Robust-first Computing Wiki - people:taylor_berger

biologically_inspired_nn

anonymous@undisclosed.example.com (Anonymous) — Mon, 08 Sep 2014 22:56:19 +0000

@@ -3,9 +3,9 @@
  
  ===== Biological Structure Overview =====
  Neurons are made up of (essentially) four components: the Soma, axons, dendrites, and synapses.
  
- {{:people:taylor_berger:nb_neuron.gif?200|}} 
+ {{:people:taylor_berger:nb_neuron.gif|}} 
  
    * **Soma**: stores the current charge of the neuron but also leaks charge over time. Once it charges up to a certain point, the charge is released via the axon hillock and sent down the rest of the axon to the synapses
    * **Axon(s)**: the output trajectory (path?) of the neuron. Think of it as an extension cord from the Soma to the synapse
    * **Dendrite(s)**: the input to the Soma. These tree-like structures collect and disperse the charge throughout the rest of the dendrites until it is collected by the Soma or dissipates. 
@@ -43,4 +43,12 @@
    * Moving the network? Exponential decay function again? So maybe the outside edges "move" more than the internal edges?
    * Sensory neurons? Create separate elements designed to just look for things in their event window and have them stimulate themselves? Have other elements that stimulate them? I.E.: rods, cones and flash of light elements?
    * Output neurons? What can I do to interpret the output? WHAT DOES IT ALL MEAN?!?!?
    * Where can I find a u-shaped curve? Can I borrow yours?
+ 
+ ===== Goals & Suggestions =====
+   * Developing a network that grows would be rewarding in and of itself
+     * Use DRegs and Regs to construct a robust network of semi-ordered structure
+     * Soma elements consume regs to create new dendrites/axons
+     * dentrites and axon elements consume regs to create branches
+   * Trivial computation (nand gate?)
+   * Nix Element_synapse

infinite_brain_paper

anonymous@undisclosed.example.com (Anonymous) — Mon, 20 Oct 2014 18:35:29 +0000

@@ -1,3 +1,3 @@
  ===== An Indefinitely Scalable Brain: Implicit Neural Networks in a Spatially Distributed System =====
  
- //Abstract: // We define an implicit, spatially distributed neural network and show pattern recognition is not only viable, but can be reconstructed in volatile systems. We use the Moveable Feast Machine architecture to investigate a neural network with implied connections between neurons based on their proximity. We show that this type of neural network can be scaled indefinitely and learns arbitrary patterns despite adverse learning conditions. We also found the most important factor for successful pattern recognition is the density of neurons.  
+ //Abstract: // ...(problem statement)... We define an implicit, spatially distributed neural network and show pattern recognition is not only viable, but robust in its classification tasks in a volatile system. We use the Moveable Feast Machine architecture to investigate a neural network with implied connections between neurons based on their proximity. We show that this type of neural network can be scaled indefinitely and learns effectively despite adverse learning conditions. We show this type of neural network is capable of identifying patterns and performs better than making a random decision in a two-class classification task.

low_fidelity_data_compression

anonymous@undisclosed.example.com (Anonymous) — Tue, 11 Nov 2014 06:30:26 +0000

@@ -31,5 +31,5 @@
  ===== Summary =====
  These blocks are meant to be the low-fidelity signature of the incoming data. This is a small step towards higher-fidelity data compression that is hopefully the start of better methods in data compression and/or transportation of data in the MFM
  
  ===== Figure 1 =====
- {{:people:figure1.png|}}
+ {{:people:taylor_berger:figure1.pdf|}}

more_realistic_nn

anonymous@undisclosed.example.com (Anonymous) — Mon, 29 Sep 2014 16:58:19 +0000

@@ -1 +1,9 @@
+ ====== Indefinitely Scalable Brain ======
  
+ After reviewing my last attempt at coming up with a project, this is my attempt to come up with a more realistic goal by vastly simplifying the structure of the neural network. This project will better tie into how the Movable Feast Machine works and focuses more on creating elements with more stigmergic actions as opposed to basing them off of state.
+ 
+ ==== The Elements ====
+ Really, the only element needed in the new architecture is a neuron element. The behavior is as follows:
+   * Creates a new neuron element if there are any Res in its event window
+   * If its internal charge is greater than some threshold then it 'fires' and transfers some amount of charge to other neurons in the right side of its event window
+   *