- Home
- Technology
- Deep Learning
Discipline
Neural networks, and the discipline they demand.
Deep learning is the right tool for perception, language and signal — and an expensive way to underperform gradient-boosted trees on a spreadsheet.
01 / What this covers
Where neural methods genuinely win.
Deep learning earns its cost where the input is unstructured and the relationships are not something a human can enumerate: images, audio, language, sequences and signals. In those domains nothing else is close.
On structured tabular data — the majority of business problems — it is usually slower to train, harder to explain, more expensive to serve and no more accurate than a gradient-boosted tree. We use it where it wins and say so where it does not.
02 / In practice
What this looks like in a build.
- Transfer over training from scratch
- Almost always the right starting point. Pre-trained models fine-tuned on your data reach a better result with a fraction of the data, compute and time — and training from scratch is justified rarely and explicitly.
- Fine-tuning strategy
- Full fine-tuning, parameter-efficient methods such as LoRA, or careful prompting with retrieval — chosen on data volume, how far your domain sits from the base model, and what you can afford to serve.
- Retrieval-augmented generation
- For language tasks over your own documents, retrieval usually beats fine-tuning: it is cheaper, updates instantly when the documents change, and produces answers you can trace to a source.
- Serving economics
- Quantisation, batching, distillation and caching to bring inference cost and latency into the budget. A model that is 2% better and four times the cost per call is frequently the wrong trade, and that is a decision to take with numbers.
- Compute discipline
- Training runs are planned, budgeted and checkpointed. GPU spend is one of the easiest costs in this field to lose control of, and it is almost always lost through unmonitored experimentation rather than a single decision.
03 / Honest limits
Where we would tell you not to.
Deep models are difficult to explain in the way a regulator or an adverse-decision notice requires. Where that is a hard constraint, it narrows the architecture choice considerably, and that constraint belongs in the scoping conversation rather than the launch review.
They are also data-hungry and confidently wrong outside their training distribution. Neither property is fixable by tuning; both have to be designed around with abstention thresholds and human review.
04 / Related
The rest of the method.
These disciplines are applied together far more often than separately. Most engagements draw on all three.
Data Science
Experiment design, causal inference and the statistics behind the claims.
Robotics & Autonomy
Perception, planning and control for machines that move.
Technology overview
The engineering standards that apply across every discipline — evaluation, reproducibility, failure behaviour and cost control.
Next step
Have a technical question about this?
Technical enquiries reach an engineer rather than a sales inbox. If you want to interrogate the approach before committing to anything, that is a conversation we would rather have early.