Computer Vision and Natural Language Processing
What this domain tests
Computer Vision and Natural Language Processing tests how machines turn pixels and text into usable representations. Expect pipeline ordering, basic convolution arithmetic, and the difference between classification, detection and segmentation, plus NLP preprocessing and evaluation.
Core rules and concepts
- A digital image is a grid of pixel values; colour images have channels such as red, green and blue.
- Convolution slides a small kernel over the image, multiplying and summing local values to produce feature maps. Output size for one dimension is (W − K)/S + 1, with padding added to W where used.
- Vision tasks: classification names the main object; detection locates and labels objects with bounding boxes; segmentation labels pixels or regions.
- NLP pipelines tokenize text, normalize it, remove noise where appropriate, convert tokens to vectors, then model sequences or context.
- Bag-of-words ignores order; embeddings capture similarity; transformers use attention to weigh context.
- Evaluate with task-appropriate metrics: precision, recall and F1 for classification; intersection-over-union for detection; human or task scores for generation.
Worked example
A 6×6 image is convolved with a 3×3 kernel, stride 1, no padding. Output width = (6 − 3)/1 + 1 = 4, so the feature map is 4×4. Each output cell is the sum of the nine products between the kernel and the local patch. The kernel therefore acts as a learned or hand-designed feature detector, such as an edge finder.
Traps that cost marks
Do not forget the stride term or add padding incorrectly. Do not apply augmentation to validation or test data. Do not assume tokenization is trivial across languages. Do not treat fluent generated text as automatically factual.