Questions about training Neural Networks from the maker

The maker : For who’s pleasure do we train the models?

Who is the curator/ the audience?

What is the purpose?

“If we employ machine intelligence to augment our writing activities, it’s worth asking how such technology would affect how we think about writing as well as how we think in the general sense.” (Ross Godwin )

How to start

Coding Resources for the installation

Machine Talking

Machine Learning

  • Riverscript to process the input text to output text
  • Text Generator, All about Chatbot
  1. Markov Chain
  2. RNN
  3. Deep Learning
  4. TensorFlow’s sequence to sequence library

Neural Network Text

App Recommanded by Janelle Shanne:

1. textgen-rnn – an open-source framework by Max Woolf, written in Python and powered by Tensorflow. It’s the easiest to install (though you still have to know your way around command line a bit) and comes pre-trained so you can get interesting results even from tiny datasets. It runs fine on an ordinary computer’s CPU, and lets you train the same network successively on different datasets, which is fun. It’s not as powerful as the other frameworks, but just fine for simple lists of names. (tested 15th may )


2. tensorflow char-rnn – an open-source framework by Chen Liang, written in Python. It has tons of flexibility, including the ability to adjust dropout, save frequency, and number of saved snapshots during training, and the ability to adjust temperature during sampling. There’s a tutorial here for running it on AWS, and I’m hoping to find a good tutorial for Google Cloud as well.

3. Andrej Karpathy’s char-rnn, an open-source neural network framework for torch (written in Lua). This one has great flexibility in training/sampling parameters, and it seems to run faster on my 2010 Macbook Pro’s CPU than the python/tensorflow models. I’ve been using this one for the simpler datasets.

Ross Goldwin :

General Method

  1. prepend the seed with a pre-seed (another paragraph of text) to push the LSTM into a desired state.
  2. Use high quality sample of output from the model you’re seeding with length approximately equal to the sequence length (see above) you set during training.
  3. Seed the LSTM with a meaningful text, that the machine would complete
  4. build a data set , corpus
  5. Choosing the right settings for a given corpus
  6. train the model
  7. generate output
  8. train again

James Loy : How to build your own Neural Network from scratch in Python- A beginner’s guide to understanding the inner workings of Deep Learning

Chatbot Personality by 5agado , github

Looking for the” language” of the installation


Looking for a language for the installation, digging into Twitter bot ( Shakespeare sonnet from Twitter feed) , and generative text algorithm ( Markov Chain).

Started to learn Python , and Tensor Flow to manipulate text data sets

Goldsmiths : IS53051A: Machine Learning (2017-18)IS71074A: Data and Machine Learning for Artistic Practice (2017-18), IS71068A: Data Programming (2017-18)


Kadenze :     machine-learning-for-musicians-and-artists with Rebecca Friendrick   , deep-learning-with-tensorflow by P. Mital

Some of the very good blogs/ tutorials to get inspired or start prototyping

Narrative Reality by Ross Goodwin
Natural Language Processing Wikipedia

N grams and Markov Chain with Daniel Shiffmann


Tutorial Electronics


Modelling in 3D digital

  • using 3D modelling for having repetition
  • export to laser cutting or 3D printed
  • example of use : bits and bots to make a plotter controlled by arduino and processing sketches
  • “object is embedded by information, fabrication process has to be honest” Lior Ben Gai

Daily experiments, sensing the world

Imagine Communication between objects through their tactile properties

sensors to the world (other machines, human and environment)


plan to use a spin drum on the top of the servo A. work with catch dreamer format as it spin, and automata spinner Robert Race, carousel toy , spin


Plan to encapsulate Two servo motors, A and B

  • Environment produce shadow/ sound/ light …… and trigger A
  • A motor on
  • A turn and hit B ( every time or randomly)
  • B motor on
  • B produce sound


<When Sensing EMF from electronic devices you have on you
 The nodes move A / Nodes = textile + SERVO   / Osc control  / Nodes  sound B , if no EMF>


<Building a matrix ( resisitive or capacitive) sensors > , in order to map touch by machines,  exchanging  OSC messages

Tutorial, inspirations

Etching on Textile

Pattern for a etched PCB?

  • DataPaulette soft Circuits, les circuits souples 

    Dans notre cas, la gravure au vinaigre est une solution accessible qui fonctionne bien (Ref : Tuto sur Instructables ). Cette solution ne fonctionne pas avec les tissus au nickel. Ci-dessous : la composition et les proportions du mélange. Nous avons constaté que la solution fonctionne de mieux en mieux après plusieurs gravures. Nous avons donc imaginé activer le mélange avec une pièce de cuivre ou une pincé de sulfate de cuivre.

