Google DeepMind’s algorithm flunks in math test for 16-year-olds
Moving over from training artificial intelligence (AI) in the game of chess, or for assessing MRIs, Google’s AI firm DeepMind decided to train its algorithms on a high school Math test, according to the UK national curriculum.
However, its deep learning neural networks failed to clear a 16-year-old level math problem as they couldn’t even translate the problems, revealed a paper published by Google’s DeepMind AI researchers last week.
Google’s DeepMind decided to put its algorithms to test by training it on UK national school mathematics curriculums (up to age 16), which covered Arithmetic, Calculus, Comparisons, Measurement, Numbers, Manipulating Polynomials, and Probability. It then made the models attempt the same level of questionnaires that a high school student would typically face in an exam.
While one model performed slightly better than others, but on most occasions, they faced problems in translating the questions, full of words, symbols, numbers, and functions, into actual operations needed to obtain results.
Out of the 40 questions directed at different algorithms, the neural networks managed to get only 14 answers correct, which is equivalent of an E grade for a 16-year-old UK student, researchers noted. For instance, one of the incorrectly answered questions was: What is the sum of 1+1+1+1+1+1+1?
While the Math test result may be one of the surprising disappointing performance by DeepMind’s AI, it has, however, done pretty well in the past.
For instance, DeepMind’s AI has not only tackled games like Go and StarCraft like a professional but also helped Google in cutting down the cost of power at its data centers by 15 percent. Besides these, DeepMind algorithms can leave doctors behind in detecting diseases and can also make their own images from random data.
We, humans, apply a variety of cognitive skills to solve a simple math question, such as automatically learning of mathematical operations, memorizing the order in which to perform them, make sense of a problem, with all the numbers, variables, arithmetic operators, and words.
However, machines are not yet fully equipped to perform all these functions and need to be upgraded. They need to be trained to not only understand problems but also be trained on applying the knowledge of rules, transformations, and processes to solve problems.