Shulou(Shulou.com)06/03 Report--

What is the future research direction of artificial intelligence and machine learning, and how do people use this technology to solve problems in many scientific fields?

At the 2019 World artificial Intelligence Congress (WAIC 2019), which ended not long ago, Eric Grimson, Honorary President of MIT, brought a keynote speech on "stepping into the Beauty of AI Science" in the artificial intelligence topic forum of Shangtang Technology "Zhihuanxin Kinetics", sharing the latest research and views of MIT. Here are some wonderful excerpts to share with you.

What suggestions can modern cognitive science provide for the development of AI?

First, observe the evolution of artificial intelligence over the past 60 years. Since the term artificial intelligence was first proposed in 1956, AI is not only a computing system capable of performing intelligent tasks, but also takes human understanding of human intelligence as its development guide, which was a great innovation at that time. We have seen a lot of trends in artificial intelligence research in the past 60 years, and some of the current trends are to combine modern neuroscience and cognitive science research with artificial intelligence research.

Therefore, we can't help but ask, what suggestions have modern neuroscience and cognitive science put forward for the future AI algorithm?

The current results of deep learning are amazing, such as AlphaGo defeating human professional go players or computers have been able to recognize facial features better than people, but these systems often require hundreds of millions of training samples and a lot of computing power. By contrast, however, a young child can learn or reason with very few examples.

The research of modern cognitive development shows that the learning style of human children is much more complex than previously thought. Young children not only carry out pattern matching, but also explore and form a cognitive structure. For example, Laura Schulz, a professor in the Department of brain and Cognitive Science at the Massachusetts Institute of Technology, has shown that children often create hypotheses related to objects or situations and then test them through observations or experiments. Her research shows that children can form and test hypotheses through experiments, and can be summarized and summarized through final results and causality. This learning theory, especially the ability to learn from a few examples, provides us with different methods of artificial intelligence evolution algorithms.

Based on this idea, let's start with an unusual example of Josh Tenenbaum, a professor in the Department of brain and Cognitive Science at MIT. Suppose there are 50 different characters, containing 1600 simple visual concepts, but only 20 examples are provided for each concept, so this training data is very limited. The experimental results show that people are very good at recognizing and matching characters, and machines can be the same as or surpass humans. Taking a single hand-drawn character as an example, the computer has been able to find the matching character correctly.

Professor Josh Tenenbaum took advantage of the logic of character generation and expanded it. Suppose the user builds a physical simulation engine that can generate a sequence of images for any operation, not just modeling the stroke generation of characters, including allowing random variables to be created. The user can then predict the outcome of the operation like a child building blocks. Under this logic, he created a human intelligence model to simulate the behavior of young children through a series of engines and probability programs. By using perceptual observations and constantly improving the understanding of the system, the whole system can learn over time, start learning like a newborn, and gradually grow into a two-or three-year-old child. This approach may introduce a very different way of learning-a way that requires few examples but is achieved through simulated training.

Can mathematics improve the current machine learning methods?

In addition to using neuroscience and cognitive science to better understand artificial intelligence and machine learning systems and to build new methods, the mathematical basis behind machine learning is also important. How do you know whether to trust the answers provided by artificial intelligence systems? How powerful is the system? Is there an implicit bias in the system? Here are examples of current research projects at MIT, including these questions.

If the training data can well represent the test data, then the deep learning system will be very reliable. However, if you attack the system with a selected jammer, AI may be fooled. Is there a mathematically reasonable way to create a more powerful system?

Recently, many studies have shown that deep neural networks may be very fragile, and adding small interference to a sample may lead to obvious classification errors. To solve this problem, Aleksander Madry, an associate professor in the Department of Electronic Engineering and computer Science at the Massachusetts Institute of Technology, created a more robust classifier. He proposed to treat the problem as a game theory problem.

The key of this method is to train the model not only on the real data, but also on the disturbed samples. Ideally, we would train for all possible disturbances, but this is impossible because there are too many. Through experiments, Madry observed that it is enough to select a limited number of adjacent points in the feature space, because the value of the loss function does not change much in a small region. By selectively training for disturbance, a more robust neural network can be constructed. Experiments on standard data sets show that his method does significantly improve the robustness of the classification system.

How do artificial intelligence systems help traditional science?

In addition to considering the science behind machine learning, we also need to think about how artificial intelligence systems can help solve some problems in traditional science.

Materials scientists and computer scientists have jointly created an artificial intelligence system that can read scientific papers and extract "recipes" to produce specific types of materials. Their system theoretically provides a formula for the creation of a material, such as the physical properties of the material and the method of designing and manufacturing the material, while the system can identify higher-level patterns, for example, can identify the correlation between the precursor chemicals used in the material formulation and the crystal structure of the resulting product. In addition, the system also provides a natural mechanism for generating original formulations, through which researchers can suggest alternative formulations for known materials. The application of this research work is particularly important in renewable energy.

The second example is the study of converting a group of reactants into target compounds. Synthetic chemistry is an art usually performed by experienced chemists.

MIT research scientist Caleb Harper and colleagues combine environmental and climate control, metabolic phenotypic measurement and machine learning to better understand the biological processes of food growth. By using machine learning to study the correlation between light conditions and the molecular concentration of volatile flavor compounds, we can have a better understanding of the chemical process of food production. The work aims to develop plants with higher levels of compounds that can be used to help fight diseases such as diabetes.

The above series of examples show how MIT researchers study the science that underpins current and future artificial intelligence systems, and how these systems help scientists answer basic questions about the physical world.

AI is showing more and more charm!

Https://www.toutiao.com/a6737136664151327243/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.