-
The basic regression model is represented as shown above.
-
By increasing the number of layers or increasing the units in between, learning is achieved.
-
Each arrow has a weight w, which has different values for each arrow.
-
These weights are adjusted through learning.
-
However, if that’s all, it would be the same as regular regression, so filters like ReLU or Tanh are applied.
-
It is also possible to regularize the weights to make them closer to zero, like in Ridge Regression.
-
By default, regularization is rarely applied.
-
Initially, the weights are determined randomly.
-
Analyzing the learned content is difficult, and one way to do it is to look at the heatmap of the weights.
-
There are Adam and LBFGS algorithms available for parameter learning, which are suitable for beginners. #Getting Started with Machine Learning in Python
Once the model is completed, when actually performing predictions, you just need to perform this calculation, simple. (x is the input, W is the weights for each layer, y is the output, and σ is the sigmoid function).
- Among the many perceptrons in a single layer, there may be one that has a very strong influence.
- To avoid this, dropout is randomly applied.
- This is done to avoid overfitting.
#udacity_intro_to_deep_learning_with_pytorch #Deep Learning