Try Boost
Accelerate your research with up to 20 hours of computing
Contact Us
Williams et al. |

EEG-GAN

16
1
Image Analysis
    More
Step 1: Upload your data

EEG Training Data

Drag your file(s) or upload
  • Your file can be in the following formats:csv, txt, tsv
    • Files must be in .csv, .txt, or .tsv format
or
Don’t have a file?
Use our demo data to run
Use Demo Data
View example data

EEG Testing Data

Drag your file(s) or upload
  • Your file can be in the following formats:csv, txt, tsv
    • Files must be in .csv, .txt, or .tsv format
or
Don’t have a file?
Use our demo data to run
Use Demo Data
View example data
Step 2: Set Parameters
Fine-tune pretrained model
Min: 1
Max: 10000
100
1000
50000
100
1000
5000
Step 3: Complete run profile
GANs are machine learning frameworks that consist of two adversarial neural network agents, namely the generator and the discriminator. The generator is trained to create novel samples that are indiscernible from real samples. In the current context, the generator produces realistic continuous EEG activity, conditioned on a set of experimental variables, which contain underlying neural features representative of the outcomes being classified. For example, depression manifests as increased alpha oscillatory activity in the EEG signal, and thus, an ideal generator would produce continuous EEG that includes these alpha signatures. In contrast to the generator, the discriminator determines whether a given sample is real or synthetically produced by the generator. The core insight of GANs is that the generator can effectively learn from the discriminator. Specifically, the generator will consecutively produce more realistic synthetic samples with the goal of “fooling” the discriminator into believing them as real. Once it has achieved realistic samples that the discriminator cannot discern, it can be used to generate synthetic data—or in this context, synthetic EEG data.
Limitations: How realistic synthetic data may be can vary. Artifacts or patterns found in the training data may be reflected in the synthetic data. More training records, and a larger number of training epochs generally will improve performance. Further analysis is recommended as the classification method used here provides some metrics, but does not fully address how realistic the synthetic data may be. The classifiability of datasets might vary greatly between datasets, particularly when data is imbalanced, for example if there is a 9:1 ratio between two classes.
 
Use cases: Synthetic data can be used to augment downstream analysis. For example, a classification model trained on 1000 synthetic and 100 original records may outperform a model trained only on the 100 original records. Synthetic data can also be used for testing purposes, or for protecting sensitive data.
 
Tasks: Three tasks are available: (i) synthesize, (ii) training, or (iii) finetune. Select synthesize if you only want to use the pretrained GAN to generate synthesised data, select training to train a brand new model, and select finetune to fine tune the pretrained GAN. Note that the train_data and n_epochs parameters are not used if synthesize is selected, but the default settings can be used.
Citation:
Augmenting EEG with Generative Adversarial Networks Enhances Brain Decoding Across Classifiers and Sample Sizes Williams, Weinhardt, Wirzberger, & Musslick (2023)
Released:
Aug-18-2023
Previous Job Parameters
Your previous job parameters will show up here
so you can keep track of your jobs
Results
Parameters