From data to decisions: ML, deep learning, and generative AI

Three ways to see it

1. 1

   Machine learning, ML, is the broad family. You take labelled data, run an algorithm over it, and end up with a model that maps inputs to outputs. A spam filter, a credit-scoring engine at HBL, a delivery-time predictor at FoodPanda, a chest-X-ray triage system, all of these are ML. The model can be simple (a logistic regression a statistician would recognise) or very large. ML is a way of producing decisions from data.

2. 2

   Deep learning is a subset of ML where the model has many layers, hence 'deep'. Each layer learns intermediate representations, the next learns over those, and so on. Deep learning unlocks problems where the input is raw and messy: images, audio, free-flowing language. The fingerprint match at NADRA, voice transcription in Punjabi at a courier call centre, Urdu OCR on Sindh land records, these need deep learning because no human can write a rule for what a fingerprint looks like.

3. 3

   Generative AI, GenAI, is a younger branch of deep learning, where the model does not just classify but produces new artefacts. A new paragraph of Urdu policy text, a new picture of Lahore Fort at sunset, a new draft of a sale deed, a new line of code. GenAI is the kind of AI that flooded the public conversation after ChatGPT in late 2022. Behind it sit large language models, image diffusion models, and audio models. The trick is the same as ML: pattern from data. The output is different: a sample, not a label.

The why-tree

Try this with Claude

AI-edge prompt: 'Here is my daily workflow in [your role]. Break it into three buckets: classification, regression, generation. For each, suggest one Pakistan-available tool I could try this week and one risk I should plan for.' Run this in Claude or ChatGPT. Bring the answer to your team meeting.