Artificial Intelligence & Deep Learning Course Training

The Perceptron Algorithm is defined based on a biological brain model. You will talk about the parameters used in the perceptron algorithm which is the foundation of developing much complex neural network models for AI applications. Understand the application of perceptron algorithms to classify binary data in a linearly separable scenario.

• Neurons of a Biological Brain
• Artificial Neuron
• Perceptron
• Perceptron Algorithm
• Use case to classify a linearly separable data
• Multilayer Perceptron to handle non-linear data

Neural Network is a black box technique used for deep learning models. Learn the logic of training and weights calculations using various parameters and their tuning. Understand the activation function and integration functions used in developing a Artificial Neural Network.

• Integration functions
• Activation functions
• Weights
• Bias
• Learning Rate (eta) - Shrinking Learning Rate, Decay Parameters
• Error functions - Entropy, Binary Cross Entropy, Categorical Cross Entropy, KL Divergence, etc.

• Artificial Neural Networks
• ANN Structure
• Error Surface
• Gradient Descent Algorithm
• Backward Propagation
• Network Topology
• Principles of Gradient Descent (Manual Calculation)
• Learning Rate (eta)
• Batch Gradient Descent
• Stochastic Gradient Descent
• Minibatch Stochastic Gradient Descent
• Optimization Methods: Adagrad, Adadelta, RMSprop, Adam
• Convolution Neural Network (CNN)
• ImageNet Challenge – Winning Architectures
• Parameter Explosion with MLPs
• Convolution Networks
• Recurrent Neural Network
• Language Models
• Traditional Language Model
• Disadvantages of MLP
• Back Propagation Through Time
• Long Short-Term Memory (LSTM)
• Gated Recurrent Network (GRU)

Learn about single-layered Perceptrons, Rosenblatt’s perceptron for weights and bias updation. You will understand the importance of learning rate and error. Walk through a toy example to understand the perceptron algorithm. Learn about the quadratic and spherical summation functions. Weights updating methods - Windrow-Hoff Learning Rule & Rosenblatt’s Perceptron.

• Introduction to Perceptron
• Introduction to Multi-Layered Perceptron (MLP)
• Activation functions – Identity Function, Step Function, Ramp Function, Sigmoid Function, Tanh Function, ReLU, ELU, Leaky ReLU & Maxout
• Back Propagation Visual Demonstration
• Network Topology – Key characteristics and Number of layers
• Weights Calculation in Back Propagation

Understand the difference between perception and MLP or ANN. Learn about error surface, challenges related to gradient descent and the practical issues related to deep learning. You will learn the implementation of MLP on MNIST dataset - multi class problem, IMDB dataset - binary classification problem, Reuters dataset - single labelled multi class classification problem and Boston Housing dataset - Regression Problem using Python and Keras.

• Error Surface – Learning Rate & Random Weight Initialization
• Local Minima issues in Gradient Descent Learning
• Is DL a Holy Grail? Pros and Cons
• Practical Implementation of MLP/ANN in Python using Real Life Use Cases
• Segregation of Dataset - Train, Test & Validation
• Data Representation in Graphs using Matplotlib
• Deep Learning Challenges – Gradient Primer, Activation Function, Error Function, Vanishing Gradient, Error Surface challenges, Learning Rate challenges, Decay Parameter, Gradient Descent Algorithmic Approaches, Momentum, Nestrov Momentum, Adam, Adagrad, Adadelta & RMSprop
• Deep Learning Practical Issues – Avoid Overfitting, DropOut, DropConnect, Noise, Data Augmentation, Parameter Choices, Weights Initialization (Xavier, etc.)

Convolution Neural Network are the class of Deep Learning networks which are mostly applied on images. You will learn about ImageNet challenge, overview on ImageNet winning architectures, applications of CNN, problems of MLP with huge dataset.

You will understand convolution of filter on images, basic structure on convent, details about Convolution layer, Pooling layer, Fully Connected layer, Case study of AlexNet and few of the practical issues of CNN.

