2016 Computer conference PostScript-Machine learning: development and future

PostScript reprint please indicate the source

Report title: Machine Learning: Development and the future

Reporter: Zhou Zhihua

Abstract: Over the past 20 years, the human ability to collect, store, transmit, and process data has developed rapidly, and there is a need for computer algorithms that can effectively analyze and utilize data. Machine learning, as the source of intelligent data analysis algorithm, conforms to the urgent demand of the great times, so it has made great development and has received extensive attention.

Machine learning is an important branch of science from artificial intelligence, and it is the key to realize intelligence. Its classic definition is: Use experience to improve the performance of the system itself. Transform your experience into data. With the development of this field, the main research of intelligent data analysis theory and algorithm, and has become one of the sources of intelligent data analysis technology.

Article Filter Story: Invite experts to read a small number of articles, experts mark the article as "relevant" or "irrelevant", based on this information to establish a classification model, and then according to the model to the other articles to predict.

A typical machine learning process: Data is collected first, data is in tabular form, each row represents an object or an instance, and each column depicts an attribute of an object, with a column that we call a category tag.

We train the data to get the model. In the future, when we get a data we have not seen, we know its input, input into this model, this model will give you a result (such as watermelon good or bad). So we can abstract the classification and speculation in real life. It is more important to learn the data to get the model (using the Learning algorithm).

Deep learning

1, enhance the model complexity, improve learning ability

Increase the number of hidden neurons (model width) Increase the number of functions

Increasing the number of hidden layers (model depth) increases the number of functions and increases the number of layers of the function: increasing the number of hidden layers is more effective than increasing the number of hidden neurons, not only increasing the number of neurons that have activation functions, but also increasing the number of layers in which the activation function is nested.

2. Increase the complexity of the model, increasing the risk of overfitting (because the model is too complex), increase the computational overhead

Over-fitting risk resolution can use a lot of training data, and complex models use brute force computing devices to calculate

Deep learning also requires the know-how.

Future machine learning problems: difficult to adapt to environmental changes, difficult to understand models, difficult to obtain sufficient samples, difficult to obtain expert-level results, and difficult to avoid data leakage.

In addition, even with the same data, it is difficult for ordinary users to live machine-learning expert performance.

The humble opinion about the future: The robustness is the key to the open environment Learning task.

The concept of a learning piece (Learn Ware) is presented.

Learning Piece (Learnware) = Model + Statute (specification)

The machine learning application has been done by others and is happy to share the model with you on a platform. Other people can find their own model in this platform. Partially reuse the results of others and use their own data to polish the model. The protocol needs to be able to give a suitable characterization of the model. The model needs to be fulfilled: reusable, evolutionary, and understanding.

Reusable: The pre-training model of a learning piece needs to be updated or enhanced with only "small amounts of data" for new tasks.

Evolution: The pre-training model of a learning piece should have the ability to perceive changes in the environment, and to adapt to the changes actively and adaptively.

It can be understood that the model of the learning piece should be understood to some extent by the user (including its objectives, learning results, resource requirements, performance on typical tasks, etc.), otherwise, it will be difficult to give the model of the functional specification, through the reuse, the evolution of the model obtained by the validity and correctness is also difficult to protect.

Machine Learning Summary:

1, deep learning may have winter, it is only a machine learning technology, the more tidal technology will always appear.

2, machine learning will not have winter: unless we no longer need to analyze data.

3. About the Future:

Technology: Efficient use of computing devices such as GPUs

Task: Open Environment Machine learning task (robustness is key)

Morphology: from "algorithm + data" to "learning" (Learn Ware)

2016 Computer conference PostScript-Machine learning: development and future