644 Homework 2

In this homework, you will choose from one of the existing, functional code base and (1) replicate previous results, and (2) make small modifications and run a small set of new experiments. Some of these are small in scale, suitable to be developed later into an individual project. The more complex ones can become a team project.

Choose one task from below, and I will provide you with the details on an individual basis (or in groups of those who are interested in the same task), plus the working code. Please choose your preference by Tuesday, 3/9/2010.

Codebase Environment Replicate previous results Modifications Experiments

Choe's Thalamocortical circuit model in XPPAUT:
[Download: thalcor.tar.gz]
[Dir] XPPAUT: use your own code from homework 1 Run all experiments in Choe (2004) Instead of exponentially decaying PSP, use alpha function. Run the same simulations with the revised model, adjusting parameters to get the same qualitative behavior.

Sensory-invariance-driven action code:
[Download: sida-nat.tar.gz]
[Dir] Octave Run the basic simulation for autonomous learning of internal state, on synthetic and natural images. Learn RF using a random policy and Hebbian weight adaptation. After RF learning is complete, learn R(s,a). Run the same simulations with the revised model.

Simple neuroevolution code for feedforward neural networks:
[Download: ga.m] Octave Run the basic simulation for learning boolean functions. Modify network to have context layer (hidden layer feeding back to itself). Experiment with temporal sequence learning (supervised).

NEAT code (in JAVA):
[Download: anji-yc.tar.gz] Java Run the basic examples (XOR and navigation). Develop methods for analyzing the function of the resulting topology.

Topographica
[Download: download page] Topographica Run two tutorial simulations: Tutorial page Add integrate-and-fire neuron types to the code base. Extend the coupled-neuron experiment and get synchronization/desynchronization behavior

Sarma and Choe's orientation energy code for contour saliency detection and associated code
[TBA]
Octave Run code on new images and compare to your own perceptual experience. Derive and implement a neural network to learn response threshold. Use a square-root activation function at some stage during the process. Train and test using the human data set.

Your own codebase, relating to cortical networks ? ? Discuss with the instructor Discuss with the instructor

Tue Mar 2 16:40:17 CST 2010 Yoonsuck Choe

Codebase	Environment	Replicate previous results	Modifications	Experiments
Choe's Thalamocortical circuit model in XPPAUT: [Download: thalcor.tar.gz] [Dir]	XPPAUT: use your own code from homework 1	Run all experiments in Choe (2004)	Instead of exponentially decaying PSP, use alpha function.	Run the same simulations with the revised model, adjusting parameters to get the same qualitative behavior.
Sensory-invariance-driven action code: [Download: sida-nat.tar.gz] [Dir]	Octave	Run the basic simulation for autonomous learning of internal state, on synthetic and natural images.	Learn RF using a random policy and Hebbian weight adaptation. After RF learning is complete, learn R(s,a).	Run the same simulations with the revised model.
Simple neuroevolution code for feedforward neural networks: [Download: ga.m]	Octave	Run the basic simulation for learning boolean functions.	Modify network to have context layer (hidden layer feeding back to itself).	Experiment with temporal sequence learning (supervised).
NEAT code (in JAVA): [Download: anji-yc.tar.gz]	Java	Run the basic examples (XOR and navigation).	Develop methods for analyzing the function of the resulting topology.
Topographica [Download: download page]	Topographica	Run two tutorial simulations: Tutorial page	Add integrate-and-fire neuron types to the code base.	Extend the coupled-neuron experiment and get synchronization/desynchronization behavior
Sarma and Choe's orientation energy code for contour saliency detection and associated code [TBA]	Octave	Run code on new images and compare to your own perceptual experience.	Derive and implement a neural network to learn response threshold. Use a square-root activation function at some stage during the process.	Train and test using the human data set.
Your own codebase, relating to cortical networks	?	?	Discuss with the instructor	Discuss with the instructor