Deep Learning - Artificial Intelligence
- Igor Aizenberg - 2000
- Algo inspired by structure and function of human brain
- Uses interconnected nodes or neurons in a layered structure that resembles the human brain
- Uses Backpropagation in Training modal
- Input is represented in array of higher dimensional real numbers called Tensor, getting transformed through different Layers
- Weights => Modal parameters, Brain of the modal, Learned during training
- Parameters interact through Data only using Weighted sums
- Comparing with ML
- Training time is more and Testing time is less compared to ML
- When amount of data is small, ML works better, but when data is large DL is better
- Can handle huge amount of data and Unstructured data, Can solve complex problems
- Less manual input required for feature extraction
- Like Linear regression has 2 parameters, GPT3 has 175 Billion parameters
- Artificial Neural Network
- Artificial creation of neural networks of brain
- Calculate weighted sum and send it to activation function to generate a particular output/action for that particular input
- First Artificial Neuron
- Buffer, AND, OR like basic operation representation
- Activation Function
- Linear Function
- Heviside Step Function
- Sigmoid Function
- Non-linear & Non-zero centered activation function
- Vanishing Gradient
- Hyberbolic Tangent Function
- tanh
- Non-linear & Zero centered activation function
- Vanishing Gradient
- Range = [-1, 1]
- Domain = [-1, 1]
- Rectified Liner Unit (ReLU)
