In recent years, the popular Neural networks (neural network) and deep Learning (learning) are branches of machine learning (machines learning). In the definition of computer science, learning can be understood as: "A computer program was said to learn from experience E with respect to some class of tasks T and P Erformance measure P, if its performance at the tasks in T, as measured by P, improves with experience E "(Mitchell 1997).
T-task: The goal of Machine learning (Task). From an engineering point of view, the machine learning approach can be used to solve many of the problems that "programs" in conventional sense cannot solve. Common machine learning tasks are:
- Classification
- Classification with missing inputs
- Regression
- Transcription
- Translation
- Structured Output Task
- Anomaly Detection
- Synthesis and sampling
- Imputation of missing value
- Denoising
- Density of probability function estimation
p-performance: We need to measure the quality of machine learning algorithms. For example, in classification's task, we can use the accuracy accuracy rate to measure the performance of the algorithm. Typically, we set up a test set to test network performance. The test set does not intersect with the training set, the validation set (the training set contains the data for training learning, the validation set is used to select the optimal parameters, etc.). Many times we will find that performance is a difficult problem to quantify. In supervised learning (supervised learning), examples like recognition, speech recognition, etc., we can find a cost function (loss function) to measure the gap between the network output y ' and the target output y. But not all learning can be done through supervised learning. For example, in the task of computer music creation, we cannot find the "correct answer" for the next output note under the current input note. Unsupervised learning is required (unsupervised learning). And that leads to our experience, E.
e-experience : The experience of machine learning E can come from the two learning modes mentioned above, supervised learning (SL) and unsupervised learning. Both types of learning require a data-set, which contains sample examples. These samples contain many characteristics, and the task of machine learning in many cases is to learn the characteristics of datasets. Unsupervised learning requires a unique algorithm to learn features, common algorithms include clustering Algorithms (Cluster), and RBM based on energy models, Autoencoder and so on. Supervised learning differs from unsupervised learning in that a sample of supervised learning has a label for each example, and we need to learn the characteristics of example so that the network learns to assign the correct label to example similar to the test set.
Machine learning also includes other algorithms, such as reinforcement learning (intensive learning), which has been widely used for nearly one or two years. The next blog is about a branch of machine learning, the network and algorithms in deep learning.
Reference: Deep Learning -Yoshua Bengio etc.
Machine Learning-basics