Analysis for Figures 1-3:

• Check the Correlation_and_Modulation_Analysis folder for the following demos with expected outputs:
  • PSD_LFP.m: Calculates the power spectral density of LFP (Fig. 1, Supplementary Fig. 1).
  • CFC_Phase_Amp.m: Calculates cross-frequency phase-amplitude coupling of LFP (Fig. 2, Supplementary Fig. 1).
  • Corr_EnvCorr_Env.m: Calculates the correlation of instantaneous amplitude of LFP across 3 regions (Fig. 3, Supplementary Fig. 2).

LightGBM Analysis:

• For non-freezing or freezing data analysis, visit MemoryAge_WinLen_5_Step_0.5/code or MemoryAge_WinLen_5_Step_0.5_FREEZING_ONLY/code. These folders contain files for modeling and analysis notebooks with expected outputs for figure generation (notebooks contain 'Figure' in their names). Follow the STEPS provided.

Transformer Analysis:

• For non-freezing or freezing data analysis, visit MemoryAge_Transformer_xLARGE/code or MemoryAge_Transformer_xLARGE_using_FREEZING_ONLY/code. Key files include:
  • MemoryAge ViT CrossSubject NonFreezing.ipynb: Main file for Transformer training.
  • Using_Attention_to_analyze_Transformer_decision - Comparing Recent Remote in Accurate (or Inaccurate) - BroaderWelch - without Corr - Figure 5_5.ipynb: For attention-guided analysis.
  • Perturbation_results_analysis_Unnormalized - Figure6_2.ipynb: For perturbation analysis.
  • Please also check the analysis in notebooks with the expected outputs (notebooks contain 'Figure' in their names).
• Note: It is recommended not to rerun the notebooks as the outputs are already displayed.

• OS: Windows 10 Professional (Version 22H2)
• Code Editor: VSCode (Version 1.88.1)
• Package Manager: Conda (Version 4.10.3)

• CPU: Intel(R) Core(TM) i5-9400F @ 2.90GHz
• GPU: Nvidia GeForce 2080Ti

• Conda environment installation: Run conda env create -f LFP_TRANS_LOCAL.yml (takes approx. 10 mins).

• LightGBM: Approximately 30 minutes.
• Transformer: Approximately 12 hours (depending on the hardware).

• We utilize the backbone IO strategy from TorchEEG.