GINNI GARG

Highlighted Projects

Name Similarity & Intelligent Name Matching using Fine-Tuned MPNet (Contrastive Loss)

Fine-tuned the MPNet Base v2 transformer model for a high-accuracy Name Similarity task to solve real-world name matching challenges such as token shuffling, spacing inconsistencies, honorific prefixes, and partial name variations. Expanded the dataset from 60K to 80K+ records through targeted augmentation applied to matched-name pairs (label = 1), including space removal (e.g., “RAHUL SHARMA” → “RAHULSHARMA”), token shuffling ("RAHUL KUMAR SHARMA" → "SHARMA RAHUL KUMAR"), middle-name dropping ("RAHUL KUMAR SHARMA" → "RAHUL SHARMA"), and random concatenation strategies ("RAHUL KUMAR SHARMA" → "RAHULKUMAR SHARMA") to improve model robustness. Implemented intelligent preprocessing by normalizing case, trimming whitespace, and removing prefixes such as mr, mrs, late, sh, lt, ensuring consistency across training and inference pipelines. Designed a dual-embedding strategy where two vectors were generated per name—(1) preprocessed original name and (2) preprocessed concatenated name—to handle structural name variations. Experimented with multiple similarity aggregation strategies and finalized a hybrid scoring formula: Similarity Score = max(V1, V2, mean(V1, V2)), which outperformed individual and mean-based approaches. Optimized contrastive loss by auto-computing the margin (0.62) from training/validation distributions instead of using a fixed heuristic margin (0.5), improving class separation. Achieved outstanding performance with AUC: 0.9991, Accuracy: 99%, and optimal threshold 0.8383, demonstrating production-grade reliability for large-scale identity matching systems such as TSP-based embedding pipelines.
Tech: mpnet-base-v2, Fine-Tune, Augmentation, Preprocessing, AUC-ROC, Recall/Precision, Hyper-parameter Tuning (Margin) for Contrastive Loss, Google Collab (T4)

AI-Based Name Matching System using Fine-Tuned MPNet (Triplet Margin Loss)

• Developed an AI-driven name matching system by fine-tuning the sentence-transformers/mpnet-base-v2 model on a custom dataset of 75K+ samples, consisting of all possible matching name combinations within clusters along with randomly generated unmatched name pairs.
• Implemented Triplet Margin Loss with Adam Optimizer for training and performed hyperparameter tuning, identifying an optimal margin value of 0.1 for improved embedding separation.
• Achieved strong model performance with Precision: 0.9906, Recall: 0.9853, F1 Score: 0.9879, and ROC-AUC: 0.9992, with an optimal similarity threshold of 0.694 for match classification.
• Designed an efficient inference pipeline with domain-specific preprocessing including normalization and removal of honorific tokens (late, sh, lt, mr, mrs, etc.) to improve embedding consistency.
• Generated two embeddings per name (preprocessed original name and concatenated variant) and computed similarity using the maximum similarity score across both vectors to improve matching robustness.
• Evaluated data augmentation strategies but determined best performance without augmentation, leading to improved real-world inference accuracy.
• Delivered a high-precision entity resolution solution suitable for applications such as identity matching, deduplication, and record linkage in large-scale datasets.

Tech: mpnet-base-v2, Fine-Tune, Preprocessing, AUC-ROC, Recall/Precision, Hyper-parameter Tuning (Margin) for Triplet Margin Loss, Google Collab (T4)

Semantic Search Engine and Question-Answering over Github DataSets/issues

Developed a Semantic Search Engine to get best n=5 matching issues for user search from Github huggingface DataSets repo using FAISS index, further In-context learning is used to have question answering system over n matching issues for particular user search using T-5 model. App is deployed over Hugging Face Spaces. Further preprocessed data of github hugging_face dataset repo is pushed to huggingface hub.
Tech: Gradio, DataSets, Cosine Similarity, Cls_Pool (model: sentence-transformers/multi-qa-mpnet-base-dot-v1), Hugging Face Hub, T-5 for QA, FAISS.

