Courses / HEC NSCT – National Skill Competency Test: Complete Preparation for Computing Graduates
8.5 Deep Learning, NLP, Computer Vision and Responsible AI
Area 8 – AI, Machine Learning & Data Analytics (10%) · 50 min read · 14 MCQs to check yourself
Neural networks
Neurons in input, hidden and output layers; each computes a weighted sum + bias, then an activation function (ReLU in hidden layers; sigmoid for binary output; softmax for multi-class probabilities). Training = forward pass → loss → backpropagation → gradient descent updates weights; learning rate controls step size; an epoch = one pass over all training data; batch = subset per update.
| Architecture | Best for |
|---|---|
| CNN (convolution + pooling) | Images, computer vision |
| RNN / LSTM | Sequences, time series, older NLP |
| Transformer (self-attention) | Modern NLP and LLMs (BERT, GPT) |
| GAN | Generating images |
| Autoencoder | Compression, anomaly detection |
| Regularize with dropout; use transfer learning (fine-tune a pretrained model) when data is limited. |
NLP
Pipeline: tokenization → lowercasing → stop-word removal → stemming/lemmatization → features (bag of words, TF-IDF, word embeddings such as Word2Vec) → model. Tasks: sentiment analysis, named-entity recognition, translation, chatbots. Large Language Models (GPT, Gemini, Claude) are transformers trained on huge text; risks: hallucination, bias.
Computer vision
Image classification, object detection (YOLO), segmentation, face recognition, OCR. Images are arrays of pixels (height × width × 3 channels).
Big data and MLOps
Big data 5 Vs: Volume, Velocity, Variety, Veracity, Value. Tools: Hadoop (HDFS + MapReduce), Spark (in-memory). MLOps: version data and models, deploy as an API, monitor data drift, retrain.
Responsible AI
Fairness and bias (biased training data → unfair decisions), transparency/explainability (SHAP, LIME), privacy (anonymisation, consent), accountability, security (adversarial examples, data poisoning), deepfakes. Pakistan's draft National AI Policy stresses ethical, inclusive AI.
But what is a neural network? (3Blue1Brown)
This lesson ends with a quiz of 14 MCQs. Open the lesson on PrepOasis to take it and save your progress.
Previous: 8.4 Unsupervised Learning · Next: Area 8 Review, Checklist and Area Test