TFM-Tokenizer · EEG Tokenization

Turning single-channel EEG into time-frequency discrete tokens for a Transformer.

Start from this prompt

Turning single-channel EEG into time-frequency discrete tokens for a Transformer.

Create a 16:9 landscape conceptual method diagram of TFM-Tokenizer, illustrating how single-channel EEG is converted into time–frequency discrete tokens and fed into a Transformer.

Title: "TFM-Tokenizer: EEG Time–Frequency Discretization | Conceptual Example".

The EEG waveform passes sequentially through STFT and overlapping Patching; time-masked input goes into the Temporal Encoder, frequency-band-masked input goes into the Localized Spectral Window Encoder, and the two branches merge into Gated Patchwise Aggregation; this is followed by Quantization to form the Token Vocabulary and EEG-Tokens, with Masked Reconstruction restoring the spectrogram; on the right side, show learned EEG tokens → Masking → Transformer Encoder → Token Prediction Head → Masked Token Prediction.

Use purple for the spectrum, blue for the temporal path, orange for the vocabulary and tokens, gray hatching for masking, on a white background with clear arrows and module boxes.

Keep the qualitative notes that the method is model-agnostic and adaptable to foundation models and multi-channel settings.

Add loss annotation

Annotates the Masked Reconstruction branch with its reconstruction loss, leaving the rest of the pipeline as-is.

Try Kimi Design

Create a 16:9 landscape conceptual method diagram of TFM-Tokenizer, illustrating how single-channel EEG is converted into time–frequency discrete tokens and fed into a Transformer.

Title: "TFM-Tokenizer: EEG Time–Frequency Discretization | Conceptual Example".

The EEG waveform passes sequentially through STFT and overlapping Patching; time-masked input goes into the Temporal Encoder, frequency-band-masked input goes into the Localized Spectral Window Encoder, and the two branches merge into Gated Patchwise Aggregation; this is followed by Quantization to form the Token Vocabulary and EEG-Tokens, with Masked Reconstruction restoring the spectrogram; on the right side, show learned EEG tokens → Masking → Transformer Encoder → Token Prediction Head → Masked Token Prediction.

Use purple for the spectrum, blue for the temporal path, orange for the vocabulary and tokens, gray hatching for masking, on a white background with clear arrows and module boxes.

Keep the qualitative notes that the method is model-agnostic and adaptable to foundation models and multi-channel settings.

Switch palette

Recolors spectrum, temporal path, and tokens to indigo, teal, and vermilion without changing the structure.

Try Kimi Design

Create a 16:9 landscape conceptual method diagram of TFM-Tokenizer, illustrating how single-channel EEG is converted into time–frequency discrete tokens and fed into a Transformer.

Title: "TFM-Tokenizer: EEG Time–Frequency Discretization | Conceptual Example".

The EEG waveform passes sequentially through STFT and overlapping Patching; time-masked input goes into the Temporal Encoder, frequency-band-masked input goes into the Localized Spectral Window Encoder, and the two branches merge into Gated Patchwise Aggregation; this is followed by Quantization to form the Token Vocabulary and EEG-Tokens, with Masked Reconstruction restoring the spectrogram; on the right side, show learned EEG tokens → Masking → Transformer Encoder → Token Prediction Head → Masked Token Prediction.

Use purple for the spectrum, blue for the temporal path, orange for the vocabulary and tokens, gray hatching for masking, on a white background with clear arrows and module boxes.

Keep the qualitative notes that the method is model-agnostic and adaptable to foundation models and multi-channel settings.

Add multi-channel sketch

Adds a small sketch of the multi-channel extension next to the qualitative notes.

Try Kimi Design

Create a 16:9 landscape conceptual method diagram of TFM-Tokenizer, illustrating how single-channel EEG is converted into time–frequency discrete tokens and fed into a Transformer.

Title: "TFM-Tokenizer: EEG Time–Frequency Discretization | Conceptual Example".

The EEG waveform passes sequentially through STFT and overlapping Patching; time-masked input goes into the Temporal Encoder, frequency-band-masked input goes into the Localized Spectral Window Encoder, and the two branches merge into Gated Patchwise Aggregation; this is followed by Quantization to form the Token Vocabulary and EEG-Tokens, with Masked Reconstruction restoring the spectrogram; on the right side, show learned EEG tokens → Masking → Transformer Encoder → Token Prediction Head → Masked Token Prediction.

Use purple for the spectrum, blue for the temporal path, orange for the vocabulary and tokens, gray hatching for masking, on a white background with clear arrows and module boxes.

Keep the qualitative notes that the method is model-agnostic and adaptable to foundation models and multi-channel settings.