Designed a Fast Tokenizer from Scratch over WikiText Dataset

Designed a Fast Tokenizer over WikiText dataset of batch size 1000 as Generator, further various tokenization steps such as Normalization (NFD, Lowercase, StripAccents), Pre-tokenization (Whitespace and Punctuation), Model and Trainer (BPE, WordPiece, Unigram), Post Processing ([CLS], [SEP]), Wrapped Tokenzier using PreTrainedTokenizerFast. End to End build Wrapped Tokenizer is further pushed to hub over huggingface models.
Tech: Transformers, WordPiece, WikiText Dataset, Python, HuggingFace hub.

English to French Translator by Fine-Tuning MarianMT and KDE4 Dataset

Developed a English to French Translator by fine-tuning existing model MarianMT by using KDE4 en-to-fr dataset. Fine Tuning using Trainer API, invovles 3 epochs, sacrebleu evaluation metric, split into train and validation in 9:1 with seed 20. Sacrebleu score improved from 38 to 52 approx 14% improvement due to fine-tuning. Further app deployed over huggingface spaces.
Tech: Gradio, Transformers, KDE4 dataset, Sacrebleu Metric, Model : Helsinki-NLP/opus-mt-en-fr, data_collator (Dynamic Padding)

Detoxify Dialogue Summarization by Fine Tuning using Reinforcement Learning (PPO and Reward Model)

Developed end-to-end LLM application to Detoxify Dialogue Summarization using RLHF, which involves using LoRA PEFT Flan-T5 as base model, further Value Head for PPO added additional parameters 768 + 1 (Bias). Facebook Roberta hate speech detection model is used as Reward Model. KL Divergence is used between ref_model and ppo_model to have additional reward. Comparative study of detoxification between ref_model to ppo_model is done by measuring reward score respectively.
Tech: Dataset : knkarthick/dialogsum, LoRA PEFT Flan-T5, Facebook Roberta Hate Speech as Reward Model, KL Divergence, PPO, Amazon SageMaker AI

AI powered Semantic Search engine using RAG for Fashion Store

Developed end to end AI powered Semantic Search engine using RAG for Fashion Store. RAG system involves FAQ’s, and Product information (Technical or Creative accordingly choose Temperature and top_p model params), Cls_pool (model: sentence-transformers/multi-qa-mpnet-base-dot-v1) as Semantic Embedding in Weaviate Db (Hybrid Search using weight, Semantic Search and Keyword Search), Flan-T5 for QA after semantic search , Phoenix and Open-Telemetry for observation and evaluation of project.
Tech: Gradio, Cls_pool (model: sentence-transformers/multi-qa-mpnet-base-dot-v1) as Semantic Embedding in Weaviate Db, Flan-T5, Phoenix and Open-Telemetry

Designed Naive Bayes Classifier from Scratch

Architected a Naive Bayes classifier from scratch for spam email classification using Bayesian probability theory. Applied feature independence assumptions while implementing conditional probability distributions for word occurrences. Integrated Laplace smoothing to ensure all vocabulary elements appear in both spam and non-spam training sets, eliminating P(word|class)=0 scenarios. Enhanced numerical stability through logarithmic transformation of probability scores, preventing misclassification due to floating-point underflow. Delivered a robust classification system achieving 98% accuracy on testing datasets
Tech: Python, Bayesian Probability, Increasing Function (Log)

FakeFinder – AI-Generated Image Detection System (CNN Trained from Scratch)

Developed FakeFinder, an end-to-end deep learning system to distinguish AI-generated images from real images using flexible convolutional neural networks (CNNs) and automated hyperparameter optimization.

• Designed dynamically configurable CNN architectures using nn.Sequential, enabling flexible adjustment of network depth, filter sizes, kernel dimensions, and regularization parameters without hard-coding model structures.
• Constructed a robust hyperparameter search space and implemented a custom Optuna objective function to systematically optimize model performance.
• Executed Optuna studies to automatically explore and identify effective CNN configurations across multiple trials.
• Evaluated models using both predictive performance and efficiency metrics (accuracy, model size, inference time) to support deployment-aware decision making.
• Applied advanced model selection strategies, including:
  • Weighted scoring to balance accuracy and computational efficiency
  • Constraint-based filtering to enforce strict limits on resource usage
• Refined and selected optimal CNN models by analyzing trade-offs between accuracy and performance constraints.

