Time series · 28 September 2026
When not to use a forecasting foundation model
Chronos and similar models are pretrained on many series. A short, well-specified local model is still the right tool for some problems.
Blog
Notes on vision, language, retrieval, and time series
Short notes on image processing, vision transformers, explainable AI, natural language processing, large language models, retrieval-augmented generation, and time-series foundation models. These are teaching and research notes, not journal papers.
RAG · 22 September 2026
When RAG returns the wrong passage
A language model can quote a retrieved chunk and still miss the claim. Retrieval quality, not prompt tone, decides whether RAG helps.
Image processing · 17 September 2026
Augmentation does not create new patients
Flips, colour jitter, and GANs can regularise a small set. They do not replace an external cohort or a second scanner.
Vision transformers · 12 September 2026
Data leakage in medical image splits: the accuracy you cannot trust
If slices from the same patient appear in train and test, a transformer can look brilliant and still fail on the next person.
Explainable AI · 8 September 2026
Explainable AI for medical scans: a check, not a diagnosis
Saliency maps and attention overlays can show where a model looked. They do not prove that the look was medically valid.
Image processing · 3 September 2026
Class imbalance in medical images: accuracy is the wrong headline
Rare findings dominate clinical risk and vanish in an accuracy score. Sampling, loss design, and the right metric have to be stated together.
Time series · 28 August 2026
Chronos: versions of the time-series foundation model, and how it works
Amazon Chronos treats a series as tokens a language model can read. The original T5 models, Chronos-Bolt, and Chronos-2 differ in speed, context, and whether they can use related series and covariates.
RAG · 24 August 2026
Retrieval-augmented generation: grounding language models in sources you can name
RAG retrieves passages first, then generates. It is a method for reducing unsupported claims, not a guarantee of truth.
LLM · 23 August 2026
Large language models: what they are good at, and what they must not pretend to be
LLMs are strong at drafting and summarising. They are not a substitute for a trained medical-image model or for a cited clinical source.
NLP · 22 August 2026
Natural language processing beyond the chat window
NLP is tokenisation, representation, and sequence modelling. Those steps still matter when the text is a report, a paper, or a clinical note.
Vision transformers · 21 August 2026
Vision transformers for medical images: patches, attention, and when CNNs still help
ViTs model long-range context well, but medical scans are noisy and small. Hybrid CNN–transformer designs are often the practical choice.
Image processing · 20 August 2026
Why image processing still matters before a deep model sees the scan
Noise, contrast, and colour shifts in medical images can mislead a classifier. Careful preprocessing is still part of a trustworthy pipeline.