MOOC-AI: Week - 2

Topic 2a – Overview of Types of AI

Highlighted Hidden from students Not available

Edit section Add subsection Highlight Display as a link Merge with parent Move Hide section Delete section
- Edit settings Move Move right Move left Hide Duplicate Assign roles Delete
  
  So, what is artificial intelligence? Artificial intelligence is the overarching name for a branch of computer science that deals with smart machines or robots that can think and perform tasks like a human. In artificial intelligence you can teach a computer to learn and make its own decisions. AI is the broad concept of machines being able to carry out tasks in a way that we would consider ‘smart’, and we encounter these technologies on a daily basis: from voice assistants in our home, to Netflix recommendations or autonomous cars.
  
  AI relies on data and on training algorithms. In this context, an algorithm is a set of rules for a computer to follow. In AI, data and training algorithms are what makes the computer seem intelligent, because they allow it to make decisions based on the rules that it has learned to follow.
  
  For a computer to learn it needs to find links between things, develop understanding and make decisions based on the training data that they are given. Sometimes, people can think of AI as a ‘black box’ – where you put everything into a machine and a magic formula will give you the exact results that you want. However, it is very important to remember that you only get out of AI what you put into it. So, if you enter poor quality data into a training algorithm, you will get poor results. Later in this week and throughout the course, we will go into more detail about training data and why it is so important in Earth observation.
  
  Machine Learning is a subset of artificial intelligence. It is the art of teaching a computer to learn and improve from experience rather than being explicitly programmed. Machine learning is an ‘empirical approach’, which means it learns from what is seen or experienced rather than theoretical models. We encounter machine learning in our day-to-day lives: for example, using face-detection to unlock our phones, or getting real-time traffic updates on Google Maps.
  
  The next subset of machine learning is deep learning. This is a kind of machine learning that aims to solve problems by mimicking the biological structure of the brain. To do this, they use a type of model called a neural network, which follows the way that the brain makes connections between different pieces of information. Later in this course, you will learn more about different kinds of neural networks.
  
  Today, deep learning is reaching such a high level of accuracy that it has the ability to replace many difficult, expensive and time-consuming tasks, which opens up huge opportunities for businesses and science, including in Earth observation.
  
  Furthermore, another subset of machine learning is reinforcement learning. Reinforcement learning is the training of machine learning models to make a sequence of decisions, taking into account interaction with the environment. In reinforcement learning, the AI faces a game-like situation: the computer employs a trial and error system to come up with a solution to the problem that it has been given. To get the machine to do what the programmer wants, the artificial intelligence will either be rewarded or penalised based on the actions it performs. Its goal is to maximise the total reward and solve the problem.
  
  An example of the use of reinforcement learning is the AlphaGo programme which was able to beat a professional human Go player for the first time.
  
  In this week, we will go into more detail about the kinds of problems that can be worked on using AI, and the different kinds of algorithms that can be used to solve these problems. Later in the course, you will see how these concepts are applied in the field of Earth system monitoring to solve problems that we face in the real world.
Topic 2b Part 1 – Types of Machine Learning Problem: Supervised Learning

Highlighted Hidden from students Not available

Edit section Add subsection Highlight Display as a link Merge with parent Move Hide section Delete section
- Edit settings Move Move right Move left Hide Duplicate Assign roles Delete
  In machine learning there are three kinds of learning problems (or tasks): these are supervised learning, unsupervised learning and reinforcement learning. The next two videos will cover supervised and unsupervised learning, and you will have a look at reinforcement learning in the first Jupyter Notebook task.
  
  Supervised learning is a method of training an algorithm where a data set consists of two parts: the data itself and a label. For example, a picture of a strawberry and a label which describes it as a strawberry.
  
  The algorithm is rewarded when it predicts the correct label, and it can rearrange itself if it predicts the wrong label.
  
  Similarly, the data can be associated with a real number, for example representing a geophysical parameter, rather than with a label. This is an example of a regression problem, where the task is to predict a continuous quantity.
  
  This can be used to do two things:
  
  Classification: sorting data into classes
  
  Regression: parameter estimation
Topic 2b Part 2 – Types of Machine Learning Problem: Unsupervised Learning

Highlighted Hidden from students Not available

Edit section Add subsection Highlight Display as a link Merge with parent Move Hide section Delete section
- Edit settings Move Move right Move left Hide Duplicate Assign roles Delete
  
  Next up is unsupervised learning: this is where a ML is given a training data set that does not have any labels.
  