Tech: PyTorch, Flexible CNN Architectures, Optuna (Tree-Structured Parzen Estimator (TPE) is a Bayesian optimization algorithm), Model Selection (Weighted & Constraint-Based), AI-Generated vs Real Images Dataset (subset)

Optuna Hyperparameter Optimization Results

number	acc	batch	dropout_rate	fc_size	k_s_l0	k_s_l1	k_s_l2	lr	n_f_l0	n_f_l1	n_f_l2	n_layers	resolution
15	0.732	16	0.170192	384	5	3	NaN	0.000732	32	24	NaN	2	16
9	0.715	8	0.200294	384	5	3	NaN	0.001072	24	16	NaN	2	16
6	0.715	16	0.134501	384	3	3	3	0.000135	64	56	16	3	16
16	0.692	16	0.151364	256	5	3	NaN	0.000109	32	40	NaN	2	16
1	0.690	16	0.395933	128	3	5	NaN	0.001220	56	16	NaN	2	32

FakeFinder – AI-Generated Image Detection System (Fine-Tune MobileNetV3-Large)

Developed a deep learning–based image classification system to distinguish AI-generated images from real images using a subset of the AI-Generated Images vs Real Images dataset. Implemented an end-to-end training and evaluation pipeline using PyTorch, including dataset loading, image preprocessing, data augmentation, and efficient batching through DataLoaders. Fine-tuned a pre-trained MobileNetV3-Large model by freezing its feature extraction layers and replacing the final classifier head to adapt the network for binary classification. Applied transfer learning techniques to significantly improve model performance with minimal training. Achieved 82% test accuracy with just one training epoch using MobileNetV3-Large, outperforming a CNN model trained from scratch, which achieved ~72% accuracy after three epochs, demonstrating the effectiveness of transfer learning for limited-data scenarios.
Tech: PyTorch, MobileNetV3-Large, Transfer Learning, Image Classification, Data Augmentation, CNNs

Pneumonia Diagnostic AI Assistant — Chest X-Ray Classification

Built an image classification model to distinguish Normal, Bacterial Pneumonia, and Viral Pneumonia using the Chest X-Ray dataset. Implemented a clean training pipeline with PyTorch Lightning, including a LightningDataModule for preprocessing and a LightningModule encapsulating ResNet18 (fine-tuned on the classifier head and last two conv layers), loss, Accuracy metrics, and Adam optimizer. Configured a Lightning Trainer with EarlyStopping (callback), ModelCheckpoint, and LR scheduling (ReduceLROnPlateau) to automate training. Achieved ~90% test accuracy with limited epochs.
Tech: PyTorch Lightning, ResNet18 (partial fine-tuning), AdamW, ReduceLROnPlateau, EarlyStopping

ArcFace-Weaviate : Face Recognition-Based Authentication System

Developed an application enabling registration of new users using facial images and subsequent identification/authentication based on stored faces. Implemented embedding generation using ArcFace, producing 512-dimensional vectors for each registered face. Used Weaviate DB with HNSW (Hierarchical Navigable Small World) graphs to efficiently index and search embeddings using cosine similarity. Authentication logic: if cosine similarity > 60%, the input face is considered a match. Built a user-friendly Gradio interface for image input, registration, and real-time identification. Applied transfer learning for feature extraction, ensuring high accuracy with minimal data. Optimized workflow for fast search and retrieval of embeddings, supporting scalable face-based authentication.
Tech: Python, ArcFace, Gradio, Weaviate, Transfer Learning

AI Powered Visual Search Engine for Fashion Store

Developed a deep learning–based visual search system that retrieves visually similar fashion items. Trained a MobileNetV2 classifier on the clothing-dataset-small (7 apparel classes) using Cross-Entropy Loss, then leveraged the trained backbone in a Siamese Network for metric learning with TripletMarginLoss. Images were resized to 64×64, producing 276-dimensional embeddings optimized with AdamW. The system performs cosine similarity search to return the top-5 nearest items, enabling accurate and efficient visual product discovery. The approach fully utilizes class labels for initial supervised training before embedding-based similarity learning.
Tech: Python, MobileNetV2, Siamese Network, CrossEntropyLoss, TripletMargin Loss, AdamW, Cosine Similarity, DataSet : clothing-dataset-small

