Master10
Artificial Intelligence Module

Computer Vision & Natural Language Processing

Computer Vision (CV) and Natural Language Processing (NLP) constitute the two primary perceptual and cognitive pillars of applied artificial intelligence. Computer vision tasks range from image classification and semantic segmentation to real-time object detection models like YOLO (You Only Look Once) and Mask R-CNN. NLP processes human linguistic data through tokenization, part-of-speech tagging, named entity recognition (NER), dependency parsing, and semantic sentiment analysis. Vector word representations—such as Word2Vec (Mikolov et al., 2013) and GloVe—map words into dense geometric vector spaces where semantic relationships correspond to vector arithmetic. Automatic Speech Recognition (ASR) converts acoustic audio waveforms into textual transcriptions using acoustic transformer models.

Key Concepts & Examination Highlights

  • Word2Vec, introduced by Tomas Mikolov at Google in 2013, uses continuous bag-of-words and skip-gram models to generate semantic word embeddings.
  • YOLO (You Only Look Once) is an influential real-time object detection architecture that treats detection as a single regression problem.
  • Named Entity Recognition (NER) is an NLP technique that identifies and classifies key elements in text into predefined categories like names, dates, and locations.
  • YOLO (You Only Look Once), introduced by Joseph Redmon et al. in 2016, is a single-stage object detection framework that predicts bounding boxes and class probabilities directly from full images in real time.
  • Word embeddings, such as Word2Vec (developed by Tomas Mikolov in 2013) and GloVe, represent words as continuous dense vectors in a geometric space where semantically similar words are positioned closer together.
  • Semantic segmentation assigns a categorical label to every individual pixel in an image, whereas instance segmentation distinguishes between distinct individual objects of the same class.
  • Vision Transformers (ViT), introduced in 2020, apply standard Transformer architectures directly to sequences of non-overlapping image patches for visual recognition tasks without convolutions.
  • Lemmatization reduces a word to its canonical base or dictionary form (lemma) using vocabulary and morphological analysis, unlike stemming, which crudely chops off word affixes.
  • Sentiment analysis determines the emotional valence or polarity (positive, negative, or neutral) expressed in natural language texts using lexicon-based or deep learning classifiers.
  • Named Entity Recognition (NER) is an information extraction subtask that locates and classifies named entities in unstructured text into predefined categories such as person names, organizations, and locations.
  • The BLEU (Bilingual Evaluation Understudy) score and ROUGE metric are standard automated benchmark scores used to evaluate the quality of machine translation and text summarization systems.
  • Feature extraction in classical computer vision relied on handcrafted algorithms such as SIFT (Scale-Invariant Feature Transform), SURF, and HOG (Histogram of Oriented Gradients) before deep learning.
  • Optical Character Recognition (OCR) converts printed or handwritten text images into machine-encoded text, utilizing CNN-based feature extractors and CTC (Connectionist Temporal Classification) decoding.
  • Generative Adversarial Networks (GANs) and diffusion models are widely applied in computer vision for image super-resolution, style transfer, and inpainting missing regions in photographs.
  • Fast R-CNN and Faster R-CNN are two-stage object detection architectures that use a Region Proposal Network (RPN) to generate candidate bounding box regions before classification.
  • Facial recognition systems use deep convolutional networks to map facial images into compact latent embeddings (such as FaceNet embeddings), comparing Euclidean distances for identity verification.
  • Edge detection algorithms, such as the Canny edge detector and Sobel filter, identify sharp discontinuities in pixel brightness by calculating spatial image gradients.
  • Part-of-Speech (POS) tagging is the NLP process of assigning grammatical categories (such as noun, verb, adjective) to words in a sentence based on lexical context and syntactic rules.
  • Stop words are frequently occurring grammatical words (such as 'the', 'is', 'at') that are often filtered out during preliminary text preprocessing to reduce dimensionality in search indexing.
  • Term Frequency-Inverse Document Frequency (TF-IDF) quantifies the relative statistical importance of a word within a specific document compared to an entire text corpus.
  • N-grams are continuous sequences of nn items from a given text sample, where n=1n=1 is a unigram, n=2n=2 is a bigram, and n=3n=3 is a trigram, used in statistical language modeling.
  • Dependency parsing in NLP analyzes the grammatical syntax of a sentence by establishing directed dependency relationships between head words and their dependent modifiers.
  • Text summarization is categorized into extractive summarization, which selects representative sentences directly from the original text, and abstractive summarization, which generates novel rephrased sentences.
  • Machine Translation evolved from Rule-Based Machine Translation (RBMT) and Statistical Machine Translation (SMT) to Neural Machine Translation (NMT) powered by sequence-to-sequence Transformer models.
  • Bhashini (National Language Translation Mission), launched by the Government of India, is an AI-led language platform aimed at breaking language barriers across 22 scheduled Indian languages.
