Tommitchel, the most globally recognized father of

  • Detail

Tom Mitchell, the globally recognized father of machine learning: how artificial intelligence and intelligent adaptive learning are interconnected

original title: Tom Mitchell, the globally recognized father of machine learning: how artificial intelligence and intelligent adaptive learning are interconnected I dark horse & fire

original title: the globally recognized need for machine learning to design new fixtures father Tom Mitchell: how artificial intelligence and intelligent adaptive learning are interconnected

i dark horse & Matchbox news November 16 On the th, at the "global ai+ smart Adaptation Education Summit" held by Lei Feng and Yi Xue education squirrel AI, Tom Mitchell, the globally recognized father of machine learning, delivered a speech on the performance of wood-based panel experimental machines

the following is the content of his speech, which has been sorted out by I dark horse & Matchbox:

Tom Mitchell: Thank you very much. First of all, I want to thank the organizers of this meeting for giving us the opportunity to come here today. I want to tell you one thing. I think intelligent adaptive learning and related teaching have been developing, but I think the last 10 years are the best 10 years. Because in the past 10 years, we have begun to see these technologies become more mature, and companies have begun to use these technologies, so I think this meeting, especially this year, is a turning point in our field. Today's audience will invent a new generation of intelligent adaptive learning technologies

today, I want to share with you my thoughts on artificial intelligence and how it relates to intelligent adaptation learning. These three ppts are intercepted from youtube, telling us what intelligent adaptation learning is. There is a system of intellectual adaptation, in which teachers will constantly evaluate students' abilities, find students' shortcomings, set goals for them, find students' learning needs, and give them some better suggestions according to their learning speed

we found that some systems have begun to do this, but they are in the early stage. We will do more work to improve this system next. Today I want to talk about machine learning and artificial intelligence, which will be a driving technology in intelligent adaptive learning

let's take a closer look. At present, machine learning has developed very fast in the past 10 years, so AI didn't do very well in computer vision in the past 10 years, but later its recognition accuracy was very high, almost the same as human eyes. So we can see on the screen that the accuracy of computer vision has been very high since 2010. The horizontal axis here is human, next to machine recognition. There is a comparison. The accuracy is very close, and the speed of machine learning is very fast. The development speed of voice learning is also very fast. In go, AI has defeated the world champion. In robot cars, we will also have many combinations of driverless cars and robots

so looking forward to the future, we can say how such machine learning will promote the development of AI and intelligent adaptive learning? I think there are three issues worth discussing:

1 What should machines learn

today or yesterday, we have done a lot of useful discussions and discussions. We should have personalized teaching strategies based on each student. For example, students' drop out rate, students' exam results and some suggestions, when teachers should intervene, and how to intervene between students and machines

2. What data should we learn from

for a long time, we have seen that many systems calculate the data of the network, whether the students answer the question correctly, as well as the time spent on this question and the number of mouse clicks. We found different types of data, including environmental data, such as camera heads, earphones, microphones, etc. with these environmental data, we can better understand the situation of learners, such as whether there is noise in the classroom and whether the temperature is appropriate. We also saw many sensors on the body, such as smart watches and smart chairs, which can sense the learning state of learners. In the future, we will also have the invention of intelligent chair, which may monitor whether students sway around and whether they are distracted. We can also use EEG to monitor the state of the brain, so the source of data we find in machine learning is also very important. And algorithms are also important for machine learning

let's take a look at the algorithm. This is a supervised learning system, which was proposed by a doctoral student at Carnegie Mellon University. There is such a network that can predict which students will drop out or withdraw from classes. Some of the data it monitors include the time spent by students in the course and the number of mouse clicks, so supervised learning will be monitored from the input and output ends. In the middle is a network structure. Supervised learning is also very important for intelligent adaptation education. We need a series of technologies to improve the accuracy and efficiency of supervised learning, taking into account the learning situation at that time

for example, if we want to train the network more intelligently and monitor it better, which course the student will withdraw from, or whether he will withdraw from other courses, this is also the point we need to pay attention to in cross task transfer