CleanVision AI – Edge Model Optimization

Optimized a ResNet-18 image classification model for CPU-only edge devices used in smart city street-cleaning vehicles. Applied pruning, INT8 dynamic quantization, and quantization-aware training (QAT) to meet strict latency and storage constraints. Reduced model size from ~512 MB to <150 MB and CPU inference latency from ~900 ms to <50 ms per image. Maintained >95% classification accuracy across clean, litter, and recycle categories. Enabled real-time, GPU-free deployment on low-resource embedded hardware. Tech: PyTorch, ResNet, Model Pruning, INT8 Quantization, QAT, Edge AI

Prediction of Survival Time and Hazard Risk in Primary Biliary Cirrhosis (PBC)

• Developed survival analysis models to predict patient survival time and hazard risk scores for Primary Biliary Cirrhosis (PBC) using censored clinical data.
• Implemented a Cox Proportional Hazards model, optimizing the negative partial log-likelihood (events-only likelihood) using Newton–Raphson second-order optimization.
• Built a Random Survival Forest to capture non-linear effects and complex feature interactions beyond proportional hazards assumptions.
• Evaluated model performance using the Concordance Index (C-index), achieving 0.85 on Cox model test data and 0.86 on Random Survival Forest validation data, indicating strong risk-ranking accuracy.
• Compared linear and ensemble survival models to balance interpretability and predictive performance.

Tech: Primary Biliary Cirrhosis (PBC) Dataset, Cox Proportional Hazards, Random Survival Forest, Negative Partial Log-Likelihood, Newton–Raphson Optimization, Concordance Index (C-index)

Treatment Effect Analysis of Levamisole–Fluorouracil in Post-Surgical Colon Cancer Patients

• Analyzed randomized controlled trial (RCT) data to estimate the individualized treatment effect of Levamisole–Fluorouracil compared to standard chemotherapy in post-operative colon cancer patients.
• Modeled predictive risk reduction to identify patients who benefit from treatment, defining treatment utility when estimated benefit > 0.
• Implemented Logistic Regression and T-Learner (Random Forest–based) approaches to estimate conditional treatment effects.
• Performed hyperparameter tuning for Random Forest models using Grid Search to optimize treatment effect estimation.
• Evaluated model performance using c-for-benefit, a metric designed for treatment effect discrimination across patient subgroups.
• Achieved c-for-benefit scores of 0.5412 (Logistic Regression) and 0.5250 (T-Learner), indicating modest but consistent ability to rank patients by treatment benefit.
• Interpreted results to validate the clinical benefit of Levamisole–Fluorouracil in reducing residual cancer risk following surgery.

Tech: RCT Data, Random Forest, T_learner, Grid Search, Logistic Regression, Risk Reduction, C-for-benefit

Coursera and DeepLearning.AI

• Retrieval Augmented Generation (RAG) Certificate
• Generative AI with Large Language Models Certificate
• ChatGPT Prompt Engineering for Developers Certificate
• LangChain for LLM Application Development Certificate
• Machine Learning in Production Certificate
• Trustworthy AI: Managing Bias, Ethics, and Accountability Certificate
• Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization Certificate
• Artificial Intelligence: Ethics & Societal Challenges Certificate
• Securing AI and Advanced Topics Certificate
• AI Infrastructure and Operations Fundamentals Certificate
• Human-Centered Artificial Intelligence Certificate
• Unsupervised Learning, Recommenders, Reinforcement Learning Certificate
• Mathematics for Machine Learning and Data Science Specialization Certificate
• Linear Algebra for Machine Learning and Data Science Certificate
• Calculus for Machine Learning and Data Science Certificate
• Probability & Statistics for Machine Learning & Data Science Certificate
• PyTorch: Fundamentals Certificate
• PyTorch: Techniques and Ecosystem Tools Certificate
• PyTorch: Advanced Architectures and Deployment Certificate
• PyTorch for Deep Learning Specialization Certificate
• AI for Medicine Specialization Certificate
• AI for Medical Diagnosis Certificate
• AI for Medical Prognosis Certificate
• AI For Medical Treatment Certificate
• Attention in Transformers: Concepts and Code in PyTorch Certificate
• Introduction to GIS Mapping Certificate
• 10-202: Introduction to Modern AI by Carnegie Mellon University Certificate
• Introduction to Generative AI for Software Development Certificate
• Team Software Engineering with AI Certificate
• AI-Powered Software and System Design Certificate
• Generative AI for Software Development Certificate

