Cognitive psychology models built with numpy & autograd

• Linear Regression
• Logistic Regression
• Multilayer Classifier
• Autoencoder
• DIVergent Autoencoder (Kurtz, 2007) & Modifications
  • gendiscrim version
  • global average pooling version
  • shallow versions
  • detangler
• Multitasking Network
• ALCOVE (Krushke, 1992)
  • ^ a little off from original model because the similarity function is non-differentiable at zero (i think)
• WARP (Kurtz & Silliman, 2018)

git clone https://github.com/mwetzel7r/autoCog

• numpy
• autograd
• probably scipy eventually
• optional:
  • matplotlib (plotting)

• models typically utilize the same basic functions:
  • forward(...): activate model
  • loss(...): cost function of model activation
  • loss_grad(...): returns gradients to update model weights
• for most models, you update the parameters with the function: -update_params(...)

• keep the modules very light weight and consistent
  • forward, loss, loss_grad functions in every model (that allow them)
  • only use numpy & autograd, and sometimes scipy
• want to include other cognitive models besides neural nets
  • recurrent net
  • covis
  • bayesian things
  • neural turing machines & differentiable computers