• ImageNet Challenge – Winning Architectures, Difficult Vision Problems & Hierarchical Approach
• Parameter Explosion with MLPs
• Convolution Networks - 1D ConvNet, 2D ConvNet, Transposed Convolution
• Convolution Layers with Filters and Visualizing Convolution Layers
• Pooling Layer, Padding, Stride
• Transfer Learning - VGG16, VGG19, Resnet, GoogleNet, LeNet, etc.
• Practical Issues – Weight decay, Drop Connect, Data Manipulation Techniques & Batch Normalization

You will learn image processing techniques, noise reduction using moving average methods, different types of filters - smoothing the image by averaging, Gaussian filter and the disadvantages of correlation filters. You will learn about different types of filters, boundary effects, template matching, rate of change in the intensity detection, different types of noise, image sampling and interpolation techniques.

You will also learn about colors and intensity, affine transformation, projective transformation, embossing, erosion & dilation, vignette, histogram equalization, HAAR cascade for object detection, SIFT, SURF, FAST, BRIEF and seam carving.

• Introduction to Vision
• Importance of Image Processing
• Image Processing Challenges – Interclass Variation, ViewPoint Variation, Illumination, Background Clutter, Occlusion & Number of Large Categories
• Introduction to Image – Image Transformation, Image Processing Operations & Simple Point Operations
• Noise Reduction – Moving Average & 2D Moving Average
• Image Filtering – Linear & Gaussian Filtering
• Disadvantage of Correlation Filter
• Introduction to Convolution
• Boundary Effects – Zero, Wrap, Clamp & Mirror
• Image Sharpening
• Template Matching
• Edge Detection – Image filtering, Origin of Edges, Edges in images as Functions, Sobel Edge Detector
• Effect of Noise
• Laplacian Filter
• Smoothing with Gaussian
• LOG Filter – Blob Detection
• Noise – Reduction using Salt & Pepper Noise using Gaussian Filter
• Nonlinear Filters
• Bilateral Filters
• Canny Edge Detector - Non Maximum Suppression, Hysteresis Thresholding
• Image Sampling & Interpolation – Image Sub Sampling, Image Aliasing, Nyquist Limit, Wagon Wheel Effect, Down Sampling with Gaussian Filter, Image Pyramid, Image Up Sampling
• Image Interpolation – Nearest Neighbour Interpolation, Linear Interpolation, Bilinear Interpolation & Cubic Interpolation
• Introduction to the dnn module
  • Deep Learning Deployment Toolkit
  • Use of DLDT with OpenCV4.0
• OpenVINO Toolkit
  • Introduction
  • Model Optimization of pre-trained models
  • Inference Engine and Deployment process

Understand the language models for next word prediction, spell check, mobile auto-correct, speech recognition, and machine translation. You will learn the disadvantages of traditional models and MLP. Deep understanding of the architecture of RNN, RNN language model, backpropagation through time, types of RNN - one to one, one to many, many to one and many to many along with different examples for each type.

• Introduction to Adversaries
• Language Models – Next Word Prediction, Spell Checkers, Mobile Auto-Correction, Speech Recognition & Machine Translation
• Traditional Language model
• Disadvantages of MLP
• Introduction to State & RNN cell
• Introduction to RNN
• RNN language Models
• Back Propagation Through time
• RNN Loss Computation
• Types of RNN – One to One, One to Many, Many to One, Many to Many
• Introduction to the CNN and RNN
• Combining CNN and RNN for Image Captioning
• Architecture of CNN and RNN for Image Captioning
• Bidirectional RNN
• Deep Bidirectional RNN
• Disadvantages of RNN
• Frequency-based Word Vectors
• Count Vectorization (Bag-of-Words, BoW), TF-IDF Vectorization
• Word Embeddings
• Word2Vec - CBOW & Skip-Gram
• FastText, GloVe

Faster object detection using YOLO models will be learnt along with setting up the environment. Learn pretrained models as well as building models from scratch.

• YOLO v3
• YOLO v4
• Darknet
• OpenVINO
• ONNX
• Fast R-CNN
• Faster R-CNN
• Mask R-CNN

Understand and implement Long Short-Term Memory, which is used to keep the information intact, unless the input makes them forget. You will also learn the components of LSTM - cell state, forget gate, input gate and the output gate along with the steps to process the information. Learn the difference between RNN and LSTM, Deep RNN and Deep LSTM and different terminologies. You will apply LSTM to build models for prediction.

Gated Recurrent Unit, a variant of LSTM solves this problem in RNN. You will learn the components of GRU and the steps to process the information.