Hugging Face Courses

• LLM Course Certificate Certificate
• Hugging Face Agents Course Certificate Certificate
• MCP Course Certificate Certificate

Udemy Courses

• NLP with Python Certificate
• Python Programming Certificate

Non Technical Courses

• The Science of Well-Being Certificate
• Improving Communication Skills Certificate
• AI for Business Certificate
• Developing Innovative Ideas for New Companies: The First Step in Entrepreneurship Certificate
• Innovation for Entrepreneurs: From Idea to Marketplace Certificate
• New Venture Finance: Startup Funding for Entrepreneurs Certificate
• English for Career Development Certificate

NLP and Generative AI Skills

• Flan-T5, Fine-Tuning, Tokenizer, Hugging Face, SpaCy, NLP, NER, FAISS, Regex, Gradio, BERT
• LLM, Generative AI, RAG, Agentic AI, MCP, KL Divergence, Transformers, Weaviate (Vector DB)
• Keyword Search (TF-IDF and BM25), Semantic Search, Hybrid Search (Reciprocal Ranking Fusion : RRF, Weight based Fusion)
• Reinforcement Learning (PPO, GRPO), LoRA: PEFT, MLOps, Phoenix and Open Telemetry
• Smolagents, Ollama, Qwen, Chain-of-Thought Reasoning, ReAct Model, Langchain, LangGraph
• Linear Algebra, Calculus, Probability & Statistics for Machine Learning & Data Science
• Trustworthy AI, Responsible AI, Fairness, Baises, Ethics & Societal Challenges in AI
• Content based Filtering Recommendation System, PCA, Dataset, DataLoader (Pinned Memory), Gradient Accumulation, Transfer Learning, Transforms (Data Augmentation)
• TorchVision, Pytorch Lightning, Profiler, Partial/Full Fine-Tuning, Resnet18, ImageNet, Step LR, Reduce LR on Plateau, and Cosine Annealing LR
• Flexible CNN Architectures, Optuna (Tree-Structured Parzen Estimator (TPE) is a Bayesian optimization algorithm), Model Selection (Weighted & Constraint-Based)
• Siamese Network (Embedding Model, TripletMarginLoss), Resnet (Inputs added to blocks Output), Densenet (Inputs are concat to blocks Output, channel up)
• Encoder, Decoder, Encoder-Decorder NLP, Stable Diffusion, MLFlow, ONNX, Unstructered/Structured/Global Pruning, Static/Dynamic/QAT Quantization
• Cox Proportional Hazards Model, Random Survival Forest, Negative Partial Log-Likelihood, Newton–Raphson Optimization, Concordance Index (C-index)
• Mean Imputation, Regression Imputation, Soft Dice Loss, U-net for Image Segmentation
• T-learner, c-for-benefit, GradCam, Saliency Maps, bioc, Negbio, Permutation Importance and Shapely Values

Backend Skills

• Backend: Flask, Spring MVC, Django, Golang (Echo)
• Encryption (AES, RSA), SSL/TLS, Watermarking, Authentication (Keycloak, JWT, Digital Signature)
• Databases: MySQL, MongoDB, Postgres, Weaviate (Vector DB)
• DevOps: Kubernetes, Docker, Git, GitLab, CI/CD

Published Book

Modern AI Using PyTorch: Building Intelligent Systems with Transformers

• Python programming fundamentals and mathematical foundations
• Linear algebra, tensors, and automatic differentiation
• Neural networks and optimization from scratch
• Transformer architecture and self-attention mechanisms
• Building a LLaMA-style Large Language Model from scratch
• Tokenization using Byte Pair Encoding (BPE)
• End-to-end LLM pretraining
• Supervised Fine-Tuning (SFT)
• Direct Preference Optimization (DPO) for aligning language models with human preferences

