CampusLearn
← BIT 402 Artificial Intelligence in Business

Neural Networks, NLP and Computer Vision

Notes

Neural network foundations A neural network is a trainable mathematical model formed from connected layers of computational units. Each unit combines weighted inputs and applies a transformation, allowing deeper networks to learn complex patterns. Training adjusts weights to reduce a loss function, which measures disagreement between predictions and targets. Gradient-based optimisation estimates adjustments that reduce this loss. Large models often require substantial data and computing resources, so transfer learning can adapt a previously trained model to a narrower business task. A simpler baseline remains valuable for judging whether extra complexity produces worthwhile improvement. Language and visual applications Natural language processing, or NLP, analyses and generates human language for tasks such as document classification and customer support. Tokenisation divides text into processing units, while embeddings represent words or passages as numerical vectors. Computer vision extracts information from images, including object detection and document recognition. Optical character recognition converts visible text into machine-readable characters. Both language and vision systems can fail when local vocabulary, lighting or document layouts differ from training examples. Evaluate performance on representative Kenyan inputs and route uncertain cases to human reviewers rather than treating every output as reliable. Practical example A Nakuru agribusiness cooperative uses image classification to identify damaged produce. On 200 photographed crates, staff confirm that 50 contain damage. The model identifies 45 of these and incorrectly flags 15 undamaged crates. Its precision is 45 ÷ 60 = 75%, and recall is 45 ÷ 50 = 90%. Reviewing the 60 flagged crates at two minutes each requires 120 minutes, compared with 400 minutes for manual review of all 200. The cooperative still samples unflagged crates because five damaged crates were missed, and it checks whether evening lighting increases errors.

Quiz

1. Which task is optical character recognition?

2. What does transfer learning offer a business with limited training data?

3. Why sample unflagged crates in the example?

Sign in to submit and securely grade your quiz.