• Introduction to LSTM – Architecture
• Importance of Cell State, Input Gate, Output Gate, Forget Gate, Sigmoid and Tanh
• Mathematical Calculations to Process Data in LSTM
• RNN vs LSTM - Bidirectional vs Deep Bidirectional RNN
• Deep RNN vs Deep LSTM

• Seq2Seq (Encoder - Decoder Model using RNN variants)
• Attention Mechanism
• Transformers (Encoder - Decoder Model by doing away from RNN variants)
• Bidirectional Encoder Representation from Transformer (BERT)
• OpenAI GPT-2 & GPT-3 Models (Generative Pre-Training)
• Text Summarization with T5
• Configurations of BERT
• Pre-Training the BERT Model
• ALBERT, RoBERTa, ELECTRA, SpanBERT, DistilBERT, TinyBERT

You will learn about the components of Autoencoders, steps used to train the autoencoders to generate spatial vectors, types of autoencoders and generation of data using variational autoencoders. Understanding the architecture of RBM and the process involved in it.

• Autoencoders
  • Intuition
  • Comparison with other Encoders (MP3 and JPEG)
  • Implementation in Keras
• Deep AutoEncoders
  • Intuition
  • Implementing DAE in Keras
• Convolutional Autoencoders
  • Intuition
  • Implementation in Keras
• Variational Autoencoders
  • IntuitionImplementation in Keras
• Introduction to Restricted Boltzmann Machines - Energy Function, Schematic implementation, Implementation in TensorFlow

You will learn the difference between CNN and DBN, architecture of deep belief networks, how greedy learning algorithms are used for training them and applications of DBN.

• Introduction to DBN
• Architecture of DBN
• Applications of DBN
• DBN in Real World

Understanding the generation of data using GAN, the architecture of the GAN - encoder and decoder, loss calculation and backpropagation, advantages and disadvantages of GAN.

• Introduction to Generative Adversarial Networks (GANS)
• Data Analysis and Pre-Processing
• Building Model
• Model Inputs and Hyperparameters
• Model losses
• Implementation of GANs
• Defining the Generator and Discriminator
• Generator Samples from Training
• Model Optimizer
• Discriminator and Generator Losses
• Sampling from the Generator
• Advanced Applications of GANS
  • Pix2pixHD
  • CycleGAN
  • StackGAN++ (Generation of photo-realistic images)
  • GANs for 3D data synthesis
  • Speech quality enhancement with SEGAN

You will learn to use SRGAN which uses the GAN to produce the high-resolution images from the low-resolution images. Understand about generators and discriminators.

• Introduction to SRGAN
• Network Architecture - Generator, Discriminator
• Loss Function - Discriminator Loss & Generator Loss
• Implementation of SRGAN in Keras

You will learn Q-learning which is a type of reinforcement learning, exploiting using the creation of a Q table, randomly selecting an action using exploring and steps involved in learning a task by itself.

• Reinforcement Learning
• Deep Reinforcement Learning vs Atari Games
• Maximizing Future Rewards
• Policy vs Values Learning
• Balancing Exploration With Exploitation
• Experience Replay, or the Value of Experience
• Q-Learning and Deep Q-Network as a Q-Function
• Improving and Moving Beyond DQN
• Keras Deep Q-Network

Learn to Build a speech to text and text to speech models. You will understand the steps to extract the structured speech data from a speech, convert that into text. Later use the unstructured text data to convert into speech.

• Speech Recognition Pipeline
• Phonemes
• Pre-Processing
• Acoustic Model
• Deep Learning Models
• Decoding

Learn to Build a chatbot using generative models and retrieval models. We will understand RASA open-source and LSTM to build chatbots.

• Introduction to Chatbot
• NLP Implementation in Chatbot
• Integrating and implementing Neural Networks Chatbot
• Introduction to Sequence to Sequence models and Attention
  • Transformers and it applications
  • Transformers language models
    • BERT
    • Transformer-XL (pretrained model: “transfo-xl-wt103”)
    • XLNet
• Building a Retrieval Based Chatbot
• Deploying Chatbot in Various Platforms

Learn the tools which automatically analyzes your data and generates candidate model pipelines customized for your predictive modeling problem.