  The ML is left to its own devices – and its goal is to find structure in the input that it was given.
  
  This can be useful in Earth observation, because scientists are often dealing with very large datasets.
  
  Like supervised data, unsupervised data can also be used for classification, clustering and feature identification.
  
  To make things more complicated, you can also perform semi-supervised learning. This is a combination of supervised and unsupervised learning. When you don’t have enough labelled data to produce an accurate model and you don’t have the ability or resources to get more data, you can use semi-supervised techniques to increase the size of your training data. Unlabelled data, when used in conjunction with a small amount of labelled data, can produce considerable improvement in learning accuracy.
  
  Jupyter Notebook Task - Reinforcement Learning
  
  Reinforcement learning is the training of machine learning models to make a sequence of decisions. For this first Jupyter Notebook task, you will have a look at an example of a reinforcement learning workflow.
  
  For this Jupyter Notebook workflow tutorial, please follow TBC to go to WEkEO. You can view the video for this tutorial both here on FutureLearn as well as at the top of the Jupyter Notebook.
  
  You may need to log in if you are not already. If you need to register with WEkEO you can do so here.
Topic 2c - Understanding Machine Learning Workflows

Highlighted Hidden from students Not available

Edit section Add subsection Highlight Display as a link Merge with parent Move Hide section Delete section
- Edit settings Move Move right Move left Hide Duplicate Assign roles Delete
  
  So, what’s the process of setting up a machine learning workflow?
  
  In simple terms: you firstly need to understand what your problem is.
  
  After that, you need to choose appropriate data and the correct learning algorithm.
  
  Then comes the most crucial part of machine learning: preparing your training data set.
  
  Once that’s done you can train your model. It is very important to evaluate the performance of the model. You do this by testing it on a selection of the training data that you held back from the ML, and didn’t use before starting the learning process.
  
  Then if everything has gone to plan, your problem is solved! But if necessary, you can re-run your model with optimised parameters.
  
  Throughout this course, you will have the opportunity to look at and interact with many ML workflows, like the one you have already seen in the step before this!
  
  Jupyter Notebook Task - Understanding Machine Learning Workflows
  
  In this Jupyter Notebook task, you will learn about machine learning workflows.
  
  For this Jupyter Notebook tutorial, please follow TBC to go to WEkEO. You can view the video for this tutorial both here on FutureLearn as well as at the top of the Jupyter Notebook.
  
  You may need to log in if you are not already. If you need to register with WEkEO you can do so here.
  
  If you have any questions, comments or would just like to discuss this Notebook with your fellow learners, please write in the discussion below.
Topic 2d – Common Machine Learning Algorithms

Highlighted Hidden from students Not available

Edit section Add subsection Highlight Display as a link Merge with parent Move Hide section Delete section
- Edit settings Move Move right Move left Hide Duplicate Assign roles Delete
  An algorithm is essentially a set of instructions that a computer follows to solve a problem. There are many kinds of machine learning (ML) algorithms, but these are some common examples that you will encounter during this course:
  
  K-Nearest Neighbour (KNN) - This is a simple classification and regression method. It assumes that similar things exist in proximity and therefore similar things are near to each other. The principle behind nearest neighbour methods is to find a predefined number of training samples closest in distance to the new point, and predict the classification label from these. In general, the distance can be any metric measure: standard Euclidean distance is the most common choice.
  
  Decision trees - These algorithms have a flowchart-like structure, which looks like an inverted tree. Each internal node or condition in the tree represents a “test” on an attribute (such as a yes or no question). The branches then split based on this test. . The end of the branch that is the decision, or leaf. In the world of AI, decision trees are used to develop learning machines by teaching them how to determine success and failure.
  
  Random Forest - This technique is a collection of decision trees, known as forests. For classification tasks, the output of the random forest is the class, or decision selected by the most trees.
  
  Support Vector Machines (SVM) - This is a linear model for classification and regression problems. It takes in data points and creates a line or a hyperplane which separates the data points into classes.
  
  Gradient boosting - This is another ensemble technique (like Random Forest), it uses a ‘boosting’ method to combine decision trees, which means that it takes in many weaker trees to create stronger trees. It does this by each tree attempting to minimise the errors of the previous tree. Since trees are added sequentially, boosting algorithms learn slowly, but slower learning algorithms are known to perform better.
  
