A special thanks and credit to Zach Carmichael for providing many components of the core code of this project

• No work comparing reaction times CYBORG (i.e. what works best?)

• What improves model performance most? (reaction time, or CYBORG)?
• Can reaction time help CYBORG improve?
• What is the best way to use reaction time with deep learning models?
  • In the loss term?
  • In a regularization term?

CYBORG: CYBORG: Blending Human Saliency Into the Loss Improves Deep Learning ( WACV 2023)

Reaction times: Measuring Human Perception to Improve Handwritten Document Transcription (TPAMI 2021)

You can install this in a pip environment

python3 venv -m env
source env/bin/activate
pip3 install -r requirements.txt

You can run (and optionally set the log level) via the following:

CYBORG_SAL_LOG_LEVEL=INFO ./main.py ...

Run the following for help:

./main.py -h

General options:

./main.py \
  -B DenseNet121 \
  ... \
  --epochs 2 \
  --gpus 1 \
  --quick-test \
  --batch-size 64 \
  --hparam-tune \
  --stochastic-weight-averaging

CYBORG:

./main.py \
  -B DenseNet121 \
  -L CYBORG \
  -T original_data \
  --cyborg-loss-alpha 0.5

CYBORG+REACTIONTIME:

./main.py \
  -B DenseNet121 \
  -L CYBORG+REACTIONTIME \
  -T original_data \
  --cyborg-loss-alpha 0.5

You can also run this with WandB. Most of this follows their PyTorch-Lightning setup.

You can track your experiments by passing the flag --use-wandb-logger true to your run script.