Associative Learning with Artificial Neural Network (ALANN)

03/04/2024

Summary

This is an ongoing project led by Prof. Andrew Delamater in collaboration with Dr. Santiago Castiello de Obeso. The aim of this project is to understand the "microstructure of cognition" (Rumelhart, & McClelland, 1988).

Instructions

This project is programmed in R. Thus, the user will need to install the latest versions of R programming language and R Studio.

1. Open ALANN.Rproj R project

2. Open main.R within the project

3. Prepare the training.csv file. The columns are:

4. Adjust the main.R script:

4.1. Decide which model you want to run in line 53.

mod0: Rescorla-Wagner (RW); Rescorla & Wagner (1972)

mod1: Back-Propagation (BP); Delamater (2012) - Learning & Behaviour

mod2: BP dynamic alpha; Delamater & Castiello (2022) - gregynog Associative Learning Symposoum

mod3: BP; Xie & Seung (2003) - Neural Computation

mod4: Contrast Hebbian Learning (CHL); Xie & Seung (2003) - Neural Computation

mod5: CHL with random feedback; Detorakis, et al. (2019) - Neural Networks

mod6: CHL with random feedback and dynamic LR (mod2 and mod5)

4.2. Adjust parameters depending on the model (modType) in lines 63 to 98.

5. Visualize Figures from lines 139 to 155

Content

This project is programmed in R. Thus, the user will need to install the latest versions of R programming language and R Studio. In this folder you will:

-ALANN.Rproj (R project): run the open this project to automatically open the R work directory

-main.R (R script): this script uses functions to run neural network models' simulations

-functions.R (R script): here you will find all the functions needed in main.R See the function's description commented within functions.R

-figures (folder): the weights plots (heatmaps) will be stored here

-outputs (folder): activations files after a simulation will be stored here

-weights (folder): weights files after a simulation will be stored here

-*.csv (csv file): this are input files for a simulation

phase = discrete numbers

matType = three values (INPUT, OUTPUT, and TEST; the last is optional)

trialType = trial types as character, identify the trials for training (INPUT and OUTPUT) and TEST

in.$ = input units, where each column should be called different starting with "in."

out.$ = output units, where each column should be called different starting with "out."

par.$ = parameters (par)

par.nH.$ = number of hidden units (nH)
 
  par.nH.nHV = number of hidden visual (nHV)

  par.nH.nHMM = number of multiSmodal units (nHMM)

  par.nH.nHA = number of hidden auditory (nHA)	

par.nI.$ = number of input units (nI)

  par.nI.ctx = number of context units (fully connected)
 
  par.nI.vis = number of visual units (connected with multimodal and visual hidden units)
 
  par.nI.aud = number of auditory units (connected with multimodal and auditory hidden units)

Contact:

Andrew Delamater (AndrewD@brooklyn.cuny.edu)

Santiago Castiello de Obeso (santiagocdo@gmail.com)

References:

Rumelhart, D. E., McClelland, J. L., & PDP Research Group. (1988). Parallel Distributed Processing: Explorations in the Microstructure of Cognition: Foundations, 1.

Delamater, A. R. (2012). On the nature of CS and US representations in Pavlovian learning. Learning & Behavior, 40, 1-23.

Castiello, S., Zhang, W., & Delamater, A. R. (2021). The retrosplenial cortex as a possible 'sensory integration' area: A neural network modeling approach of the differential outcomes effect in negative patterning. Neurobiology of learning and memory, 185, 107527.