Sonar Data & Continual Learning at Exail

Reflections from my internship: the challenges of real-world sonar classification, what domain shift looks like in practice, and what continual learning means outside the textbook.


Where I ended up

For my internship I joined Exail Belgium, a company operating in the defence and maritime sector. Exail builds systems for underwater robotics, autonomous vehicles, and navigation — the kind of infrastructure that runs on sonar and inertial sensors rather than cameras and GPS.

My work sat at the intersection of signal processing and machine learning: taking sonar data and building models that could classify, detect, and extract meaningful features from it. It was a different world from the computer vision and tabular data I'd worked with in academic projects — and that gap turned out to be the most educational part.

What makes sonar data hard

Sonar data is not like image data, even when you render it as a 2D spectrogram. The signals vary significantly depending on the hardware configuration, the water conditions, the depth, the angle of incidence, and the target material. A model trained on data from one environment can perform very differently on data from a slightly different setup — even if the objects being classified are identical.

This is domain shift, and it's a known problem in academic literature. Seeing it in production is something else. The validation metrics on the training distribution looked solid; performance on data from a different deployment context was noticeably worse. The model hadn't learned the underlying structure of the objects — it had learned correlations with the recording conditions.

Pretraining for feature extraction

One of the approaches I worked on was using pretraining as a feature extractor rather than fine-tuning end-to-end. The idea: build feature representations from unlabelled data that are general enough to transfer across different environments, then use those features as input to a smaller task-specific model trained on the labelled data available.

The challenge at Exail was data volume. Getting enough diverse recordings to make the pretraining meaningful required careful curation of what was available, and the domain is narrow enough that true diversity is limited.

Continual learning in practice

The other major thread was continual learning — the problem of updating a model with new data without losing what it learned from old data. In defence applications, a model may be deployed for years and encounter new object types or environments well after its initial training. Retraining from scratch each time is expensive and often not feasible.

In practice, the challenge is that the definition of "what the model needs to remember" is fuzzy. How do you balance plasticity — learning the new — against stability — retaining the old — without a clear benchmark to optimise against? The approaches I explored each came with trade-offs between memory usage, training time, and how well they preserved earlier knowledge.

What the internship actually taught me

The technical lessons were real, but the deeper lesson was about the gap between benchmark AI and production AI. On a benchmark dataset, you control the distribution, the labels, and the evaluation protocol. In a real deployment, none of those are stable. The data changes, the labels are inconsistent, and the evaluation metrics you used during development may not reflect the thing that actually matters to the end user.

Working on sonar data at Exail gave me a clear-eyed view of where machine learning is genuinely useful in hard domains, and where it still requires significant investment in data quality and domain expertise before it becomes reliable. That calibration is something academic projects — however well-designed — can't fully provide.