Monkey See, Monkey Do (Unless the Monkey is a Computer)
The terms machine learning or AI (artificial intelligence) have become commonplace in today’s tech filled world. Even though we accept these terms we usually do not stop to consider what they really mean in practice. As apps like Google Translate become more powerful through machine learning, the natural question arises, , “How exactly do you make a computer ‘learn’?” You can’t exactly sit your laptop down and show it a series of videos from National Geographic and expect a Wikipedia page about lions to be spontaneously created the following day. despite the common belief, computers don’t just do things. Computers have to be given the ability to perform a function.
In our last Cat’s Meow blog post we talked about the capabilities of machines and provided an example of a robotic arm and how each step in its movement had to be pre-programmed so the machine knew what to do when. The complexity of programming a robotic arm to pick up an object or make a drink may seem like a “technological marvel”, but inputting a command into a computer is easily executed; automation - getting the machine to understand a process without always having human inputs - is harder.
In order to make a completely automated system all possible inputs and outcomes must be considered; hence the complexity behind automated translation.
If you were to program a computer that could translate a single phrase from English to Urdu, the computer could only expect one input and give one output.
As you may have realized by now, language does not work like that. Language is flexible; there are many ways to say the same thing and predicting these variations isn’t easy; that mastery comes with time and exposure to the language.
If you ask me to speak Swahili, I would not be able to, but I can learn to. Currently I have no exposure to Swahili, and it would take time and practice to facilitate that learning.
Unlike the human brain, computers do not have an innate ability to learn, they must be programmed for this task. . Computers must be made to learn.
Despite the differences in “hardware” between humans and computers, they can learn in similar fashions. Not surprising since computers have been built and programmed by creatures (humans) that have “brains”. Programming complex “neural networks” is how computers have been made to learn language.
That’ll Learn Ya
The axiom goes, “Learning is a process”. Learning can be measured by change in behavior over time, but nobody taught humans how to learn. Learning is, for human beings, more or less automatic as long as some effort is applied. We are still, at our for lack of a better term, finite creatures. We only have so much ability to retain what we learn, just as computers do. The main difference, no one had to teach us how to learn or how to be intelligent. In order to create computer systems capable of learning, the brilliant minds behind artificial intelligence had to model these learning processes after human intelligence. If you Google the definition of intelligence, you will be offered the answer, “The ability to acquire and apply knowledge and skills.”
This is exactly what the “intelligence” in artificial intelligence means. The piece in this system that had to be constructed first in theory and then in application was the ability to acquire knowledge and skills. We learn through the formation connections of neurons in our brains associated with particular pieces of knowledge and skills. The stronger the connection, the more proficient in that skill, or cemented that knowledge becomes. Recently the focus of artificial intelligence has shifted towards neural models of machine learning also known as “deep learning”. Essentially, we are giving computers their own connection of “neurons” to form the same sort of links that we do when we learn. Essentially the process looks like this:
An example of the input and the desired outcome is provided to a neural network of a computer.
The computer uses its own established “bias” to try to “figure out” the connection.
This process is repeated as the computer adjusts its “bias” as it is shown more data from the same sample
After a sufficient sample has been taken and trained the A.I., it can be set to work using it’s training as a guide.
Essentially, the computer is made to analyze patterns and then predict an answer or generate a response based on the data it analyzed. Translating, “How are you” from English into Italian first requires the computer’s mathematically based system to find the answer and then produce a result that we can understand based on a match that it learned. Again, this isn’t very different from the process that human beings undergo in order to process anything. There aren’t words physically stored in my brain; it’s all chemicals and bioelectrical signals transferring back and forth until I come up with “Come stai?” as our answer.
The machine works the same way. Once it has been given that relationship between the words like I was “Come stai? = How are you?” it understands that relationship and like humans, builds upon it. If you add the word “today” to the sentence, the computer may very well add, based on other data it has come up with, the word “oggi”. After referencing the grammar rules, the computer can easily produce the answer “Come stai oggi?” just as a human speaker of the language might. This is a simplified example, of course. That flexibility we had mentioned before plays a large role in the difficulty, but just like the people (or machines) learning them, languages change and expand all the time.
This is where it gets a little more complicated, as if it wasn’t complicated enough already. Language branches in countless directions. Creating an AI that can tell the difference between a dog and a cat is easy. Eventually, the program will have enough data where it is consistently able to differentiate the two when offered the data it needs. Language however, changes all the time. In the last couple years, the word “lit” has become slang in English to mean fun or exciting. A wrench like this in the system will radically distort a machine language unless it has already accounted for this outlier. The computer would have to then learn to differentiate between things that are flammable and things that are fun. Though, you may have a conundrum when the machine attempts to translate a news article about an amusement park catching fire.
This is why the translation industry is confident that machines will not be replacing us any time soon. Computers analyze data you ask them to. They exist within a vacuum. Aside from the input they receive, they can’t form new ideas, but the human beings who use language are organic and dynamic. Language is a wonderful reflection of the beings that use it. As we change and adapt, we adapt language to fit us and we will always be the one in charge of that. Language is more fluid than algorithms can predict because each person who uses language, uses language slightly differently. Emphasis, implied meaning, colloquialisms, idioms, and the like are all endless twists that no analyzing of words can predict any time soon. Of course, machine learning makes our jobs easier, but we have only just begun to fight. (AI are also confused by allusion as well).
Want to read more about this topic. Check out the following articles: