● ●Magnitudeoferror−αθ●● ●By repeating the update process, the error gradually becomes smaller.Determining the update direction based on the gradientEPILOGUE SCIENCE, SCIENCE ! Rion is supported by many science-loving and math-loving staff mem-bers. In this series, our science-minded staff members write about their enthusiasm for their respective fields of interest. Part 9 is about the gra-dient descent method used in AI machine learning. Are machines similar to humans or completely irrelevant?Article by Ryo SatoRyo SatoAI Research & Development Group, Research & Development Department, Research & Development Center. Joined Rion in 2020. After working on signal processing and firmware development for hearing aids, he is currently engaged in research and development of hearing aid processing using deep learning.20A Column by Rion’s Staff on Their Obsession with ScienceSince joining Rion, I’ve been involved in research on deep learning, a key area of AI technology. Although I didn’t study deep learning during my student days, I found myself fascinated by its appeal in my daily work. Especially intriguing is the process by which AI becomes more intelligent through the gradient descent method. Watching how AI—nominally merely a machine—learns through this method almost feels like observing the growth of a human.With gradient descent, the AI’s current predictions are compared to answers provided by humans, and then the AI is adjusted to minimize the differences. In other words, we guide the AI in a direction that reduces the discrepancies between its predictions and the correct answers. By repeating this process, the AI gradually improves itself and eventually becomes capable of making more accurate predictions.Gradient descent methodA method used in machine learning to optimize pa-rameters. First, obtain the gradient for the function to be minimized (the objective function), and then use this gradient to find parameters that minimize the function. The formula, where α is the learning rate and J is the objective function, indicates how updating parameters decreases the objective function.No. 009The Gradient Descent MethodThe similarities don’t end here. Like us, AI sometimes fails to improve. We’ve all experienced situations where the actions we choose with good intentions make things worse. The same can happen to AI. Even if you guide it in the right direction, if the gap is too large, the AI’s performance may de-cline. Still, by continuing the trial-and-error process, the AI can overcome temporary * *One of the most common learning methods in AI is supervised learning. Here, the AI is provided with large volumes of data along with the corresponding answers prepared by humans, which are used to teach it how to predict correct answers. Ideally, we want the AI to accurately predict the answers to given questions. However, since AI uses a massive number of internal parameters, it’s difficult for humans to directly determine how to adjust them. That’s where gradient descent comes into play.(k+1)(k)What fascinates me about the gradient descent method is how closely the learning process of AI resembles the trial-and-er-ror process we apply in our daily lives. We don’t always know the perfect answer to daily issues, which is why we repeat the process of making small improvements to our current situation to achieve better results. For example, when we’re cooking, we often taste the food and gradually add some seasoning to create better flavors. That’s exactly like repeating the trial-and-error process to make gradual improvements toward something still unknown. In this way, AI improves its abilities through a process similar to our trial-and-error approach we unconsciously use.dips and make progress. * *So, the way AI learns through gradient descent is much like human trial and error. In our daily lives, both at work and at home, we can see many examples of this approach of starting from an imperfect state and making improvements step by step to come closer to a goal. Understanding the AI learning process may give us a good opportunity to think about our own behavior.The humanity I feel in AIθ = ∂ J∂θ Because We’re Science and Math Lovers
元のページ ../index.html#22