1) Machine Vision

• Biological Vision techniques are inspired to build modern vision techniques[Deep Learning]

Advantages of Machine Vision:

• Accuracy
• Consistency
• Efficiency
• Fastness
• Reduction in Human Error
• Saves time and cost

Libraries used for Machine Vision:

• OpenCV
• TensorFlow
• Keras
• PyTorch
• MATLAB

2) NLP

• NLP stands for Natural Language Processing

NLP tools:

• NLTK (Natural Language Tool Kit)
• SpaCy
• OpenNLP
• Gensim
• Cloud-based NLP:
  • AWS
  • Azure
  • Google Cloud

NLP Algorithms:

• Tokenization Algorithm
• Parsing Algorithm
• Part-of-speech Algorithm
• Classification Algorithm
• Clustering Algorithm

Role of AI in NLP:

• Understanding text:
  • AI reads the text and understands its meaning
• Understanding speech:
  • AI listens to the speech and understands its meaning
• Responds:
  • AI responds to the text and speech
• Speech Recognition:
  • AI recognizes the speech and converts it into text format
• Language Translation:
  • AI converts from one language to another language
  • Example:
    • Google Translate: English to Telugu

Example:

• AI Chatbots:
  • ChatGPT
  • perplexity
• Voice Assistants:
  • Siri
  • Alexa

3) Generative Adversarial Networks[GAN]

• GAN stands for Generative Adversarial Networks

Generator:

• A Generator creates fake data from random inputs
• The Goal of the Generator is to produce fake data that looks like realistic one

Real Data and Fake Data:

• Real Data: Real sample from the dataset
• Fake Data: Fake sample generated by Generator

Discriminator:

• Discriminator is the binary classifier that classifies whether the data is real or fake data

Types of GAN:

• There are so many types of GANs. Some of them are:
  • Vanilla GAN
  • Conditional GAN
  • Cycle GAN
  • Deep Convolutional GAN
  • Super Resolution GAN

Applications of GAN:

• Image Generation
• Face Generation
• Image-to-Image translation
• Enhancing Image Quality

4) Autoencoders and DAE

• Autoencoders are used to compress the input data and reconstruct it to the Original data
• The Main goal of an Autoencoder is to rebuild the existing data
• There are mainly two components. They are:
  • Encoders
  • Decoders

Encoders:

• It compresses the input data into a latent space

Decoders:

• It reconstructs the compressed data into the Original data

DAE:

• DAE stands for Denoising Autoencoders
• DAE is a type of Autoencoder designed to remove noise from the input data

5) What are the steps involved in a typical deep reinforcement learning algorithm?

• Reinforcement Learning is a technique for training the machine by using an agent and an Environment
• This Technique is used to produce actions and rewards in a given situation

Steps:

• Initialize the Agent and Environment
• The agent selects an action
• Execute the action in an Environment
• Environment returns the next state and Rewards
• They are stored in the Memory
• They connect this to the Neural Network and update the policies

6) Restricted Boltzmann Machines

• RBM stands for Restricted Boltzmann Machines
• RBM mainly consist of two units:
  • visible units
  • hidden units
• RBM trained based on the CD Algorithm
• There are mainly two phases in RBM:
  • Positive Phase
    • Updates the weight based on visible units
  • Negative Phase
    • Updates the weight based on hidden units

Types of RBM:

• There are mainly two types
  • Binary RBM
  • Gaussian RBM
• Binary RBM:
  • Input and Hidden Units are binary variables
• Gaussian RBM:
  • Input and Hidden Units follow a Gaussian Distribution

Difference between Boltzmann Machines and Restricted Boltzmann Machines

Boltzmann Machines	Restricted Boltzmann Machines
Visible to Visible is possible	Visible to Visible is not possible
Hidden to Hidden is possible	Hidden to Hidden is not possible
Visible to Hidden is possible	Visible to Hidden is possible
It is a fully connected bipartite graph	It is a bipartite graph
Hard to train	easy to train
works slower	works faster

7) Deep Belief Network and Deep-Net in Deep Learning

• DBN stands for Deep Belief Network
• Deep-Net:
  • Deep-Net(Deep-Network) is a neural network that has multiple hidden layers between the input and output layers
• A Deep Belief Network is made by stacking multiple Restricted Boltzmann Machines
• The Output of one RBM is an input for the next RBM
• There are mainly two ways to train the DBN. They are:
  • UnSupervised Pre-Training
  • Supervised fine-tuning