Research Papers

Land Cover Classification using Satellite Images and Hybrid MLP-SVM Classifier

Research Paper

Main objective of this project is to take satellite/Hyperspectral images of any area and then perform classification into different classes like – crop classification using Hybrid MLP-SVM Classifier.
Tech: Python, MLP, SVM, Hyperspectral Images, Feature Extraction

Brain Tumor Segmentation and Detection using Ensemble Classifier

Research Paper

Main objective of this project is to take MRI Images of brain, which are further segmented and classifier using ensemble classifier. Various technologies are used like Otsu’s method is used for segmentation, feature extraction is done using PCA+SWT+GLCM, then segmentation is done using KNN+DT+RF ensemble classifier.
Tech: PCA, SWT, GLCM, KNN, Decision Tree, Random Forest, Machine Learning

Real Time Plant Disease Detection and Classification

Research Paper

Research project based on Real Time plant leaf disease identification and classification using various techniques like ANN, K-means segmentation.
Tech: Machine Learning, Image Processing, K-means, Neural Network

Research Reviewer

• Springer Nature
• SAGE Publishing
• Public Library of Science (PLOS)
• FACTA UNIVERSITATIS : Electronics and Energetics
• IAES : International Journal of Artificial Intelligence Certificate
• IAES : International Journal of Informatics and Communication Technology Certificate
• The Institute of Engineers (India) : Series A Certificate

Achievements

• 2× Academic Excellence Awards
• 3× Published Research Papers
• 15× Research Reviewer
• 2x Direct PhD Selection at IIT Ropar and Jodhpur
• Department Rank: 4 | University Rank: 5
• JEE Main: 8123 (231/360) | GATE: 2562
• Ranked 252 in National Science Talent Search Examination
• Selected in Indo-Asian Research Week with Google. Selected among top 50 Students from all the applicants in Computer Vision Track.
• Member of Institution Innovation Council under the ageis of MHRD’s Innovation Cell established at NIT, Kurukshetra for academic year 2018-2019.

Books (Recreational Activities)

• "The Psychology of Money" by Morgan Housel
• "The Power of Now" by Eckhart Tolle
• "Ikigai: The Japanese Secret to a Long and Happy" by Héctor García and Francesc Miralles
• "The Power of Your Subconscious Mind" by Dr. Joseph Murphy
• "Atomic Habits" by James Clear
• "How To Stop Worrying And Start Living" by Dale Carnegie
• "How to Win Friends and Influence People" by Dale Carnegie
• "The 21 Irrefutable Laws of Leadership" by John C. Maxwell
• "Vivekanand Ki Atmakatha: An Autobiography of Vivekananda" by Shri Sankar
• "The 80/20 Principle: The Secret to Achieving More with Less" by Richard Koch
• "The 5 AM Club" by Robin Sharma
• "Who Will Cry When You Die" by Robin Sharma
• "Elon Musk" by Isaacson Walter
• "Range" by Epstein David
• "Ratan Tata a Complete Biography" by A.K. Gandhi,
• "The Thinking Machine: Jensen Huang, Nvidia, And The World’s Most Coveted Microchip" by Stephen Witt

Profile

• Linkedin
• GitHub
• Hugging Face
• Google Scholar
• Publons
• Kaggle

Contact

Email: gargginni01@gmail.com

Evolution of NLP

1. Frequency based Vectors: TF-IDF, BM25 etc, for each sentence one Frequency vector, we donot call it embedding vector as no semantic meaning. Size of Vector is V : Vocabulary Size and Sparse, we cannot compress to small dimension D becoz it donot contain any semantic information.
Remove Stopword/lemmitization because they are noise as embedding depend upon frequency of token

2. Static Embeddings (Semantic but not contextual): Word2Vec, Glove etc, for each token one embedding vector. (Single One Co-occurance Matrix for entire training corpus) As each token represents static embedding, donot have context from other tokens of sentence, so to get Sentence Static Embedding we apply Pooling (Mean/Max/Sum) etc over all computed token embeddings of sentence.
Drawback: Its same for every token irrespective of words-context how and where it is used.
Size of Co-occurance Matrix is V*V : Size of Vocabulary, these sparse vectors are convert for each token from V size to D, Dense Vectors which are used as inputs to MLP.

