Machine-learning types, applications and hardware
Classify the learning approach from the task, then choose hardware from a standard laptop to specialized and cloud/HPC infrastructure.
What you need to be able to do
- describe deep learning, reinforcement learning, supervised learning, transfer learning and unsupervised learning;
- connect approaches to market basket analysis, medical imaging diagnostics, natural language processing, object detection/classification, robotics navigation and sentiment analysis;
- describe processing, storage and scalability requirements;
- recognize application-specific integrated circuits (ASICs), edge devices, field-programmable gate arrays (FPGAs), graphics processing units (GPUs), tensor processing units (TPUs), cloud-based platforms and high-performance computing (HPC) centres.
Rapid Recall Deck
Attempt each card first. If a term is new, flip and study it, then answer it again without looking. Mark only secure recall as “Know it”.
Classify the learning signal
Four tasks
Choose the most relevant approach: A sentiment from labelled reviews; B customer segments from purchasing features without labels; C robot navigation learned from rewards; D adapt an existing image model to a small related labelled dataset.
A supervised learning. B unsupervised learning. C reinforcement learning. D transfer learning. Deep learning can be used inside several approaches when multi-layer neural networks are the model.
Use the approach in context
Medical imaging
Describe a supervised-learning setup for medical imaging diagnostics.
Use labelled medical images as training inputs and diagnostic categories as target labels; the model learns patterns that support predictions for unseen scans.
Market basket / NLP
Give one plausible use of ML for market basket analysis and one for natural language processing.
Market basket analysis can discover purchasing relationships or customer groups. NLP can classify sentiment, categorize text or learn patterns in language data.
Choose hardware from constraints
Three deployments
A: small 5,000-row experiment. B: phone camera classifier requiring low-latency local inference. C: very large image-model training on millions of examples.
A standard laptop can suit A. B needs efficient edge processing, potentially using specialized mobile/ASIC-style acceleration. C is suited to powerful GPUs/TPUs or other accelerators with substantial storage and scalable cloud/HPC resources. The exact choice balances processing, storage, scalability, latency, power and cost.
Advanced infrastructure must all be recognizable
GPU / TPU
Parallel or tensor-specialized acceleration for suitable ML workloads.
FPGA / ASIC
FPGA favours reconfigurability; ASIC favours fixed specialized implementation/efficiency.
Edge / cloud / HPC
Edge reduces latency/network dependence; cloud/HPC supplies scalable remote compute and storage.
Diffusion creates a compute problem
Explain the hardware connection
Visionary Studios uses iterative diffusion image generation. Explain why this workload can benefit from GPU-style parallel processing, large memory/storage and scalable cloud or HPC resources, and state one latency, cost or power trade-off.
Repeated neural-network denoising involves many tensor operations suited to GPU/accelerator parallelism. Large models and image datasets need substantial memory/storage. Cloud or HPC resources can scale training/generation, but Visionary Studios must weigh output quality against latency, infrastructure cost and energy/resource use.
Generation methods and Paper 1 transfer
Conditional vs unconditional
Distinguish them and explain where text-to-image, segmentation maps and image-to-image translation fit.
Conditional generation is guided by supplied information such as text, labels, sketches or segmentation maps; text-to-image and image-to-image translation are conditional examples. Unconditional generation has no predefined conditioning input and generates from patterns learned from training data.
Paper 1 mini-set
1. Describe one reason a GPU suits CNN training. 2. Distinguish reinforcement and supervised learning. 3. Suggest hardware for very large image-model training and justify.
GPU: high-throughput parallel tensor computation. Supervised: labelled examples; RL: state-action interaction and rewards. Large training: GPU/TPU-equipped cloud/HPC with large processing, memory/storage and scalability.
Finish the learning cycle
Case study / exam preparation
Add one research note on the computational efficiency of generative-image systems.
Paper 1 — short practice
Reproduce the five ML approaches and seven advanced hardware categories from memory.
Case-study checkpoint
Connect conditional, unconditional, text-to-image and image-to-image generation to Visionary Studios.