    • Eau oxygénée (10 Vol) – 700ml
    • Sel – 70gVinaigre – 2L
  • Etching with vinaigre

Pseudo codes R&D


Network of artifacts.

02/05/2018 ChatBot exploration

  • <create a surrealist “conversation”between the nodes and with the nodes>
  • use the bibi dictionary as a base, binary system with letters , “Ho, He, BikEDa” , translated into shapes
    • code the binary letters, still to be done, shape equivalence bibiDico
  • work with pattern making in processing , to create the “landscape” of the cell? 

01/25/2018 Embedded Signal

Signal can be embedded in an object by adding textured patterns or by modifying an object’s natural texture.This form of watermarking is currently employed to track sensitive or high-value machine parts and to identify containers carrying toxic or hazardous materials. Three-dimensional objects also raise the possibility of encoding signal by arranging artistic elements in space”

  • No one has trespassed the border, the nodes are “talking”/ interacting with each other.
  • Someone crossed the border, his face is snapshot by the LAO mask, the nodes stopped talking to each other, mirroring/ framing/hiding/revealing the human.
  • Human is crossing back, when he is out of the “view”, the nodes are “talking” again, they are talking about him, something about this human remains with the nodes to “talk” about.
  • inspiration : chinese whispers, 1 ,2 ,3 soleil, gossip, bid/offer and how news are spread across exchange markets and transformed into (non)tangible assets, lemurs and their group behaviour
  • border is a framing a door for accessing the nodes world , behind the screen, nodes , projecting their shadows on screen . on the order side the viewers trying to pass

01/22/2018 Secret conversations

UCCC : Ekman,U Ubiquitous Computing, Complexity and Culture, 2016 Edited by Ulrik Ekman, Jay David Bolter, Lily Díaz, Morten Søndergaard, Maria Engberg – Routledge, p125 Irene Mavrommati :  “Ubiquitous computing (Ubicomp ) are complex computing systems; they can also be understood as ecologies of artifacts, services, people, and infrastructure. Systems with such complexity can be approached as component-based systems. …Component based systems are independent systems with loose coupling of interchangeable parts, which communicate via interfaces.”

AmI : ,“The encrypted tunnels allow the private network to communicate privately over public networks.” “The way devices communicate, in a mesh network of devices”

 create a Network of artefacts , interacting with each others and with people
Slip in two groups, on two different locations

-When “alone”, the artefacts are moving accordingly to patterns generated by one algorithm and data previously collected, input/ parameters should include time and location
-When a presence is detected, new data are collected and new patterns, movement, sound or visual are generated, informations are processed , some of them shared to the viewer, so he could modify its input.
-The two groups will evolve to separate species after few iterations, even if they started identical

(Uncanny space).

– using the programming publish-suscribe concepts?

– artefacts moving accordingly to patterns generated by cellular automata algorithms ( ref simulating neural transmissions, synaptic caguama by Lozano-Hemmer)?  make them evolve with genetic algorithm ( the more people looking at them, ref Jessica Field)?

– repulsion / attraction systems, 10 prints, L systems, machine learning, MaxMSP , using OpenData to map to a visual. Use of OSC touch on my mobile to draw.artefacts mixing digital and analog making, embedded with sensors and OSC technology, talking to each other while receiving data from external input ( open data?)


  • algorithms to work on : GA, CA, particle systems,L system
  • technology for input : OSC touch , kinect,OpenData access, web Scraping
  • technology for output : machine learning, rasperry pi, 

draft scenario

  • Input is human , through sensors or data collecting. Output group 1 is visual- Output group 2 is sonic
  • If there is no input, the group 1 are drawing some shapes and machines of group 2 are reacting to the drawings by making sounds,
  • If there is some external input group 1 is taking the input and changes the patterns of its drawing , sending informations to the group 2,  having a private conversation between them . machine 2 output as sounds responding,
  • If the two groups have people , they count the people, the group with more people will have the possibility to evolve and/ or making more noise/ expanding drawings….

01/29/2018 the machine talks to each other in a foreign language

the machine talks to each other in a foreign language, we gave and forgot about it , they still talk this language< Check the ASCII control characters table from old teletype days>


<Define the Network of artefacts contolled via OSC touch?  Network is the Mask, its artefacts are the nodes>

01/20/2018 Listen to the Other via EMF

<the machines will feel the EMF on visitor ‘s body and react with sounds or visuals>