Computing Culture; a humanities approach to artificial neural networks

Kieran Browne — Thesis Proposal Review

Artificial neural networks are notoriously difficult to understand.
Are a class of algorithm inspired by the brain
They allow us to extract patterns from data and repeat these in future
Critical approaches to data
I'm going to argue that:
Neural networks are a cultural way of knowing
They are vulnerable to cultural biases
They can enact learned biases on a grand scale
They are the algorithm behind recent success in solving problems that were previously only possible by humans.
They are universal approximators
Most commonly, they are trained on an example dataset and mathematically approximate the patterns it contains.
Their notable successes and generality has
Drive cars
Compose poetry
Translate between languages
Describe images
Grant parole to inmates

Neural networks are considered to be black boxes because it is difficult to understand its internal processes
A black box refers to a system whose internal patterns and process are inscrutable
We give them a bunch of data, and they transform it into some other form
Training a network involves running a bunch of data through the network telling it where it went wrong and making small updates to improve its performance
This process can approximate any pattern contained in data, even ones the researcher is not aware of
This is a powerful technique because it gives us a handle on systems which are yet to be fully understood
Neural networks are black boxes
BB problem cited many times as a barrier to adoption
But after some major successes in image recognition tasks people have started using them anyway
Levels of abstraction
It's not that we can't see inside neural networks as such, just that relationships between input and output are obfuscated in complexity.
Nowadays we come into contact with neural networks all the time
You might not care if google translate inexplicably gives you a bad translation every now and again
You might feel differently about getting into a self-driving car, if no one knows how it works.
We want our networks to explain themselves
Semantic meaning is needed to make sense of these things
For deep neural networks, the raw text or pixel data is rarely a we want our explanation for a self-driving car in terms of semi-trailers, pedestrians and cliffs
We would like technicians to be able to explain the decisions of the network but this isn't always possible

The problem

Semantics derived automatically from language corpora necessarily contain human biases

Caliskan, Aylin, Joanna J Bryson, and Arvind Narayanan. (2017).

The nightmare scenario

Explaining visually

Theory

Johanna Drucker

Humanities Approaches to Graphical Display

Rethinking our approach to visualization and the assumptions that underpin it

Constructivist notions of data
Representing ambiguity
Designing for interpretation and sense-making

The way we use and understand neural networks is built on realist models of knowledge

Proposed methods

Critique current visual language of neural networks with care for the assumptions they bring.
Design experimental visualisations and interfaces to explore the possibilities of visual representation

Questions

Artificial neural networks are a potent technology capable of extracting and reproducing complex patterns from data, including cultural, linguistic and visual processes previously beyond the reach of computation. However, the internal functions of neural networks are notoriously difficult to understand. The researchers and technicians who design and train neural networks can rarely explain their behaviour. This is particularly problematic in cases where neural networks learn, and thus perpetuate cultural biases. My research inherits its theoretical framework from the (digital) humanities, particularly the ideas of Johanna Drucker regarding graphical display and constructionist notions of data. I intend to make use of mixed methods in my research. This includes critical analysis of existing representations of neural networks and practice-led research, specifically developing experimental visualisations and novel exploratory interfaces. - introduction - what's a neural network - what's a black box - why is the neural network a black box - the black box getting darker - semantics - we want our explanations in terms of meaningful concepts - extracting semantics - problems extracting semantics - marked and unmarked categories - networks often uncritically applied to cultural expressive forms - why is this a problem - examples of cultural bias in neural networks - computer science is illequipped handle questions of culture and bias - code needs criticism - scope - - theory - Johanna Drucker - humanities approaches to (graphical display/ interface theory) - data vs capta - precedents (why visualisation) - knowledge as visually mediated - when you have a hammer everything looks like a nail - attempts to visualise neural networks with techniques for data vis and graphical user interfaces; but neural networks are not data and not much like normal digital tech - methodologies - practice led research - research through making - interfaces - visualisations - critical artworks - critical approach to neural networks - critical approach to data (CAPTA) - redescribing a neural networks as cultral computing - neural networks as cultural actors - training as enculturing - what are the semantics of internal representations - originality - related work