Deep Learning
- It's a class of Machine Learning models that uses the same training algorithm technique to learn from data:
backpropagation
- Requires large amount of input data
- Requires GPUs/NPUs for processing
- Use cases
- Computer vision
- Natural Language Processing (NLP)
Tensors
- The whole input is transformed into a
multidimensional array of numbers called Tensors
Layering
- The tensor suffers several transformations on multiple
layers and the final layer is the desired output
- On each layer the tensor is tunned by applying the weights
Weights
- Weights are the paramters of the model
- On each layer, the parameters (
weights) are applied (by weighted sum) to the current tensor, tunning the tensor
- The
weights is organized as a matrix, that is used to multiply the tensors
- E.g., GPT-3 has 175 billion weights (parameters) organized into 27 thousand matrices
Backpropagation