a very specific problem that you are trying to solve, the AI will be better than a human.”
For instance, the Google Brain group successfully made a Google Neural Machine Translation (GNMT) system that can translate languages without being taught to translate between these languages. The AI technology was able to create its own language that was used as an intermediary language between the two languages that it was trying to translate. Because of the efficiency of the technology, Google was able to use the neural machine translation system in order to translate between some of the newest language pairs.
So it can be seen that the ability for AI to train itself independently is useful. But how near in the future is this technology? The likelihood of such a technology becoming a reality is increasing because of the increase in the number of resources allocated to improve AI. OpenAI and the Google Brain project are two groups that have allocated resources for improving AI. OpenAI a non-profit organization that currently has 45 members and works on the improving the efficiency of AI technology.
Ankita explained, “I think we are going to see this technology in two to three years. And the reason I say that is because in the past couple of years we are seeing more non-profits and organizations that are devoted to making AI and probably automated AI. Companies like OpenAI did not exist until a few years ago.”
Although in its early stages, this technology may reduce the amount of work the AI programmer has to do. Eventually it may replace the AI programmer altogether.