Editor's note: Mehmet Alican Noyan, a machine learning engineer at ICFO, suggested that the way we educate and learn can draw on some experience in machine learning in the field of AI. The comparison of artificial intelligence (AI) and human intelligence has been a source of intense debate. Is it possible for machines to think like humans? How far is it from intelligent machines to rule the world? Do artificial neural networks borrow from the brain? Such questions focus on shaping the future of AI. But why don't we imagine how AI can be used to improve human intelligence? I know your skepticism, don't worry, this is not an article about editing genes to upgrade the brain. Human intelligence is not just about the brain, education is also a fundamental part of our intelligence. We can improve human intelligence through better education. But it looks like we are far more successful at training machines than we are at training humans. There are several possible explanations for this phenomenon. AI is a mathematical construct, and in most cases we can come up with a better-defined performance measure, and education contains economic, social, political, and religious components, and a better definition becomes subjective. Also, in AI, we have more freedom to experiment to find out which learning method works best. On the other hand, experiments in education have many constraints (economic, time, etc.). Finally, we have benchmark datasets to help people around the world compare their machine learning methods. In education, it is difficult to make such a general comparison. These challenges do not mean that we are cornered. Asimov said, “The saddest thing right now is that science accumulates knowledge faster than society accumulates wisdom.†To make a difference, let’s use knowledge from AI to improve human intelligence. Rule-based and self-learning There are two approaches to AI: a rule-based system, which hard-codes the rules that the algorithm follows; and self-learning (ie, machine learning), where data is provided to the algorithm, and the algorithm learns patterns, relationships, and transformations on its own. On vision tasks, machine learning is widely recognized to outperform rule-based algorithms. In other words, we prefer to show data rather than tell machines what to do. But we do it differently when it comes to human education. Instead of showing data to students to learn the truth themselves, we inform and instill so-called truth. This prevents students from internalizing concepts. For some limited problems, such as repetitive tasks, this may be sufficient. However, dealing with new problems requires a habit of adapting and mixing different ideas. This can only happen if you understand the essence of the concept and not just memorize it. Students are not required to know, but only to memorize what the teacher taught. —Paulo Freire As in machine learning, we should rely on self-learning, aka self-education. Schools should create an environment that motivates self-education. I firmly believe that self-education is the only form of education. The only function of a school is to make self-education easier; if it doesn't, it does nothing. —Isaac Asimov How to motivate self-learning? Even if we agree that self-learning is the way forward, how do we execute it? AI researchers put a lot of effort into machine learning, and we have a wide range of knowledge to draw from. In many machine learning tasks, we use an optimization algorithm called gradient descent. This is how machines actually learn. Understanding the basics of this algorithm is easy. It is an iterative algorithm that gradually approaches the answer. It starts by making a prediction, then gets feedback on how far it is from the truth, and then makes a slightly improved prediction. This sequence continues until we are satisfied with the gap between prediction and truth. In other words, learning is an active, step-by-step process in which the algorithm reconsiders its assumptions at each step and gradually improves. I cannot think for others, and I cannot think without the help of others, and others cannot think for me. Even if people's ideas are superstitious, or naive, they can only make changes when they rethink their assumptions in action. In the process, one must produce one's own ideas and act on them, rather than consume the ideas of others. —Paulo Freire (Isn't this a lot like Gradient Descent? Or did I get carried away by reading too much about AI?) As you can see, gradient descent might help us understand how to perform self-learning. We can also get some experience from the testing phase. A point that everyone working in ML keeps in mind is to train an algorithm with one dataset (called training data) and then test it with another dataset (called test data) to ensure that the algorithm is not memorizing (overfitting). ), indeed learning. Of course, the training data and test data must come from the same distribution. You can't teach math and expect algorithms to answer historical questions well. For example, if we create a cat classifier, we train the algorithm by showing pictures of Garfield, Hello Kitty, Tigger, etc., and then use different kinds of cats: Felix the cat, Cosmo cat, Figaro... If the algorithm can say Fei li A cat is a cat, so it learns what "cat sex" is. If an algorithm says Garfield is a cat, then it may have learned "feline nature," but it may have just memorized the fact that Garfield is a cat. Therefore, every practitioner in this field agrees that we should not use training data for testing. Do you think this principle applies to human learning as well? When teaching children, we often train and test with a specific set of questions. However, problems in life do not have a predefined strict structure. They keep evolving. We can only deal with life's problems by internalizing