  • AlexNet, which won ImageNet in 2012, utilized an 8-layer architecture with 60 million parameters, demonstrating the power of GPU-accelerated convolutional neural networks.
  • VGGNet (VGG-16 and VGG-19), introduced by Simonyan and Zisserman in 2014, proved that stacking small 3x3 convolution filters throughout deep networks improves visual feature representation.
  • ResNet-50 introduced identity residual shortcut mappings, allowing gradients to propagate directly through skip connections and enabling training of networks with 50 to 152 layers.
  • MobileNet utilizes depthwise separable convolutions to reduce model size and latency, enabling high-performance computer vision inference on mobile and embedded edge hardware.
  • U-Net, designed by Olaf Ronneberger et al. in 2015, features an encoder-decoder contractive and expansive path with cross-layer skip connections, serving as the gold standard for biomedical image segmentation.
  • Mask R-CNN extends Faster R-CNN by adding a parallel branch for predicting pixel-level segmentation masks alongside bounding box coordinates and class labels for instance segmentation.
  • Non-Maximum Suppression (NMS) is a post-processing algorithm in object detection that filters out redundant overlapping bounding boxes based on Intersection over Union (IoU) scores.
  • Intersection over Union (IoU), or the Jaccard Index, measures the spatial overlap between predicted bounding boxes and ground truth boxes (IoU=Area of OverlapArea of Union\text{IoU} = \frac{\text{Area of Overlap}}{\text{Area of Union}}).
  • Mean Average Precision (mAP) is the standard benchmark evaluation metric for object detection, calculated by averaging the Average Precision (AP) across all object classes at specified IoU thresholds.
  • Anchor boxes in object detectors (like Faster R-CNN and early YOLO versions) are predefined bounding boxes with various aspect ratios and scales used as reference templates for object localization.
  • Keypoint detection locates semantic body landmarks and joint coordinates (such as elbows, knees, eyes) on human bodies for real-time human pose estimation.
  • Optical flow algorithms (such as Lucas-Kanade and Farnebäck) compute the pattern of apparent motion of image objects between two consecutive video frames.
  • Structure from Motion (SfM) is a photogrammetric range imaging technique that reconstructs a 3D point cloud scene geometry from a series of overlapping 2D photographs.
  • Neural Radiance Fields (NeRF), introduced by Mildenhall et al. in 2020, represent complex 3D scenes as continuous 5D neural functions (x,y,z,θ,ϕx, y, z, \theta, \phi) to render photorealistic novel viewpoints.
  • 3D Gaussian Splatting is a real-time radiance field rendering technique that represents 3D scenes using millions of 3D Gaussian ellipsoids, achieving photorealistic rendering at over 100 FPS.
  • Global Vectors for Word Representation (GloVe), developed at Stanford in 2014, constructs dense word embeddings by factorizing the global word co-occurrence matrix of a text corpus.
  • FastText, developed by Facebook AI Research in 2016, extends Word2Vec by treating each word as a bag of character nn-grams, allowing it to generate embeddings for out-of-vocabulary (OOV) words.
  • ELMo (Embeddings from Language Models), introduced by Peters et al. in 2018, generated the first deep contextualized word representations using bidirectional LSTM networks.
  • Coreference resolution in NLP is the task of identifying all mentions and pronouns in a text that refer to the same real-world entity (e.g., linking 'Marie Curie' with 'she' and 'the scientist').
  • Semantic role labeling (SRL) identifies the predicate-argument structure of a sentence, determining 'who did what to whom, when, and where' for information extraction.
  • Morphological analysis decomposes complex words into their constituent root morphemes, prefixes, and suffixes, essential for agglutinative languages.
  • Word Sense Disambiguation (WSD) determines the correct semantic meaning of a polysemous word in context (e.g., distinguishing 'apple' the fruit from 'Apple' the corporation).
  • Information Extraction (IE) transforms unstructured text into structured relational database entries using named entity recognition, relation extraction, and event detection.
  • Text classification algorithms categorize documents into predefined thematic classes, utilized in spam filtering, legal discovery, customer ticket routing, and sentiment analysis.
  • Cross-lingual language models (such as XLM-RoBERTa and mBERT) are pre-trained on text from over 100 languages, enabling zero-shot cross-lingual transfer learning across global languages.
Curriculum & Reference Sources: Association for Computational Linguistics (ACL), IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Sample Solved Questions & Concept Explanations

8 Verified Concept Questions
Q1.EASY

What type of neural network specializes in computer vision tasks like image recognition, classification, and object detection?

Q2.EASY

What does 'NLP' stand for in artificial intelligence and computational linguistics?

Q3.MEDIUM

What architecture did ResNet (Deep Residual Learning) introduce in 2015 to enable training of networks with over 150 layers?

Q4.MEDIUM

In natural language processing, what are 'Word Embeddings' (such as Word2Vec or GloVe)?

Q5.MEDIUM

What is 'Beam Search' in sequence generation and machine translation decoding?

Q6.EASY

What does the subfield acronym "NLP" stand for in computer science and artificial intelligence?

Q7.EASY

What does the optical character recognition acronym "OCR" stand for in computer vision?

Q8.EASY

What is "Tokenization" in natural language processing (NLP)?