Developers, machine learning engineers, and researchers who want deeper control over model behaviour and deployment.
PROGRAM 03 / UNDERSTAND, ADAPT, AND OPERATE MODELS
Neural Architecture & Local Models
Go beneath the API. Understand neural networks, build transformer components, and adapt smaller language models for a defined task.
- PyTorch
- Transformers
- Fine-tuning
- Quantisation & serving
Confident Python, basic linear algebra, and familiarity with machine learning concepts.
A locally served language model, adapted for a bounded use case and accompanied by quality, memory, and latency measurements.
THE CURRICULUM
From foundations
to a working project.
Open a module to see its agenda and practical outcome. The learning path builds across three stages.
STAGE 01
Neural network foundations
01The model as an engineering system
- Define the task, evaluation set, and resource constraints.
- Compare model size, context, latency, and memory needs.
- Build a reproducible experiment environment.
02Tensors and automatic differentiation
- Work with shapes, broadcasting, and matrix operations.
- Follow gradients through a small computation graph.
- Diagnose numerical instability and device errors.
03Training a neural network
- Implement a training and validation loop.
- Compare losses, optimisers, and learning-rate schedules.
- Track overfitting, checkpoints, and reproducibility.
04Text, tokens, and representations
- Prepare a text corpus and inspect tokenisation choices.
- Build embeddings and positional representations.
- Measure sequence length and vocabulary trade-offs.
STAGE 02
Language model architecture
05Attention from first principles
- Implement queries, keys, values, and attention masks.
- Compare single-head and multi-head attention.
- Inspect computation and memory growth.
06Build a compact transformer
- Combine attention, feed-forward layers, and normalisation.
- Implement a small decoder and a next-token objective.
- Inspect activations, training loss, and generated samples.
07Data quality and training experiments
- Inspect duplication, contamination, and unsuitable content.
- Design a bounded training experiment and record its inputs.
- Separate validation evidence from appealing examples.
08Adapt an existing model
- Prepare task examples and a held-out evaluation set.
- Use parameter-efficient adaptation with LoRA.
- Compare adapted output against the original baseline.
STAGE 03
Adaptation & local operation
09Quantisation and memory budgets
- Compare precision formats and compression trade-offs.
- Estimate memory requirements before choosing hardware.
- Measure changes in quality, speed, and resource use.
10Local inference and serving
- Run a local model and expose a controlled interface.
- Manage context, batching, and generation settings.
- Test concurrency, timeouts, and operational limits.
11Evaluation, safety, and maintenance
- Evaluate difficult cases, regressions, and unwanted output.
- Record dataset, model, and configuration versions.
- Plan rollback, access controls, and model replacement.
12Final project: an adapted local model
- Package the model, interface, and reproducible experiment.
- Demonstrate quality and resource measurements.
- Explain when a hosted model or retrieval would be a better choice.
NEURAL ARCHITECTURE / WAITLIST
Your next step
starts here.
Join the waitlist for Neural Architecture & Local Models. We will send the cohort dates, delivery schedule, and enrolment details when they are ready.
AED 9,999 full program fee
This waitlist is only for this program. You can join another program separately.
EXPLORE ANOTHER DIRECTION