  As well as these ML algorithms, you will also come across deep learning algorithms or methods, which are a modern update to Artificial Neural Networks (ANN) that exploit abundant cheap computation. They are concerned with building much larger and more complex neural networks. A neural network, or ANN, is an artificial system that mimics the way a human brain works, with neurons connected to synapses that transmit information. Neural networks can be trained so that they can recognise images, numbers and patterns, and they learn and improve their accuracy over time.
  
  Deep learning algorithms include:
  
  Convolutional Neural Network (CNN or ConvNet) - These are used for image recognition, pattern recognition, and computer vision.
  
  Recurrent Neural Network (RNN) - These are mostly used for time-series data to make predictions about future outcomes.
  
  Multi-Layer Perceptron (MLP) - These are used for classification prediction problems where inputs are assigned a class or label and regression prediction problems where a real-valued quantity is predicted given a set of inputs.
  
  Generative Adversarial Network (GAN) - These can be utilised for translating data from images, such as text to image translation or image to photo translation. They can also generate new images from a given set of images that are similar to the given dataset, yet individually different.
Topic 2e - Python Libraries for Machine Learning

Highlighted Hidden from students Not available

Edit section Add subsection Highlight Display as a link Merge with parent Move Hide section Delete section
- Edit settings Move Move right Move left Hide Duplicate Assign roles Delete
  
  More general frameworks:
  
  scikit-learn - Open-source ML framework in Python
  
  Tensorflow - Comprehensive ecosystem of ML tools, libraries and community resources. Allows to easily build and deploy ML powered applications. Keras is a Deep Learning Framework built on top of Tensorflow 2.0
  
  PyTorch - Open-source ML library based on the Torch library
  
  More specialised frameworks / libraries:
  
  xgboost - Open-source software library which provides a gradient boosting framework. Integrated into scikit-learn
  
  Jupyter Notebook Task - Python Libraries for ML
  
  This Jupyter Notebook task will provide you with an overview of the four most recurrent Python Libraries and frameworks.
  
  You can view the video for this tutorial both here on FutureLearn as well as at the top of the Jupyter Notebook.
  
  For this Jupyter Notebook workflow tutorial please follow this link to go to the WEkEO Jupyter Hub or click on the link in the ‘See Also’ section below.
  
  If you have any questions, comments or would just like to discuss this Notebook with your fellow learners, please write in the discussion below.
Topic 2f – Ethical Considerations

Highlighted Hidden from students Not available

Edit section Add subsection Highlight Display as a link Merge with parent Move Hide section Delete section
- Edit settings Move Move right Move left Hide Duplicate Assign roles Delete
  
  The use of AI often sparks debates about ethics and bias. AI can represent most human insight, but some human insight is inherently biased. It is important that we have a high degree of confidence that the data being used for AI is correct and accurate.
  
  EO research makes use of various types of data, such as satellite images and social media data. However, improved image resolution raises questions under the new European General Data Protection Regulation (GDPR) which came into force in 2018. The collection, storing, processing and transfer of such data needs to be done with much greater attention to legal and ethical requirements by adopting supporting technologies that give data contributors greater control over their personal data. Therefore, technologies such as blockchain are also likely to be researched and deployed, including from an ethical perspective, for fundamental AI4EM research.
  
  There are other ethical issues, such as algorithmic bias, resulting in discrimination, which need to be checked and addressed at early stages of technological development. Building responsible AI4EM that embodies social norms and values while ensuring sustainable and inclusive development is very important. Accordingly, we will address issues of data protection, privacy, data portability, and equality at the level not just of data collection, but also at the level of data use and dissemination within the sphere of ethics in AI4EM
  
  Data Insights: In Conversation with Joana Miguens
  
  In this first data insights video, we hear from Joana Miguens, who is Copernicus Data Services System Architect at EUMETSAT.
End of week summary

Highlighted Hidden from students Not available

Edit section Add subsection Highlight Display as a link Merge with parent Move Hide section Delete section
- Edit settings Move Move right Move left Hide Duplicate Assign roles Delete

Jupyter Notebook Task - Reinforcement Learning

Jupyter Notebook Task - Understanding Machine Learning Workflows

More general frameworks:

Jupyter Notebook Task - Python Libraries for ML

Data Insights: In Conversation with Joana Miguens