Alan Turing, famous mathematician and computer scientist, is well known for many accomplishments and ideas; among them is the Turing Test. The test defines the point at which one may consider a machine intelligent as when the machine is able to imitate a human in conversation. Does Clever Bot, an online chat bot, constitute an intelligent machine? Most people would likely answer no.
While Clever Bot may not be the best artificial intelligence that exists today, humans still seem to be very far from a self aware computer. Nevertheless, artificial intelligence plays an increasing role in our lives today. Namely, machine learning is becoming much more popular and prevalent.
Machine learning is the computational practice of allowing a computer to discover how to accomplish a task without being explicitly programmed to do so in a predetermined matter. At its most basic implementation, machine learning works by beginning with a random model of a situation and then finely tuning this crude approximation of data in accordance with the error it encounters with training data. For example, a machine learning algorithm with the goal of predicting whether a house is located in a rural or urban area based on square footage and cost may begin by plotting each house on a graph of square footage and cost. Then, it will randomly generate a straight line. The algorithm aims to make this line roughly divide the two types of houses—rural on one side, urban on the other. So, the algorithm will test how far off it was from dividing the two types of houses and use this information to adjust its slope and y-intercept. While this will be imperfect due to outliers and overlap among the rural and urban houses, the algorithm will, with much training data, be able to predict whether a house is one or the other based purely on square footage and cost without human knowledge or programming. Of course, this is a simple case, but it shows what a powerful tool machine learning can be.
Machine learning is frequently applied to ASCII text generation. This has created fascinating results when trained on different data. For example, when trained with a digital anthology of Shakespeare’s plays, this type of algorithm has artificially created Shakespeare, which reads eerily well. The concept governing this development is using the machine learning algorithm to predict the next character of text based off of the previous ones, a more advanced implementation than the houses example. This same concept can be utilized with MIDI data of classical music to train a machine learning algorithm to write its own music which is admittedly still lacking when compared to the works of Mozart. Another interesting example of machine learning is a generalized algorithm for automating the play of Nintendo Entertainment System games.
Additionally, machine learning algorithms are extremely proficient at many games, one of which has been of particular interest to computer scientists. Go is an ancient Chinese board game. The game is simple: each turn a player places one of their pieces—one player has the white pieces, the other has the black pieces—onto a 19 by 19 board, then players can capture the other’s pieces by surrounding them. The game ends when both players pass their turn and the winner is deemed by who encloses the most area of the board with their stones by the end of their game. While the game may sound simpler than others, it proved much more difficult for computers to play.
IBM’s DeepBlue supercomputer was able to best chess champion Gary Kasparov in 1997 during a rematch following its 1996 loss. Meanwhile, Go appeared out of reach using brute force computation and humans remained ahead of computers due to a ridiculously large number of legal board positions, two times ten raised to the power of 170. With the advent of machine learning technology, however, the AlphaGo computer was able to beat the legendary Go player Lee Sedol in March 2016. This marked a milestone in technological advancement because AlphaGo did not rely on examining several moves into the future or preprogrammed instructions, instead AlphaGo used machine learning to build an intuition for the game.
These examples may seem of little real-life importance unless you are heavily invested in artificial music or a professional Go player, but machine learning can become much more exciting when one considers the broader uses and implications thereof, like clinical diagnosis or medical research. Outside of research and healthcare, machine learning is utilized in industry. The most prevalent example is that of the infamous YouTube algorithm. Nearly everyone has interacted with it at some point; this algorithm appears as a mysterious black box producing video suggestions and shaping the revenue streams of online video creators. At its core, YouTube’s algorithm utilizes machine learning to predict which videos will keep people on the platform, watching videos and generating revenue for as long as possible. It does not promote what anyone tells it to and instead finds its own trends within the petabytes upon petabytes of training data fed to it. Google implements similar algorithms in their other products to varying effects such as boosting the accuracy of Google Translate and categorizing all of your photos by the people included.
Some people believe that Google has gone too far in their use of customer’s data for machine learning, others believe it creates vibrant innovations and progress available for free. Meanwhile, another application of machine learning is increasingly identified as sinister: Deepfake technology, its name being an amalgamation of “deep learning”, another term for machine learning, and “fake”. This technology operates by taking video or audio of a person speaking as training data and creating a model for simulating their likeness and voice. This has some perfectly innocent and fun uses, like creating movies without actors present—either to revive the dead or just to save money—or putting your friends in their favorite films.
This technology, however, also poses a substantial threat. One of the most upsetting consequences is revenge porn with 96% of deepfake videos circulating the internet being pornographic in nature. This poses an obvious threat to the privacy, autonomy, and safety of those targeted, most often women. Additionally, there are fears that deepfake technology could be used to impersonate others and wreck political or financial havoc; which has already been demonstrated. This year, a voice deepfake was used to make a CEO believe his boss from his parent company was requesting a bank transfer over the phone and it was accurate enough that the CEO believed the scammer and transferred $243,000 to the scammer’s account.
Machine learning is a radical, relatively new technology and we are dealing with all the fun, innovation, trials and tribulations that come with that. We must work towards a future in which machine learning matures, both as a technology and as a part of life. This means that scientists must work to hone this tool for good in research and saving lives and legislators must work with the private entities to create safeguards and best practices against its abuse. One thing is for sure, machine learning is here to stay. We all know how hard it is to get rid of intelligent machines.
How could you apply machine learning? Let us know in the comments below.
