jpjacobs

NB. Neural Network Demo
NB. (roughly based on chapters 1-3 of http://neuralnetworksanddeeplearning.com/ by Michael Nielsen)
NB. The main differences with the book are:
NB. - the removal of the minibatch function and loops in favor of matrix functions
NB. - joining bias and weight calculations.

NB. *** Preparations ***
NB. install addons needed from github for reading CSV files
{{install^:(-.fexist'~addons/tables/',y)'github:jsoftware/tables_',y}}&>;:'dsv csv'
NB. require the needed addons

jpjacobs

% Machine learning toolbox
module ml.
import util.

% calculates only half of the distances: symmetry
pdist(M) = M3 =>
    M3 = new_array(M.length, M.length),
    foreach(I in 1..M.length, J in I..M.length)
        M3[I,J] = sum([(M[I,K]-M[J,K])**2 : K in 1..M[1].length]),
		M3[J,I] = M3[I,J] 

jpjacobs

NB. knn graph
knn =: 4 : 0
d =. (+/@:*:@:-)"1/~ y NB. distances between vectors in y
n =. x ({. }.@/:)"1 d  NB. get nearest neighbors (exclude self)
W =. n (_"_)`(<@<@<@[)`]}"1 d NB. set all non-neighbors to _
W =. (<.|:) W          NB. make symmetric
)

jpjacobs

NB. Make square grid of side y
grid =: |:@:(,"2)@(,: |:)@(#"0 i.)

plotgraph =: 4 : 0 NB. todo: implement autoplacement of locations not given
pd 'reset'
pd 'type line;pensize 1;color blue'
edges =. ; (i.#x) ;/@,.&({&x)"0 &.> y
edges =. <"1 ~.@:(({~"1 /:@{.)"2) > edges NB. normalize: sort points on ascending x, eliminate double edges.
edges =. (#~ -.@({.&>-:{:&>)"1) edges NB. eliminate loops. Empties automagically disappear.
pd"1 edges

jpjacobs

NB. Self-Organising map
require 'math/lapack/geev'
require 'tables/csv'
require 'viewmat'
NB. closeall_jviewmat_ ''
NB. parameters:
N         =: 30     NB. side of neuron map. (N -> (*:N) neurons)
ntrain    =: 25     NB. number of training samples per class
niter     =: 10000  NB. number of iterations
'a0 an'   =: 25 1 NB. Starting and ending learning factor

jpjacobs

NB. Fast ICA
NB. http://en.wikipedia.org/wiki/FastICA
require 'math/lapack math/lapack/geev'
require 'plot'
center =: -"1 +/%# 			NB. subtract column means
mp     =: +/ .*			NB. matrix product
cov    =: (mp"2 1)~&.|: ] % <:^:(1&<)@#	NB. Covariance (note, does NOT center)
I      =: =@:i.			NB. unit matrix
clean  =: * |@* 			NB. clean near-to-zero values
whiten =: (mp ([ mp (I@# % %:)@] mp |:@[)&>/@(2b110&geev_jlapack_)@cov)@center

jpjacobs

NB.  Signal Processing in J
NB. ========================
NB. Mainly written as exercise, so no guarantees
NB. cocurrent <'jimproc'
NB. Assumptions : 1-cell = 1 sample of dimensionsionality $ samp
NB. set of samples in 2-cell

NB. Monad mean y
NB. Mean of data samples in y
mean =: +/ % #

jpjacobs

NB. Listen to UDP stream (phone IMU eg.)
load'socket'
load'graph'

(;:'jsocket jdefs z') copath 'base'

sk =: 0 pick sdcheck sdsocket AF_INET_jdefs_,SOCK_DGRAM_jdefs_,0
sdcheck sdbind sk;AF_INET;'';5555

NB. sensor_number check socket