Please enable JavaScript.

Coggle requires JavaScript to display documents.

AI (Machine Learning (This detects patterns in very large amounts of data.…

- - - - The most simple type of machine learning and already pretty widely used. It describes what has happened.
    - - This analyses lots of data and anticipates what will happen next. Think 'most probable outcome'
    - - This provides a recommendation on what to do next, in order to achieve a goal. Most complex variety of machine learning
  - - - An algorithm uses data and human feedback to learn the relationship between different inputs and the output
      - You would use this when you know the classification of the data you need to input and how this relates to the output - you just need the computer to crunch the numbers
      - You 'train' the algorithm on a data set, then you can use it over any new data set
      - Buzzwords and descriptors are 'simple neural network', 'decision tree', 'quadratic discriminant analysis' 'linear regression'
    - - An algorithm looks at data without any explicit output defined.
      - You'd use this when you don't know what patterns there are in your data and you want the algorithm to find the patterns for you
      - You run the algorithm over the data and it will infer a structure of the data and find groups that have similar behaviours.
      - Buzzwords and descriptors are 'Gaussian mixture model', K-clustering
    - - An algorithm learns to perform a task for the greatest 'reward'.
      - You would use this when you cannot define the end state, or when the only way to learn about the environment is to interact with it (Think along the lines of chess games, AlphaGo etc)
      - The algorithm learns by taking an action, receiving a reward (or not), optimising it's future actions to take the best course
      - This is how you build self driving cars and automated stock trading platforms
  - - - Convolutional neural networks
        
        extracts increasing levels of complex features as it progresses through layers
        
        Used for unstructured data - e.g. images
        
        It identifies features of what makes up an image and can use that to classify a new image as containing those features. (Think - this is what makes up 'a dog' - new picture 'this photo has a dog in it')
        
        Used for interpreting medical scans, finding defective parts on a production line
      - recurrent neural networks
        
        stores information in context nodes
        
        used for time sequenced data - e.g. audio recordings
        
        This type of neural network does not assume that all the data is independent from each other. So this one looks at what has happened before in the 'stored context' - think of predictive text
        
        Used for credit card fraud, changes to an area after a natural disaster