The Blueprints For An Emergent Personality project is back and thriving after our summer break. Theo was with the AMORE team in Barcelona this October, continuing to map and document potential sites of subjectivity and bias that could impact the model.
Here is a video with an overview of our approach, process and progress:
The sketchbook is now complete:
The next steps of channelling this preparatory work into the final series of hypothetical-emotional portraits are now well underway, beginning with visualising and analysing the data for the model's potential world view.
Theo is currently working with an AMORE postdoctoral researcher, who is applying the model’s architecture (a distributed neural network) to reverse engineering an infants’ ability to learn words.
Since the rapid development of neural networks around 2012, capacity to mimic infant development has become more feasible and yet current machine learning systems have not yet provided adequate models of infant language acquisition.
The model is fed actual (textual) raw data (mothers speaking English to their infants) and it predicts the next character in the dialog->utterance (i.e. a context prediction task). It is thus a word discovery system, which extracts similar segments of speech->language across a large corpus, identifying syllables and words. Success is thus not only when words (their shape and their duration) are discovered but also when ‘understanding’ is displayed by predicting their frequency according to context.
The resulting model, after exposure to the data received by a 6-month-old infant, learns many language patterns. It babbles.
Theo will be working intensively on the final stages of the project through January with regular updates.
The goal remains on track to produce a series of large mixed-media drawings that explore how can we represent and ultimately communicate with non-human personality. And how, in doing so, we can best combine objective analysis with subjective interpretation of both known and unknown biases.