Learning journey
Artificial Intelligence - Machine Learing - Deep Learning
Exploring the foundations of learning from data, from clear mental models to small practical experiments.
active
Current focus
Generative & Agentic AI
- Topics
- 5
- Notes
- 17
The intention
Why I’m learning this
I’m learning AI/ML to understand how data, assumptions, and evaluation shape the systems we build. The goal is not to treat models as magic, but to develop a practical intuition for when they help and where they can mislead.
Topics
The ideas I’m working through, one connection at a time.
Computer Vision
Exploring how neural networks learn visual representations for image classification, landmark detection, object detection, segmentation, and large-scale image recognition.
Generative & Agentic AI
Exploring efficient LLM fine-tuning, mixture-of-experts architectures, and systems where language models interact with external tools and applications.
Natural Language Processing
Exploring how machine learning models represent, classify, and learn patterns from human language.
Reinforcement Learning
Exploring how agents learn decision-making strategies through rewards, value estimation, Bellman equations, Q-learning, and deep reinforcement learning.
Transformers
Exploring tokenization, self-attention, decoder-only language models, and the architecture behind modern large language models.
Recent learning
Notes from the process of making things click.
LLM Tool Use with MCP and LibreOffice Draw
Exploring agentic AI by allowing a language model to interact with LibreOffice Draw through tools and GUI automation.
DeepSeek-Style Mixture of Experts with SmolLM
Exploring sparse Mixture-of-Experts language-model architecture by incorporating DeepSeek-V3-inspired MoE ideas into SmolLM2-135M.
Q-Learning and Deep Q-Networks
Moving from tabular Q-learning in GridWorld to neural-network-based value approximation with Deep Q-Networks.
Bellman Equation and Value Iteration
Exploring value functions, discounted future rewards, Bellman optimality, and iterative policy improvement.
Fine-Tuning Llama 3.2 3B with QLoRA
Exploring parameter-efficient LLM adaptation by fine-tuning Llama 3.2 3B Instruct with QLoRA on conversational instruction data.
Training SmolLM2-135M
Exploring practical language-model training by training the 135M-parameter SmolLM2 architecture on Cosmopedia v2.
Decoder-Only Transformer from Scratch
Building and training a decoder-only Transformer to understand embeddings, causal self-attention, Transformer blocks, and autoregressive text generation.
BPE Tokenizer from Scratch for Gujarati
Implementing Byte Pair Encoding from scratch to understand how raw Gujarati text becomes tokens consumed by language models.
Training ResNet-50 on ImageNet-1K
Exploring large-scale image classification and residual learning by training ResNet-50 on ImageNet-1K.
Efficient CNNs with Depthwise Separable and Dilated Convolutions
Exploring efficient CNN architectures on CIFAR-10 using depthwise separable and dilated convolutions.
MNIST Classification Under 8K Parameters
Exploring parameter-efficient CNN design by targeting high MNIST accuracy with fewer than 8,000 trainable parameters.
Sentiment Classification with Classical and Neural Models
Comparing classical machine learning, learned embeddings, and pretrained word representations for IMDB sentiment classification.
Kannada MNIST Classification
Exploring handwritten character recognition by applying convolutional neural networks to Kannada digits.
Human Detection and Background Removal
Combining object detection and semantic segmentation to detect people and extract human bodies from image backgrounds.
Facial Keypoints Detection
Exploring facial landmark localization by predicting key coordinates from face images using convolutional neural networks.
Aerial Cactus Identification
Exploring binary image classification by detecting the presence of columnar cactus in aerial imagery.
Super-Convergence and One-Cycle Training on CIFAR-10
Exploring super-convergence, learning-rate selection, and one-cycle training to train convolutional neural networks faster and more effectively on CIFAR-10.