Facebook Yann LeCun One-Hour Speech Highlights: Unsupervised learning represents the future of AI technology?

The rapid advancement of AI technology is largely due to breakthroughs in deep learning and neural networking, as well as the creation of large databases and faster GPUs. We now have an AI system with image recognition capabilities comparable to humans (such as Facebook's recognition system below). This will lead to revolutions in many areas, including automated transportation and medical image analysis. But these systems now use supervised learning, and the input data is artificially tagged.

The next challenge is how to get the machine to learn from unprocessed, unlabeled, and unclassified data, such as video and text. And this is unsupervised learning.

The traditional idea is that if you don't have a lot of data, the neural network should be controlled on a smaller scale. Yann LeCun pointed out that this is completely wrong. His team expanded the neural network with the same data and got better results. He said that the larger the neural network, the better the effect (of course, the premise is that the database size reaches a critical value). As for why this is the case, it is still a mystery, and relevant theoretical research is being carried out.

Yann LeCun highlighted the importance and application of convolutional neural networks: "We have long recognized that convolutional neural networks can be used to handle multiple tasks - not just identifying individual objects (such as alphanumeric), but also Multiple objects can be identified, and object recognition, grouping, and interpretation can be performed at the same time. For example, the convolutional neural network can be used to train the AI ​​system to identify and label each pixel in the image (camera captured) to analyze whether the front path can pass. In Nvidia's recent autopilot program, they used a convolutional neural network to train the autopilot system. The system analyzes the image provided by the camera and mimics the human steering angle."

He also introduced the application of convolutional neural networks in the Facebook image recognition system. “With it, Facebook's system not only recognizes images, but also outlines the images and classifies them according to their contours. The system can even pick out broccoli from Chinese dishes (see below).

The following is a comparison of the same image before and after recognition:

Yann LeCun says this is a huge improvement, if you asked an AI expert a few years ago: "When can we do this?", the answer will be "unclear."

“If we want AI technology to continue to improve, we must let the machine analyze, reason, remember, and turn phenomena and words into operational knowledge.”

He then made predictions that the next technology that will be very popular is the memory-enhanced neural network. It can be understood as a memory-enhanced recurrent neural network in which memory itself is a circuit that can be distinguished and can be used as part of learning for training. Yann LeCun went on to discuss the technology in depth, so I won't go into details here. See the video for details.

The required data sizes for intensive learning, supervised learning, and unsupervised learning vary by orders of magnitude. The information needed to reinforce learning for each trial may be only a few bits, supervised learning is ten to ten thousand bits of information, and unsupervised learning requires millions of bits. Therefore, Yann LeCun made a metaphor: suppose that machine learning is a cake, intensive learning is a cherries on a cake, supervised learning is a layer of icing outside, and unsupervised learning is a cake. The importance of unsupervised learning is self-evident. In order to make intensive learning work, it is also inseparable from the support of unsupervised learning.

The main problem at the moment is that the AI ​​system does not have "common sense." Humans and animals acquire common sense by observing the world, acting, and understanding the laws of nature, and machines need to learn to do so. Many experts, including Yann LeCun, use unsupervised learning as the key to giving machine common sense. The process is as follows:

The AI ​​system consists of two parts: the agent and the target (agent and objecTIve). The agent acts to observe the action's perception of the impact of reality and then use that perception to predict the reality. The motivation for the agency to carry out this series of activities comes from achieving the goal, and the ultimate goal is to achieve this goal with the highest efficiency. In reinforcement learning, the reward for agency behavior comes from the outside, and the reward for unsupervised learning comes from within (satisfaction with approaching the goal).

But there is a big problem with this process: the way agents perform unsupervised learning is to make various attempts in real life, which is dangerous and inefficient. For example, a driverless car cannot try all possible driving methods, which can pose a safety hazard. This type of attempt is limited by time and cannot be run thousands of times per second like a computer program. So, Yann LeCun explains that in order to improve the efficiency of machine learning, we need model based reinforcement learning. It consists of three parts: the world simulator, the actor and the feedback device (criTIc). The reality simulator simulates the reality, the actor generates acTIon proposals, and the feedback device predicts the effect of the action. In this way, the AI ​​system can repeatedly deduct and optimize the action without being limited by the time and cost in reality.

Molded Plastic Products

We design, engineer and fabricate mold tooling, both standard and custom. We continuously design custom tooling to satisfy our customer needs. These tools are built for machines such as Newbury, Autojector, Ameriplas, Multiplas, etc. Our Solidworks 3D design capabilities represent the leading edge in the industry.

We provide training and know-how to our customers. We offer this unique advantage to companies interested in On-site training to assist in the development of in-house capabilities. Our company can become your "over -mold engineering department" and can provide quick turn-around, high quality for customers' the complete cable set with wire harness, plastic, silizone o-ring, metal terminal, or plate, etc. Try to give you the whole supporting service.

Molded Plastic Products,Plastic Box For Cables,Waterproofing Plastic Box,Plastic Connectors,Plastic Cap,Plastic Bushing

ETOP WIREHARNESS LIMITED , http://www.oemwireharness.com