# MI-020: Learning Temporal Information for Brain-Computer Interface Using Convolutional Neural Networks

> Generated from a local paper card. Do not edit this page directly; edit the source card and rerun `vp run docs:generate`.

> Internal wiki boundary: paper-card evidence and user-provided PDF downloads are available for private reading. Extracted full-paper text is not published.

> Source: `library/paper_cards/MI-020.md`

## Paper Access

* Internal PDF: <a href={"/papers/MI-020.pdf"} download style={{ display: "inline-flex", alignItems: "center", justifyContent: "center", minHeight: "2.25rem", padding: "0.45rem 0.8rem", borderRadius: "6px", backgroundColor: "#047857", color: "#ffffff", fontWeight: 700, lineHeight: 1, textDecoration: "none", boxShadow: "0 1px 2px rgba(15, 23, 42, 0.22)" }}>Download Paper</a>
* DOI / official page: [10.1109/TNNLS.2018.2789927](https://doi.org/10.1109/TNNLS.2018.2789927)
* Deployment boundary: these PDF links are intended only for a private/protected internal wiki.

## MI-020: Learning Temporal Information for Brain-Computer Interface Using Convolutional Neural Networks

## Metadata

* ID: MI-020
* Title: Learning Temporal Information for Brain-Computer Interface Using Convolutional Neural Networks
* Year: 2018
* DOI / URL: 10.1109/TNNLS.2018.2789927
* Local PDF: see Paper Access section above
* Text artifact: local-only path withheld from docs site
* Review status: `extracted`

## Study Type

* Track: MI
* Task: CNN classification using temporal representation derived from FBCSP-style processing
* Participants or dataset: BCI Competition IV Dataset 2a with nine subjects
* Device/electrode setup: BCI Competition IV MI EEG dataset; channel details follow Dataset 2a
* Protocol/task: four-class MI classification

## Methods

* Signal processing or analysis: modified FBCSP representation plus CNN architecture
* Comparator: CSP/FBCSP and other MI classifiers
* Online/offline: offline benchmark evaluation

## Key Results

* The abstract reports improved average subject accuracy.
* The paper contributes a transition from hand-engineered FBCSP features to CNN-based temporal representation learning.

## Limitations

* Offline dataset experiment.
* Does not evaluate no-control false activation or robot command safety.
* Exact improvement values should be extracted from result tables before quantitative citation.

## Relevance To Current Review

* Useful deep-learning bridge between FBCSP and EEGNet-like compact neural decoders.
* Supports testing whether temporal information improves MI mode-control classification.

## Evidence Status

| Claim | Status | Evidence Note |
| --- | --- | --- |
| CNNs have been applied to MI EEG using FBCSP-derived temporal representations. | verified | Abstract and methods describe the representation and CNN classifier. |
| Dataset 2a remains a standard MI benchmark for deep models. | verified | Local text describes nine subjects and four-class MI data. |
| This method validates online stop/cancel control. | needs confirmation | Only offline benchmark evidence is present. |

## Open Questions

* Should this be a baseline or only trend evidence given the added complexity?
* Does the project have enough MI trials to train a CNN reliably?