Let's Learn
Ask why an AI denied your loan and the answer is usually “it just did.” XAI exists to change that. Intrinsic vs post-hoc, model-specific vs model-agnostic, local vs global — plus the headline techniques (LIME, SHAP, Grad-CAM, counterfactuals), real deployments, the accuracy-vs-interpretability trade-off, and regulations (GDPR, EU AI Act).
A student who memorises last year’s answer key scores perfectly on the old paper and fails on the new one. Neural networks fall into the same trap — regularisation stops them. Dropout (and its variants), Batch Normalisation, plus the wider landscape (L1/L2, early stopping), and how to combine both.
From reusing a giant pretrained vision or language model on a small custom dataset to adapting GPT-class LLMs for a specific business task. The three types (inductive, transductive, unsupervised), the popular pretrained models (ResNet, ViT, BERT, GPT, CLIP, Whisper, SAM), parameter-efficient techniques (LoRA, adapters, QLoRA), and common pitfalls.
Every time your phone unlocks by looking at your face, a quiet piece of AI has just understood what is in a picture. How images are stored as pixel grids, the historical milestones, the end-to-end workflow, the role of CNNs and Vision Transformers, and the core tasks (classification, detection, segmentation, OCR).
Hold up a photo and a person names what is in it in less than a second. Image classification teaches a computer to do the same. The classification pipeline, the four types, classical ML on HOG/SIFT features, and the deep-learning milestones — LeNet, AlexNet, VGG, ResNet, EfficientNet, Vision Transformers.
Find every object in a photo, draw a tight box around each one, name what it is — all at once, fast enough for real-time video. The hand-crafted era (HOG, SIFT, Haar), one-stage vs two-stage detectors, the headline algorithms (R-CNN, YOLO v1→v8, SSD, RetinaNet, DETR), and where it is heading.
Instead of just naming what is in a photo, image segmentation colours in every single pixel. The three core types (semantic, instance, panoptic), classic techniques (thresholding, watershed), deep-learning models (U-Net, Mask R-CNN, DeepLab, the Segment Anything Model), evaluation metrics (IoU, Dice, mAP), and where pixel-level vision is heading.
A tiny grid of numbers, slid across an image, becomes a network’s eyes. Convolutional filters and the feature maps they produce — the two tools that power every CNN. The convolution operation step by step, stride and padding, classic filters (Sobel, Gaussian, edge), pooling, and visualisation techniques.
One photo of a balloon, twisted into a hundred convincingly different ones — without a new camera or a new dataset. Geometric transforms, colour-space tweaks, filters and random erasing, mixing strategies (Mixup, CutMix), AI-driven and generative augmentation (AutoAugment, RandAugment) — the cheapest, highest-impact lever in computer vision.
Why train from scratch when someone already taught a neural network to see, read, or hear? The two strategies (feature extraction vs fine-tuning), the size-vs-similarity matrix, the headline pre-trained models (ResNet, EfficientNet, BERT, GPT, T5, ViT), step-by-step workflow, and where the field is heading with foundation models.