circuit-visualizer-p5

🚧 Work in progress 🚧
Experimental visualizer for studying networks / graphs / circuits and properties of their adjacency matrices. The big-picture goal for this project is to inform how best to stimulate neural circuits in order to infer their connectivity.

Try demo:

• clicking the adjacency matrix
• moving nodes around (m)
• typing new circuits into the text field

Keyboard Commands: ⌨️

Editting modes

n: enter "add new edges" editing mode, click a node and drag to another to connect them, do the same again to remove an edge
m: enter "move" mode, drag nodes to reposition them
o: enter "open-loop stimulation mode", clicking on a node delivers stimulation

Display layers

s: toggle showing edges
r: toggle showing indirect connections
w: wiggle nodes, display outputs

x: export graph to text field
<space>: create a new random, sparse network

Completed Features:

- graph import from text field
    - can use bidirectional arrows <->
    - and multi input multi output lines a,b,c<->d,e
- graph export to text field
- toggle directional edges! (with s)
- highlight edges from a second adjacency matrix
- basic binary matrix reps
- can import connections from binary string (check order / convention)
- can bitshift to "rotate connections"
- can toggle connections via adj mat
- can add connections via drag arrows
    - will also remove redundant connections!
- clearMat()

Similar Frameworks:

• xcorr-visualzier-p5
  • improves on visualizing outputs of nodes
• NetworkX
• Gephi gephi network datasets, could be scanned for candidate networks to ID
  • https://github.com/gephi/gephi/wiki/Datasets
  • has lots of layout options, including "force atlats"

• Cytoscape

  Early Findings: 🔍

closed-loop control of a node reveals a lot if node N has many (non-reciprocal) inputs

for two nodes A,B,

if corr(A,B) > th for both ctrl(A) and ctrl(B) A ⟷ B

chains contain both colliders and forks

reach(un(M)) is insufficient to predict correlation in forks

closed-loop control can eliminate colliders (by serving inputs to junction) closed-loop control interrrupts chains

Major Issues: 🐛

• [ ] all node creation functions reference global variables, this makes it difficult to extend

• [ ] signal propogation for correlation inspection is "too correlated"

Feature Requests: 🌠

• [ ] should fork-shaped and collider shaped reachability be undirected?

• [ ] calculate "passes through node X" reachability

  • [ ] this is like a ternary reachability i suppose!
  • [ ] and precursor for "what if condition"

• [~] split edge drawing and matrix highlights into 2 functions

  • dangerous because of semi-arbitrary visualization criteria

• [ ] have data generation, but currently only works with pure sources

  • can't handle reciprocal nodes as upstream sources
  • probably depends on (cycle-compatible) topological sort for resolution

• [ ] have adjacency measure as anonymous function that can be reused? or non-anonymous is fine too ...

• [ ] topological sort -> y position

• [ ] print to console control severance score!

  • [ ] just requires rowSum, colSum, matSum

• [x] visualize where we're controlling

  • diamond around node?
  • [x] used rotating circle segments

• [~] in / out degree quant & viz

  • should we color / size by 1st order degree?
  • or reachable degree?
  • calculate betweenness?
  • this helps build intuition for which points are good to intervene
  • cytoscape scales nodes by total degree

• [~] - actually! this phase propogation thing is what we want to do instead of reachability(Unidirect(M))
  which doesn't identify "common cause" correlation

• [~] after calculating reachability, topologically sort the network

  • all "leftmost" nodes propogate their "phase" to all their children
    • easy! all nodes with no in degree get their own phase
  • maybe have intrinsic noise only move in Y dimension
  • control only move in x direction
  • then addition can cause correlation in both?
  • wiggling should simply be a sum of all input wiggles?
    • normalize by num inputs?
    • verify this is what's implemented
  • [ ] control sets phase to follow control oscillator?

• [ x] calculate reachability (floyd warshall ??)

  • [x.] vis reachability (in adjMat only?)

    • [x] have direct connections
    • [x] can highlight Nth order connections
    • [F] can do offset connections to overlay multiple metrics
    • [ ] vis reachability with something directional!

  • [ ] system for drawing wiggly high order connections

    • "bouncing" arrows?
    • blob communities together?
      • metaballs :O
        • add edges to blobbables
        • subtract non-nodes and node-edges via distance field?
      • convex hull ?
    • offset connections?

• [~] delete edges / nodes gracefully

  • for now just clear everything

• [~] visualize self-connection with loop arrow

  • [x] using circle around node for now

Bonus Features: 🎁

• [ ] nodes wiggle to show correlation :)

• [ ] visualize open-loop stim

• [ ] import-export adjacency in this format:
  {a,b,c,d,e,f,g}
  a→b
  b→c,d
  d→a
  f⟷g
  export
  to matlab, python, (GML?) txt

• [ ] "mute/unmute" all inputs/outputs by toggling column / row in adjacency matrix

• [ ] nicer UI elements (with ControlP5)

• [~] viz multiple mats on one tile

• [~] highlight control-severed edges csev = m & !ctrl(m) overlay red Xs

• [ ] color nodes by community see: Clustering and Community Detection in Directed Networks: A Survey

• [ ] color nodes by K-means after force-directed layout?

• [ ] duplicate edge protection (have collection of children be Sets)

Additional caveats: ⚠️

• the force distribution piece of this is a mess
  • may want links to have forces act on them in future
  • could add "angular attractors", i.e. attract individual nodes to an angle from the center, for default ordering of nodes
• will connect "reachability" highlighting to matrix view
• have simple circuit generation lay nodes out in a static ring
• should one unified representation be managed (i.e. adjMat?)
  • or should the graph and adjmat simply be synced?

Additional Topics:

• generally, should be leveraging graph theory more, for instance
  • idea of where to control can be tied back to "minimum cut" and "highly connected subgraphs"
  • maximal clique
• quantifying degree:
  • in-degree (r)
  • out-degree (b)
  • betweenness (mediator - G) (more complicated)

Interesting Circuits: 🕸️

e->a->b->d
c->b

big fork:
a->b->c->d
a->j->i->h

big collider:
h->i->j->a
d->c->b->a

a->j
b->e
c->e
f->e,h,j
g->g,h
i->d
j->c,e