Hyperparameter Optimization Setup

[DCalhas]

Use Bayesian Optimization Algorithm.

Parameters to have into account:

• Learning rate
• Output Dimension
• Number convolution layers BOLD Encoder (do this after having a generator of layers with consistent input/output) with a desired output
• Number convolution layers EEG Encoder (do this after having a generator of layers with consistent input/output) with a desired output
• Number deconvolution layers Shared Space Decoder (do this after having a generator of layers with consistent input/output) with a desired output
• Regularization - kernel and bias
• Linear Combination loss coefficients (a x contrastive loss + b x recontruction loss) for the Encoder

[DCalhas]

Try to implement this framework: https://github.com/kirthevasank/nasbot

[DCalhas]

Conflicts with fortran compiler, maybe implement this ourselves

[DCalhas]

For now hyperparameters in consideration:

• Learning rate
• regularization kernel and bias
• linear combination coefficients PS. possibly perform a different training procedure where one applies the reconstruction loss to the coder during the first n epochs and in the next n epochs one applies the contrastive loss, and changing so forth

[DCalhas]

https://openreview.net/pdf?id=HkSm8t1PM http://papers.nips.cc/paper/7472-neural-architecture-search-with-bayesian-optimisation-and-optimal-transport.pdf https://github.com/TAMU-VITA/AutoGAN https://github.com/linnanwang/AlphaX-NASBench101

[DCalhas]

Make kernel sizes of convolutions mutable

[DCalhas]

Normalize Bold Inputs

[DCalhas]

Integrate actual optimization of each architecture

[DCalhas]

Something is wrong with the decoder network, for some reason it freezes when processing inputs

[DCalhas]

Situation:

• Transpose Convolutions have a high computation cost, and that was the reason the inputs were "freezing" (taking too long)
• First approach: Perform a MaxPooling to reduce dimension, in order to speed up the transpose convolution. (https://machinelearningmastery.com/upsampling-and-transpose-convolution-layers-for-generative-adversarial-networks/)
• Second: Replace Convolutional Transpose layer by Dense Upsampling and Reshape layers.
• Third approach: Instead of having into account the whole fMRI image, just go a ROI at a time.

https://arxiv.org/pdf/1707.05847.pdf

[DCalhas]

Check what is happening with memory usage: apparently using all the memory available in the GPU server. Reasons:

• Too many pairs of data created - unlikely because with only two instances of individuals the same thing happens.
• Type of networks being loaded in GPU?