Grokking Deep Learning (Andrew Trask)

(Stopped at page 61 file:///Users/lakesare/Downloads/Grokking_Deep_Learning_by_Andrew_W_Trask_z-lib_org.pdf)

4 flashcards · Computer Science · by lakesare

  1. Supervised learning VS unsupervised learning? As you can see, there are really four different types of algorithms to choose from. An algorithm is either unsupervised or supervised, and either parametric or nonparametric.

    supervised learning - learns how to transform one dataset into another. unsupervised learning - groups your data. All forms of unsupervised learning can be viewed as a form of clustering.Unlike supervised learning, there is no "right answer" that you’re trying to get the model to duplicate. You just tell an unsupervised algorithm to "find patterns in this data and tell me about them".The algorithm doesn’t tell you what the clusters are. How could it know? It just says, "Hey scientist! I found s…

  2. Parametric learning VS nonparametric learning? As you can see, there are really four different types of algorithms to choose from. An algorithm is either unsupervised or supervised, and either parametric or nonparametric.

    parametric learning - fixed number of parameters Note that the entirety of what the model has learned can be captured in the positions of the knobs at any given time. You can also think of this type of learning model as a search algorithm. You’re “searching” for the appropriate knob configuration by trying configurations, adjusting them, and retrying.nonparametric learning - parameters determined by data. You learned that parametric algorithms have a fixed number of parameters and that nonparam…

  3. Why do we use an error (result - goal)2 that is positive makes 0.1 smaller, and 10 larger?

    positive - because we orient ourselves by the average error. makes 0.1 smaller, and 10 larger - because it's nice to only focus on the larger errors.

  4. After we know our error, name 2 ways to adjust our weights. Knobs == weight.

    Hot and cold learning - try to reduce AND increase the weight, and determine which of these reduces the error! How do you know whether to turn the knob up or down? Well, you try both up and down and see which one reduces the error! Whichever one reduces the error is used to update the weight. Gradient descent