• AutoML Methods
  • Meta-Learning
  • Hyperparameter Optimization
  • Neural Architecture Search
  • Network Architecture Search
• AutoML Systems
  • MLBox
  • Auto-Net 1.0 & 2.0
  • Hyperas
• AutoML on Cloud - AWS
  • Amazon SageMaker
  • Sagemaker Notebook Instance for Model Development, Training and Deployment
  • XG Boost Classification Model
  • Training Jobs
  • Hyperparameter Tuning Jobs
• AutoML on Cloud - Azure
  • Workspace
  • Environment
  • Compute Instance
  • Compute Targets
  • Automatic Featurization
  • AutoML and ONNX

Learn the methods and techniques which can explain the results and the solutions obtained by using deep learning algorithms.

• Introduction to XAI - Explainable Artificial Intelligence
• Why do we need it?
• Levels of Explainability
  • Direct Explainability
    • Simulatability
    • Decomposability
    • Algorithmic Transparency
  • Post-hoc Explainability
    • Model-Agnostic Algorithms
      • Explanation by simplification (Local Interpretable Model-Agnostic Explanations (LIME))
      • Feature relevance explanation
        • SHAP
        • QII
        • SA
        • ASTRID
        • XAI
      • Visual Explanations
• General AI vs Symbolic Al vs Deep Learning
• Check out the Deep Learning Interview Questions here.

• A open-source AutoML framework based on a popular Python library Keras. It allows a non-programmer also to use advanced high-performance DL models with hyperparameter searching. Check out the AutoKeras - A New Revolution into Deep Learning here.

A Large Language Model (LLM) in the context of data science typically refers to advanced natural language processing (NLP) models, which I am based on. These LLMs are designed to understand and generate human-like text, making them useful for a variety of data science tasks.

Playgrounds provide a sandbox-like setting where users can test different algorithms, models, and methodologies to gain insights and improve their skills.

DALL-E is a groundbreaking generative model in the field of data science and artificial intelligence, developed by OpenAI. The name "DALL-E" is a combination of the famous artist Salvador Dalí and the robot character WALL-E from the Pixar film.

• Deriving insights by performing sentiment analysis with Text Mining and NLP
• Learn the Transformer architectures implemented for machine translation and automated summarization
• Learn about the groundbreaking uses of BERT and GPT in automated content creation.
• ImplementingChatbots to enhance interaction quality and operational efficiency for effective customer engagement across industries.

• Learn how LLMs like GPT and BERT are revolutionizing with improved accuracy of information retrieval across digital libraries
• Learn to build specialized models in healthcare, and the financial sector as part of Domain-Specific LLMs
• Discover the role of LLMs in developing sophisticated AI chatbots
• Understand the models that are applied in real-time translation services
• Dive into the use of generative models for original written content generation

• Exploring the World of AI-generated imagery with Dall-E 3 and Imagen
• Learn the creation of realistic video sequences from textual inputs, significantly reducing production time and costs.
• Leveraging tools like ViT and SDXL Turbo in Automated Image Analysis
• Explore the use of advanced AI models like BLIP-2 and Flamingo inenhancing user experience
• Discover how models are used in e-commerce platforms for image-based product search and recommendation systems
• Learnabout the GenAI models for AI-driven object detection in real-time decision-making.

• Explore the capabilities of Generative AI tools like Soundful and Lyria in transforming music production
• Enhancing Multimedia Experiences with the Video-Audio-Text Transformer (VATT)
• Improving audio quality and enabling innovative sound manipulation techniques with models like Distil-Whisper and SALMONN
• enhancing user engagement through tailored listening experiences.
• Enhancing user engagement through tailored listening experiences by automating Music Curation and Personalization with GenAI

• Introduction to learning platforms that respond to both text and image inputs.
• Explore advanced models like LLaVA and GPT-4V for integrating visual and textual information
• Transforming digital marketing, storytelling, and educational content generation with models like Fuyu and Fuyu-8Bthat combine text, image, and video.
• Learn about the importance of diverse datasets like PMD, COCO, and Flickr30K in training AI
• Examine how benchmarks like VQAv2 and TextVQA are used to evaluate the effectiveness of multimodal AI models in generating human-like responses
• Visual search and image-based analysis with innovative models BriVL and Qwen-VL
• Creating enriched multimedia experiences in virtual reality, video editing, and automated captioning with SALMONN

• Develop Generative AI models with AWS services: Amazon Bedrock and SageMaker
• Generating code snippets and entire functions with Amazon CodeWhisperer
• Using high-performance computing services AWS Trainium and Inferentia to Optimizethe training and deployment of AI models
• Innovating with Generative AI on the Cloud:
• Discover the broad spectrum of Generative AI applications on Cloud platforms