concepts rather than rote memorization. So we should challenge students with open-ended questions, confront them with uncertainty, and allow them to guess and explore on their own in the field. Example: How to teach derivatives? Let's take a concrete example and compare the difference between a rule-based approach and a self-learning approach in teaching derivatives. You can skip this part if you want. The goal here is to show how to motivate yourself to learn derivatives, not to teach you what derivatives are. Derivatives are traditionally taught by introducing derivative formulas and showing derivatives of several common functions. Next, students memorize formulas by solving some problems. This is the same as rule-based AI, hard-coding the rules that the algorithm needs to follow. Let's look at another method, the self-learning method. As we do in machine learning, the goal here is to create an environment that motivates self-learning. We don't instill anything, students learn on their own. The essence of a derivative is an instantaneous change, but the change occurs over the entire time period, and an instant is just one of those moments. In order to capture the idea of ​​derivatives, one should perceive the contradictions for themselves. How can we do this? We can discuss one of Zeno's paradoxes: "Suppose you wish to reach a wall that is 1 meter away from you. To reach your goal, you first have to travel half the way to the midpoint (1/2 meter). The remaining The same is true for the distance. To pass the remaining 1/2 meter, you first need to reach the midpoint of the 1/2 (1/4 meter). And so on, there is always a very small distance between you and the wall . You can get close to the wall, but never get there—or maybe you can get there after an infinite number of times. But in real life, we know we can get to the wall. Let's discuss what's going on here..." Hopefully, discussions like this will lead students to grasp the concepts of infinitesimals and infinity, or at least give them a sense of it. Otherwise we iterate this step until they grasp the concept or at least some perceptual knowledge, just like we do in gradient descent. Armed with an understanding of infinitesimals and infinity, we can ask students to discuss instantaneous velocity: "The average velocity is the displacement over a specific period of time. But how do we measure instantaneous velocity? Instantaneous means the period of time is zero, and if time stops, we cannot move. It looks like the instantaneous velocity should be 0/0 = undefined. What do you think?" Again, after a few iterations (5 epochs should be enough :) ), they can probably conclude that the time period approaches zero and the average velocity approaches the instantaneous velocity, just as we approach the wall in Zeno's paradox. That's actually the derivative (the derivative of the displacement-time function is the velocity), and when students try to figure out an answer to some question, they pick it up on their own. We can even get the above formula based on this understanding. I do not recommend omitting formulas in teaching, but students should understand the motivation behind formulas. I'm not an expert in teaching derivatives, so here I try to give a rough outline of learning derivatives on your own. The point is the method, not the derivative. You can apply this concept to any subject. For example, when teaching photography, instead of providing rules for good photos, guide students to good websites, books, and let them form their own understanding of photography in the process of looking at various photos. At the same time, you can organize exhibitions for them, have them showcase their own photos, and iteratively improve their photography techniques in constructive discussions. If you are interested in learning derivatives, check out this video by Grant Sanderson. It is a prime example of proper teaching. At certain points, he would ask questions and say "stop and reflect." This is the key, it takes pauses and reflections to really learn something. He also said, "If you're feeling weird and contradictory, fine! You're grappling with the same conflict that the father of calculus faced..." In a sense, he's helping you become a Newton, which is self-learning. in conclusion It can be said that artificial intelligence is the simulation of human intelligence. It would be a huge waste if we didn't take advantage of its results: We should promote self-learning rather than rule-based learning. We know this is the way forward for machine learning in AI. Students should improve their assumptions on their own. As we did in gradient descent, we may need to supervise the learning process, giving the student feedback at each step, but not the answer. To make sure students are learning and not rote, we should expose them to situations they haven't seen before. In machine learning, we test algorithms on unseen datasets. You might be thinking that these insights about education have long been known, and the correspondence between human intelligence and artificial intelligence is not that useful at all. But we ask you to think twice. The key here is that the AI ​​field is almost unanimously agreeing on these insights. Do you think these are also true in the field of human intelligence? I do realize that humans and machines are not the same and it is impossible to say that there is a 100% correspondence. However, it is clear that there is a strong relationship between the two. Let us use this valuable similarity between humans and machines to understand and overcome challenges in the way we educate our own children. Portable Wireless Speaker,Mini Wireless Speaker,Custom Bluetooth Wireless Speaker,Waterproof bluetooth speaker,Wholesale bluetooth speakers Shenzhen Konchang Electronic Technology Co.,Ltd , https://www.konchang.com