In the context of text-to-image generators, we discuss how a neural system could be enhanced to follow a knowledge base expressing certain moral considerations, in order to address some looming concerns. We also propose self-educating procedures that enable the system to produce rule-based approximations of its functioning, and illustrate this with an application to the creative interpretation of images.
Until relatively recently, neural image generators were developed for largely utilitarian purposes,
for instance BigGAN [
Over the last 18 months, the situation has dramatically changed and one of the main usage
of neural image generators is currently to produce highly imaginative images which could
never exist in reality. Huge neural models with billions of nodes and training costs in the
millions of dollars from OpenAI, Google and others have been produced which can, on demand,
generate images quickly with high fidelity to a given text prompt, no matter how outlandish
and imaginative the prompt is. As examples, figures 1(a) and 1(b) portray generated images
from the DALL-E 2 (OpenAI) [
In [
We look here at a particular problem with text-to-image generation, namely the lack of diversity in the images generated due to biases trained into the neural models at the heart of the systems. We explore the position that large models trained on internet-sized datasets of images/text provide breakthrough opportunities for content generation, but should be guided by rule-based systems in a neurosymbolic approach. We propose a straightforward way in which a system employing a neural generator could be further educated (rather than merely trained on data), with rules representing moral positions which may directly contradict the training data. We also suggest how symbolic knowledge can be extracted from such models to highlight biases and semantic connections. To illustrate the benefits of this, we present some results from an application to the creative interpretation of images.
When asked to generate multiple images for the prompt “a builder”, DALL-E 2 produces pictures of men only, as per figure 1(c). This is because the models DALL-E 2 employs have been trained on millions of image/text pairs where the overwhelming majority of faces associated with the word “builder” are male. If asked to do the same, many, if not most, young children would draw ifve men in a row if asked to portray five builders. Starting from TV characters such as Bob the Builder, through every trip past a building site and all but the most inclusive of books, builders are – in the experience of children – overwhelmingly male in nature. In the terminology of machine learning, children are trained with a bias towards portraying builders as male.
While it may be disappointing to progressive parents or teachers that a child expresses lack of diversity when portraying certain professions such as builders, this ofers an opportunity to educate children with a more general lesson which may expressly contradict the evidence of their own experience. In particular, parents can tell a child that a person of any gender can be a builder, and/or they may ofer a more general lesson that, in principle, a person from any background can do any job (perhaps with additional training). In future, if a child so educated is asked to draw a builder, they may recall this lesson and draw a female builder from their imagination rather than their experience.
A purely deep-learning approach to address biases baked into large neural models would be
to curate and balance the training data e.g., to make sure there are an equal number of male
and female builders represented. While this might be feasible for training a generative model
such as a GAN [
Scraping internet-scale datasets for training is very convenient, and it is already dificult/impossible for organisations to remove ofensive images from the training set; it seems highly implausible to imagine attempting to balance training data for large models. Moreover, this would miss one of the big advantages of text-to-image generators, namely that they can be directed to produce images that are not representative of their training data. That is, while DALL-E 2 and Imagen may by default produce a white, male face in a yellow builder’s hat when prompted with “a builder”, when prompted instead with “a female builder”, this would override the default and deliver an image of a woman dressed as a builder. A position we take is that the pre-trained neural models in text-to-image generators are equivalent to naïve children who have generalised from what’s around them, but who need to be educated with higher-level rules which may contradict their experience. That is, rather than blaming the neural model, the data or the training regime for the inherent bias that it can be used to portray, we should blame (and ifx) the processes surrounding the employment of the neural model for the bias.
We propose that neural systems employing pre-trained large image/text models for image
generation should be guided by what could be called a moralbase of high-level rules. In particular,
these rules could be employed for automated prompt engineering [
In the same way that it is not feasible to balance data and retrain a neural model to remove biases, it is not feasible to imagine a prompt engineering scheme which can handle all biases which may present. Hence, we expect that moralbases will be highly focused and probably specific to a particular user, organisation, project, political leaning, etc., and that individual users could build up multiple moralbases to switch between. We also expect that automated reasoning, constraint solving, planning and other classical AI techniques will be required for moralbases to be used to their full potential, and we are currently experimenting to this efect.
The building up and deployment of moralbases could be time consuming and dificult, and some
automation may improve matters by identifying biases and relationships in the pre-trained model
that the moralbase guides. To make initial proposals for this, we have recently experimented
with an implementation of CLIP-VQGAN which can perform text-to-image generation, as
described in [
As an artistic application of this kind of semantic extraction, we generated some creative
interpretations of images. To do this, we implemented a process which finds both nouns and
adjectives closest (in terms of cosine distance of CLIP encodings) to an image, then (as above),
we extract CLIP relationships between the adjectives and nouns to produce sentences which
caption the image. Given an image and a caption, CLIP-VQGAN can generate a novel image
which reflects both the original image and the caption. We used this functionality to produce
a series of creative interpretations given in the appendix. As an example, the system has
interpreted/captioned the first original image in the appendix as a “worn and forsaken ruin” and
produced a new image accordingly, which also references strongly the original. Social media
and the internet are becoming saturated with text-to-image produced imagery, and, artistically,
there are rapidly diminishing returns to the production of such images. Our aim with such
creative interpretations, and other more sophisticated generative projects we have planned, is
for generated images to be seen as artworks with an aura [
With this position paper, we propose that neurosymbolic systems, especially those generating
novel content, could reason over knowledge bases of moral considerations which educate them
with higher-level lessons. We further suggest automatically extracting symbolic knowledge
from pre-trained models to aid in this. We are currently implementing prompt engineering
processes to employ in the context of the @artbhot [
We would like to thank Amy Smith for insightful comments on the web of meaning into which artworks are situated. We would also like to thank Katherine Crowson and to the @nerdyrodent developer for their work on the CLIP-VQGAN colab notebook and github repository. We would also like to thank the anonymous reviewers for their insightful comments.
by Isaac Ganuza “Fluid and Threadbare Web” “Wiry and Torn Rope” “Worn and Forsaken Ruin” “Wrathful and Ancient Spirit” “Harmonious and
Imaginative Transportation” “Ofbeat and
Grandiose Machine” “Energetic and
Immense Revolution”
“Daniel-Henry Kahnweiler” by Pablo Picasso “Triangular Structure” “Granular and Dreary Rain” “Quarrelsome Confusion” “Strident and Jagged Shift”