3. Contextual Embeddings (Semantic and Contextual): RNN/LSTM, for each token one embedding vector. Not use positional encoding because its sequential so now the positions of tokens. (hidden layer provide contextual embedding)
Drawback: Not scalable, sequential, slow to train

4. Contextual Embedding using Attention (Semantic and Contextual):4.Transformers (Scalable, Parallel not Sequential). We stop generating next token when we reached or max_token_limit.
Static Embedding Matrix (E) : E(token_id): Lookup is fast as indexing is fast for integers over string. Token IDs exist only for speed, batching, and GPU efficiency
KV Cache: We cache K,V values for all the previous tokens, then just compute K,V values for new token at a time and reuse cached K,V values for previous token then we apply attention formula. (Speed Up)
Dimension of Static Embedding : V * d_model
Dimension of Positional Embedding : Context_Window * d_model
Static Embedding + Positional Embedding + Segment Embedding == new embedding vector + Attention == Contextual Embedding

Encoder Only (Bidirectional)
1. Sees left and right context.
2. BERT does not do next-token prediction.
3. Predict token : use that token’s embedding.
4. Predict sentence : use [CLS] embedding.
5. Middle token see more context then last token.

Decorder Only (Unidirectional : left to right) / autoregressive
1. The last token embedding has seen all previous tokens, last token itself summary token.
2. Attention mask is left-to-right.
3. No [cls] and [sep] tokens at start and end.
a) "I love machine learning"
b) ['I', 'Ġlove', 'Ġmachine', 'Ġlearning']
c) G’ means space before the word
d) [40, 1842, 4570, 4673]

Contextual Embedding single token is sufficent to sentence summary classification.
Static Embedding Pool of all token embeddings is required to get sentence summary classification.

5. In both Pruning and Quantization we donot change Bias. Pruning focus on weights only, while Quantization focus on weights + Activations both.

Pruning can be done both during inference or at fine-tune/training as well.
Unstructured Pruning: Removes individual weights to zero, makes matrix sparse, matrix shape remain same, so computation remains same, donot give real speedup, reserach purpose used, accuracy is more, at each layer applied.
Structured Pruning: Removes whole neurons / filters / channels, actual speed up, size shrink for effective pruned model, accuracy is less, at each layer applied.
Global Pruning (Prefered One): Unstructured Pruning, applied across multiple layers of model.

Quantization: Quantization Scaling Factors
Static Quantization (Post Training Method): weights (pre-compute) + activations (pre-compute) use Caliber over test_loader, Need of QuantStub and deQuantStub, Ex Computer Vision
Dynamic Quantization (Post Training Method): weights(pre-compute) + activations (dynamic-compute), Ex NLP, LLM
QAT: Quantization Aware Training, form of Static Quantization fuse_model (Conv+BN+ReLU, Conv+BN, Conv+ReLU, Linear+ReLU), Need of QuantStub and deQuantStub

6. Five types of model: Regression, Classification, Embedding, Segmentation.
6.1) Regression: Mean Squared Error Loss Function
6.2) Classification: CrossEntropyLoss
6.3) Embedding (Margin Based): TripletMarginLoss, Contrastive Loss, (Softmax with Additive Angular Margin + CrossEntropy : ArcFace)
6.4) Segmentation: Soft Dice Loss, U-Net, only Convolution Layers
6.5) Medical Prognosis: (Label : Time, Event)
a) Linear Regression (beta@X.T)
b) Negative partial log-likelihood (-log(L(beta))), here L(beta) is likelihood, only events used.
c) Newton–Raphson (second-order optimization)
d) Hazard Ratio, Survival Time, Survival Probability
e) event = death, censored = alive, till study duration
f) c-index, concordant pairs, permissible pairs, risk tie

7. We can use ROC (Receiver Operating Characteristic) to have graph between True Positive Rate, False Positive Rate, then we find closet point to (0,1), so we get threshold beyond which 1 else 0 for output class.
(threshold, (False Positive Rate, True Positive Rate)), this is how we can choose threshold, AUC is area under curve of ROC.

8. Imputations is way to fill missing data say features (Age, BP), BP is missing, how we can fill ??
a) Mean of all BP, and fill all empty BP rows.
b) Regression Imputation, BP = 10*Age + 20 (Linear Regression Model).
c) Drop rows with missing data - Bad idea - donot use it.