Deep_learning-Nano-Degree

Abstract

Advance Deep learning

Topics covered

1. ANN
2. CNN
3. RNN & LSTM
4. GAN
5. Transfer learning
6. style transfer
7. Intro to jupyter Notebooks

Introduction of Anaconda or jupyter Notebook

Jupyter Notebook is a web app

• Magic word like % or %%

• can be accessed any where by using cloud computing .

• & Anaconda is a

  Software distribution, package or Environment manager.\

Neural networks Fundamentals

>>Perceptron algorithm 
>>Error function 
>>Discrete(Step funcion) or continuous(sigmoid) Error functions
>>Softmax error function for multiclass classification
#####  Gradient Descent Algorithm    
       GDA= - learning rate * gradient 
        In simple words - ve slope for improve error function
        
  #### Cross Entropy = -(sigma of ln(probability of predictions))
                       it shows how accurate a neural network workikng .
                             
 #### Feed forwardation -->> process to take inputs and get final output 
 
 #### Backpropagation  is the central mechanism by which neural networks learn.

Implementation of Gradient Descent with Multilayer perceptron

Training of neural networks

overfitting vs underfitting ->> *in overfitting* training error is low but Testing error is high, *in underfittinng* Both errors
 are high.
 
 1./* Technique to remove overfitting*/ ->> Regularization  (L1,L2)
 2.[Dropout] : in this technique we drop some nodes which has more weight already or train only less
               Weight contain nodes. 

• [Types of Gradient descent]:
  • 1.Batch GD : take all data point in each epoch
  • 2. Stochastic GD : Take different data-sets in each epoch
  • 3.Mini Batch Gd : SGD + BGD

Introduction to Pytorch

• A DL framework develope by FB Ai research team
• Faster than Tensorflow
• Neural networks from scratch
• Torchvision library for importing different types of Vision data set Ex. MNSIT digit Recognition.. 🤗

CNN

• Better than MLP ,beacause In cnn layer each node is not connected to every node of Next layer Only Required ones
• 4 types of filters used in cnnn layer
• Applications: Style transfer, Transfer learning as VGG16,VGG19 trained with Image net dataset

Transfer learning

*pretrained models

Structure of code in Pytorch

   CNN layer -relu- > Max pooling - - >cNN layer-relu-pooling... + 3 normal layer Then apply softmax or any other activation function

RNN

Recurrent Neural network introduce memory in neural network

• Specially working area -Text processing Chat bot,Shri,GOOgle assistant
• *problem -Gradient vanishing
• LSTM Resolve this problem
• LSTM used 4 *Gates in model

GAN

• Generator Adversial Network "Produce fake images based on given images or data"
• GAN uses 2 models "Discriminator + Generator"
• Discriminator is a simple network classofy real images or produce real images but Generator produce Fake images By creating Errors in Model.

DCGAN

MOdel DEPLOY

contacts

• [email protected]
• LinkedIn- Niket kumar dheeryan

Useful Urls