3. Multitasking learning

this learning theory was put forward in the late 1990s. We want to have a good training system. Sometimes we want to improve the accuracy. We will predict some other variables. For example, we need to monitor which patients with pneumonia will become very serious at the end of the disease. It is not only the severity of pneumonia, but also some other parameters or variables, such as the number of white blood cells or whether they have been treated in ICU. We can let the system predict many different variables, the most important of which is to look at the severity of pneumonia

in this regard, we can also consider using this model to predict the dropout rate of students. In one way, we don't let the system only look at his dropout rate, but at other variables. For example, the result of his final exam. The reason why this technology can succeed is that it can train our system to predict multiple variables, so that learners can better learn the nervous system, and in this process, there can be some inductive procedures, so that this system can better predict the variables we are really interested in

in the algorithm of machine learning, another very interesting method is unsupervised learning, which is also the research done by my colleagues at Carnegie Mellon University. They looked at the clustering of different students, mainly to see their correct rate of answering questions. They gathered all students together and classified different performances. You can see that there are three very different curves, red, green and blue. For the green group, they did very well at first, but then suddenly fell, and then they didn't do so well. For this kind of unsupervised learning, we can intervene after finding a rule. For example, we can predict the final exam results of these three groups, so as to better guide other students

reinforcement learning must be very important because there are special learning algorithms. That is, at every point in time, we will give it different actions in the whole process. We can learn from alphago and make it continuously enhance its ability to defeat the go champion. In terms of reinforcement learning, it is also based on the decision-making process. In a simple way, for example, in go, every point on go and every sub is a decision. We will have a reward or incentive mechanism, that is, we will tell the system what they should do and how to achieve successful results through such a reward mechanism. We have this kind of reward mechanism. If the opponent gives up, you will add 100 points, If you give up, you will lose 100 points. We keep training this system. This process may produce millions of sequences of actions. Different sequences have been trained by playing chess with yourself in alpha

let's take a look at this way. In fact, it is very similar to our concept of intelligent adaptive learning. For example, in intelligent adaptive learning, we also have different states and different actions. This state refers to students' learning state and psychological state. The current behavior refers to teaching behavior. Should we give him an exam, video or evaluation? We can also define some reward mechanisms to tell us what actions we want to choose to better promote students' learning. If it is a positive score, it means we are doing well, and a negative score means we are not doing well. What actions should be taken in each learning state? In such repeated training, we can have the whole sequence of actions and finally get the best reward

an interesting point about reinforcement learning is that there has been a lot of progress in this area, and there has been great progress in the past five years. As you can see, we also have many variables in go. In this process, we have many changing problem models, which are also a very personalized model. From the perspective of machine learning researchers, we also have some variables and sampling methods, including algorithms and data. Now we are in the best era, for consulting education

let's take a look at the second part and look at some more specific research. First, let's take a look at the new types of collecting students' questions. Every time we think of machine learning, we will have corresponding algorithms, for example, 40%. The other 60% of the importance comes from data. If we want to do well, data is very important. I want to show you a study done by my colleague. This person is interacting with computers. "How are you?" and "I'm fine". You can see that his expression and face are constantly being tracked, "what do you want to see?" "Look at baseball", it has different monitoring. What this system does is to track the person's facial expression and his voice for 1 second. In this way, we can infer the person's emotional state. You can see that the red dot is constantly moving, that is, to see whether the person's emotional state is positive or negative, as well as his energy and the whole emotional state. Such projects are still ongoing

it was done by a teacher from Carnegie Mellon University and cooperated with another teacher. They want to apply this technology to analyze the video. This video is also an interaction between teachers and students. This video is about some students who want to solve math problems. We see that the horizontal axis is time. In this process, we can see 3 The accuracy of the force sensor may have different emotional changes every second. Students' emotions will change in a way. These small boxes represent who is speaking, or whether they have seen the teacher's action

we can see that the first part of the students are reading for college students, because you can see the display of students talking on it. We can see that at the end of this stage, the system infers that this student should have some depression and negativity

Copyright © 2011 JIN SHI