- GEOFF HILTON {Father of DNN}
- MLP {Multi Layered Perceptron} -> Vanialla Neural Network
- Problem of Vanishing Gradient {This is what RBM Solves}
- Forward/Backward Phase
- Idenfity patterns in dataset
- Denoising
- Contractive
- Unlabelled Input -> Try to Reconstruct Input as Possible
- Feature Extraction Engine
- Shallow
- Backprogation with Lost
- Deep Auto-Encoder [28 x 28] -> 30
- Better than PCA
- Combine multiple RBMs
- Hidden Layer of one RBM is Visible Layer of another RBM
- Useful in scenarios where we have small sample of Trained Data
- Training process get completed in reasonable amount of time
- Machine Vision is where CNN is used
- ImageNet competition
- Andrej Karpathy's Note {http://cs231n.github.io/}
- Layers:
- Convolution Layer
- Technical Operation {Convolution} to search for pattern
- Relu {Rectified Linear Unit}
- Used for reducing training time for early layers
- Pooling
- Used for dimensionality reduction
- Fully Connected Layer
- Allows network to classify data from samples based on patterns discovered by Polling layer
- Convolution Layer
- Since this is supervised method, take large sample of labeled data
- If patterns change over time is where RNN is useful
- Application {Speech Recognization, Driverless Car}
- Output of layer is fed back to the same layer along with inputs
- Applications
- Sequences of Input to Produce Sequences of Output
- Image Captioning {Singel Input, Multiple Output}
- Document Classification {Single Input, Single Output}
- Video Classification {Multiple Input, Multiple Output}
- Forecasting {with time Delays}
- Stacking
- Difficult to train because of Vanishing Gradients
- Solution:
- Gating {Method of deciding for network when to forget current input or consider current input}
- LSTM
- GRU
- Gradient Clipping
- Better Optimizers
- Steeper Gates
- Gating {Method of deciding for network when to forget current input or consider current input}
- GPUs prefered methods of training
- RNN suited for timeseries data
- Feedforward: Classification/Regression
- Recurrent Neural Network: Forecaster
- For applications like Parsing
- 3 Basic Component
- Root {Fires out Class and Score}
- Child Nodes {Leafs}
- Collection of Recursive Binary Tress
- Score repesent quality
- Text
- Parsing
- Sentiement Analysis
- Image Parsing {Break Image into many different components}