Machine learning or artificial intelligence

Artificial intelligence is not a new term. It has been around for decades. From the early 1980s, computer scientists began to design algorithms that can learn and imitate human behavior.

In terms of algorithms, the most important algorithm is the neural network, which is not very successful due to overfitting (the model is too powerful, but the data is insufficient). Still, in some more specific tasks, the idea of using data to adapt to functionality has achieved significant success, and this also forms the basis of today's machine learning.

In terms of imitation, artificial intelligence focuses on image recognition, speech recognition and natural language processing. Artificial intelligence experts spend a lot of time creating things like edge detection, color profiles, N-grams, syntax trees, and more. However, these advances are not enough to meet our needs.

Traditional machine learning:

Machine learning (ML) technology plays an important role in forecasting. ML has undergone multiple generations of development, forming a rich model structure, such as:

1. Linear regression.

2. Logistic regression.

3. Decision tree.

4. Support vector machine.

5. Bayesian model.

6. Regularization model.

7. Model integration (ensemble).

8. Neural network.

Each of these predictive models is based on a specific algorithm structure, and the parameters are adjustable. Training the prediction model involves the following steps:

1. Select a model structure (eg logistic regression, random forest, etc.).

2. Enter the model with training data (input and output).

3. The learning algorithm will output the optimal model (ie, a model with specific parameters that minimize training errors).

Each mode has its own characteristics, performs well in some tasks, but does not perform well in other areas. But in general, we can divide them into low-power (simple) models and high-power (complex) models. Choosing a different model is a very tricky issue.

Using a low power/simple model is better than using a high power/complex model for the following reasons:

Training high-power models will take a long time before we have powerful processing power.

Training high-power models can lead to over-fitting problems before we have a lot of data (because high-power models have rich parameters and can adapt to a wide range of data shapes, we may end up training a training data that is appropriate for the specific to the current, and It is not extended enough to make predictions about future data).

However, choosing a low-power model encounters the so-called "under-fitting" problem. The model structure is too simple, and if it is complex, it cannot adapt to the training data. (Imagine that the underlying data has a quadratic relationship: y = 5 * x ^ 2; you can't adapt to linear regression: y = a * x + b, no matter what a and b we choose.

In order to alleviate "inappropriate problems," data scientists often use their "domain knowledge" to propose "input features," which are more directly related to output. (For example, return a quadratic relation y = 5 * square(x), if you create a feature z = x ^ 2, you can fit a linear regression: y = a * z + b, by choosing a = 5 and b = 0).

The main obstacle to machine learning is the step of feature engineering, which requires domain experts to find very important features before entering the training process. Feature engineering steps are done manually and require a lot of domain expertise, so it is the main bottleneck for most machine learning tasks today.

In other words, if we don't have enough processing power and enough data, then we must use a low-power/simple model, which requires us to spend a lot of time and effort to create the appropriate input features. This is where most data scientists spend their time today.

The return of neural networks:

In the era of big data, the common development of cloud computing and massively parallel processing infrastructure has greatly improved the processing power of machines in the early 21st century. We are no longer limited to low power/simple models. For example, the two most popular machine learning models today are random forests and gradient lift trees. Nonetheless, both are very powerful and provide training data for nonlinear model fitting, but data scientists still need to carefully create features to achieve good performance.

At the same time, computer scientists are reusing many layers of neural networks to accomplish these human imitative tasks. This brings new life to DNN (Deep Neural Network) and provides a major breakthrough in image classification and speech recognition tasks. The main difference between DNNs is that you can input raw signals (such as RGB pixel values) directly into the DNN without having to create any domain-specific input functions. Through multi-layered neurons (which is why it is called a "deep" neural network), DNN can "automatically" generate the appropriate features through each layer, and finally provide a very good prediction. This greatly eliminates the hassle of looking for "feature engineering," which data scientists like to see.

DNN has also evolved into many different network topologies, so there are CNN (convolutional neural network), RNN (recurrent neural network), LSTM (long-term short-term memory), GAN (generation of hostile networks), transfer learning, attention model (attention model) All of these are collectively referred to as Deep Learning, which is drawing the attention of the entire machine learning community.

Reinforcement learning:

Another key component is how to imitate a person (or animal) learning, envisioning a very natural animal behavior of perception/behavior/reward cycle. A person or an animal first understands the environment by sensing his or her state. On this basis, he or she will choose an "action" to take him or her to another "state." Then he or she will get a "reward" and repeat it until he or she disappears. This type of learning (called intensive learning) is very different from the traditional method of curve fitting for supervised machine learning. In particular, reinforcement learning is very fast, because every new feedback (such as performing an action and getting a reward) is sent immediately to influence subsequent decisions.

Reinforcement learning also provides a smooth integration of predictions and optimizations because it maintains the current state of belief and possible transition probabilities when taking different actions, and then decides which actions can lead to the best outcome.

Deep learning + reinforcement learning = AI

Compared with the classic ML technology, DL provides a more powerful prediction model, which usually produces good prediction results. Compared with the classic optimization model, reinforcement learning provides a faster learning mechanism and is more adaptable to changes in the environment.