Machine Learning Lecture 1

 

Introduction with Boston Dynamics video. Spot and the humanoid. A Course in Machine Learning or Reinforcement Learning Books can also be extremely helpful. Some good textbooks are a course in machine learning, pattern recognition adn machine learning, a probabilistic perspective, and reinforcement learning. Written homework must be LateX. 

Right now, machine learning is a boom, and is the major topic for the next decade, lots of money and projects that are approaching, very good area to be in at this time. Another usage is machine translation, and another popular application is autonomous cars. Machines can learn to control pedals and the goal is to win a game. This is one machine learning application. You can transfer the movement of a horse to a zebra, and stuff. You can also detect joints and transfer it to another person. It's pretty freaky. This is applied in movie production.

These algorithms drastically reduce the cost of production.In machine learning, we have an example from a real world domain, lie images and understanding about these images. We feed the data to a learning algorithm and the learning algorithm will give a model of the model space to find. Now that we have the model, then we can ask the model to do some prediction, like laabel something based on prediction.

So it's like this:

Data -> Learning Algorithm -> Model. The putup table is part of machine learning, since machine learning is created based on the data we insert as an input. Machine learning is running a model against the given data. A lookup data is like a dictionary, there's just dictionary with point towards classification. We are trying to map the right characteristics to the characters. It might be yes because the computer tried to learn from an input and output and try to get some features to understand if it's a dog or a cat, and it will reuse that model at the end for unknown pictures. 

In learning, we look for a generalization. If we cannot generalize to unseen data, we cannot consider this as a machine learning approach. Some models can fit the training data perfectly, but not the other one. We call this overfitting. 

Representation is how to represent knowledge, and this is through decision trees, instances, linear models, graphical models, neural networks, support vector machines, and model ensembles. Evaluation is the way to evaluate candidate models through accuracy, prediction, squared error and likelihood. Optimization is how to search through the model space and this can be by combinatorial, convex, or constrained optimization given the inductive bias on the models. 

Machine Learning Interview, NN and deeplearning are powerful methods, but they come up with cavaets like they can overfit a lot. In an interview they give you a problem and solve it with one method. You have to carefully select your representation. How would you find the best parameter. Combinatorial is going over different possibilities, and checking and validating them. Convex Optimization is continuous optimization is continuous and constrained optimization has constraints in mind. In the domain of models, you have to select the best model that best describes your data. 

There are many types of learning. Supervised learning is with data and label with some human supervision. To collect that data, you show people the stack of images that have and label it, and this is spam this is not, learning is not supervised. Unsupervised learning is when the training data does not include desired outputs and finds interesting patterns in data. Semi-supervised learning is including few desired outputs and reinforcement learning is unsupervised learning with the algorithm that provides reward that shows the correctness of a prediction. 

Supervised learning is very expensive because you have to "pay" people to label many  datapoints for you. 

Here is supervised learning in a bit more mathematical terms: 

Expected loss is looking at the expectation of a value with the true data of the distribution.

 

As a result, we get the average loss of all the training data points. X is the raw input data and y is the label that I have.

The models we look up are usually probability distributions. 

MAP is maximum probability of the prediction which is the y hat = f(x). 

The other thing to think about is feature extraction which is how we describe the data. For example, for flowers, there's pedal width, pedal length, sepal length, and sepal width. 

We come up with multiple scatterplots based on this information. If the width is lower than the threshold, we can assume things. We can extend decision-making to a decision tree. The supervision is that you calculate the features of the data that you have. 

One approach is to come up with features based on human knowledge, or self-supervision is when we have a sentence, get a word and try to predict the word in the sentence. This part of the model attempts to predict features, and we use these features for other measures. We can learn features from labelled data and remove the sentence part after training and add a module that translate the sentences from English to German, and now the goal is to predict general words after the first features.

1. Learn features

2. Reuse the features to describe the structure of the English language and use them for translation in other tasks. 

We can measure MRI features and train features layered from own images to determine if a person has cancer or not. 

The only thing that the human can provide is a feature measurement that we have and the final label. We can have a feature as the input data and we need the label for training data. The data should be provided by any person, and the label should be provided by experts. The label always should be available for the training data. We feed whatever the algorithm is trained on. How can we get from human-designed features to raw inputs?  You have to implement everything from scratch, and numpy is the easiest way.