A government-issue model zoo for brain waves

What it does

The Army Research Laboratory’s EEGModels is a focused model zoo that implements four published CNN architectures for electroencephalography signal classification. Each network—EEGNet, DeepConvNet, ShallowConvNet, and an SSVEP-tailored EEGNet variant—is exposed as a plain Keras model you parameterize with your own channel and sample counts. The point is to give researchers a validated, citable baseline so they spend less time re-creating paper implementations and more time testing their data.

The interesting bit

Most model zoos sprawl; this one stays rigorously narrow, shipping only peer-reviewed architectures tied to specific ARL papers. That tight scope makes it closer to a living citation than a general-purpose framework. The included explainability hook—using DeepExplain to probe feature relevance—also nods toward interpretability, though the README warns it needs extra coaxing to run under TensorFlow 2.

Key highlights

• Implements EEGNet (original and revised), DeepConvNet, ShallowConvNet, and an SSVEP-specific EEGNet variant.
• Models are plain Keras objects configured by channel count, sample length, and number of classes.
• Includes a reproducibility path for EEGNet feature-relevance analysis via DeepExplain integration.
• Mixed licensing: CC0 1.0 Universal for some portions, Apache 2.0 for others.
• Requires a signed ARL Contributor License Agreement for any external contributions.

Caveats

• The code targets Python 3.7–3.8 and TensorFlow 2.0–2.3, which are now several releases behind mainstream.
• The DeepExplain-based feature-explainability workflow requires extra manual steps to function with TensorFlow 2, per the project’s own issue tracker.
• External contributors must file and return a signed ARL Form 266 CLA before submitting code.

Verdict

Worth bookmarking if you are an EEG or BCI researcher who needs battle-tested baselines without the boilerplate. Skip it if you are looking for a general time-series toolkit or need modern Python/TensorFlow support out of the box.