Artificial Intelligence Course in Mumbai

• Definitions

  • AI, machine learning and deep learning
  • narrow versus general AI
  • what changed in the last five years

• A short history

  • symbolic AI and expert systems
  • the AI winters
  • the deep learning breakthrough
  • the transformer era

• Capabilities and limits

  • pattern recognition versus reasoning
  • generalisation failures
  • adversarial examples
  • what benchmarks do not measure

• The ecosystem

  • open versus closed models
  • model sizes and what they cost to run
  • inference versus training

• Setting up

  • Python environment
  • PyTorch installation
  • free GPU notebooks
  • managing large model files

You can place any AI system in context and say what it can and cannot do.

• The neuron

  • weights, bias and activation
  • the perceptron
  • linear separability
  • why we needed layers

• Activations

  • sigmoid and tanh
  • ReLU and its variants
  • softmax for classification
  • why activation choice matters

• Forward propagation

  • matrix form
  • layer composition
  • the output layer for different tasks

• Loss functions

  • mean squared error
  • cross-entropy
  • why cross-entropy for classification

• Backpropagation

  • the chain rule applied
  • gradients layer by layer
  • implementing it manually once

• Training loop

  • epochs and batches
  • optimisers
  • learning rate
  • monitoring train and validation loss

You can implement forward and backward passes and explain every step.

• Optimisation in practice

  • SGD with momentum
  • Adam and AdamW
  • learning rate schedules
  • warmup

• Regularisation

  • dropout
  • weight decay
  • batch and layer normalisation
  • early stopping

• Initialisation and stability

  • why initialisation matters
  • vanishing and exploding gradients
  • gradient clipping

• Data handling

  • datasets and dataloaders
  • augmentation
  • class imbalance
  • train-validation-test discipline

• Debugging training

  • loss not decreasing
  • overfitting a single batch as a sanity check
  • reading learning curves
  • when the bug is in the data

You can diagnose a failing training run and fix it.

• Images as data

  • pixels and channels
  • resizing and normalisation
  • colour spaces
  • OpenCV basics

• Convolutional networks

  • the convolution operation
  • kernels and feature maps
  • stride and padding
  • pooling
  • receptive field

• Architectures

  • LeNet to ResNet
  • skip connections and why they helped
  • efficient models for edge devices
  • vision transformers in outline

• Transfer learning

  • pretrained backbones
  • freezing and fine-tuning
  • when a small dataset is enough

• Beyond classification

  • object detection concepts
  • YOLO at a high level
  • segmentation
  • annotation and dataset building

• Evaluation

  • accuracy versus per-class performance
  • confusion matrix on images
  • failure case analysis

You can build an image model using transfer learning and evaluate it properly.

• Text preprocessing

  • tokenisation
  • subword tokenisers
  • normalisation
  • handling Indian language text

• Classical approaches

  • bag of words
  • TF-IDF
  • naive Bayes for text
  • when classical is the correct answer

• Embeddings

  • word2vec intuition
  • contextual embeddings
  • sentence embeddings
  • similarity and semantic search

• Transformers

  • self-attention
  • multi-head attention
  • positional encoding
  • encoder and decoder stacks
  • why they replaced RNNs

• Using pretrained models

  • Hugging Face pipelines
  • fine-tuning a classifier
  • LoRA and parameter-efficient tuning
  • model selection by size and licence

You can solve a language task with the right tool for its size and budget.

• How LLMs work

  • next-token prediction
  • pretraining and fine-tuning
  • instruction tuning
  • RLHF in outline
  • context windows

• Prompting as engineering

  • structured prompts
  • few-shot examples
  • chain of thought
  • forcing structured output
  • prompt versioning

• Retrieval augmented generation

  • chunking strategies
  • embeddings and vector stores
  • retrieval quality
  • citing sources
  • why RAG beats fine-tuning for facts

• Generative models beyond text

  • diffusion model intuition
  • image generation
  • speech synthesis
  • multimodal models

• Evaluating generative output

  • why accuracy does not apply
  • human evaluation
  • LLM-as-judge and its biases
  • regression testing prompts

You can build a grounded LLM application and evaluate its output.

• Serving models

  • FastAPI endpoints
  • batching requests
  • quantisation and model size
  • CPU versus GPU inference
  • cold starts

• Working with model APIs

  • hosted model APIs
  • streaming responses
  • token cost management
  • caching
  • rate limit handling

• System design

  • when to call a model and when not to
  • fallbacks when the model fails
  • human review for consequential decisions

• Monitoring

  • logging inputs and outputs
  • quality drift
  • cost per request
  • user feedback loops

• Edge and on-device

  • model compression
  • ONNX
  • running a model on modest hardware

You can ship an AI feature with sensible cost, latency and failure behaviour.

• Bias and fairness

  • where bias enters
  • measuring disparate performance
  • mitigation and its trade-offs
  • documented limitations

• Privacy

  • training data provenance
  • personal data in prompts
  • the DPDP Act in outline
  • data minimisation

• Security of AI systems

  • prompt injection
  • data exfiltration through a model
  • model and output validation

• Transparency

  • disclosing AI involvement
  • explainability expectations
  • deepfakes and synthetic media
  • consent for voice and likeness

• Keeping current

  • reading papers selectively
  • evaluating a new model claim
  • distinguishing capability from marketing

You can assess an AI system for risk and explain your assessment.