Tianyi Chen

Machine learning assisted design: a design exploration of my generative symbol design practice

Abstract

Machine learning (ML) is utilized in many fields, but it threatens to replace humans in some occupations where ML is a cheaper alternative. Communication design is one such area. The development of a large body of work over many decades, much of which is available online in similar formats, is the perfect habitat for ML. How should communication design respond? That this kind of technological advancement has contributed to the trend for visual-based designers to transform into verbal-based designers (e.g., service design, copywriting, or user-experience design). However, the solution is not to avoid or delay the impact; I believe we should embrace this technology and find a way not just to cope with it, but also to strengthen communication design practice.

This research will generate new concepts, art styles, and formats that can be developed in the field of art and design, and will identify ways for designers to interact constructively with ML technology. Further, this research will provide the foundation for a greater understanding of ML-aided design (MLAD), an expanded scope of interactive and generative design, and increased utilization of MLAD in the future. The purpose of engineering a ML training structure is to ‘teach’ the machine to meet our needs. Beyond that, what benefits can designers obtain from MLAD? How should we handle results that fail the human criteria of ‘making sense‘? Through three phases of doing, reflecting, and learning, the many possibilities of MLAD proved its strength as a visual sensual detector and concept generation tool for designers.

I started with self-teaching by building a simple feed-forward Neural Network (FNN) system which is to teach the machine to understand the intensity of different videos. I made the neural network system limited to single value output to exam how much a simple machine learning (ML) structure can do in design. I generated my design concept: the process from observation to creation that I was inspired by how neural network works. Then I did two generative design output researches. This stage gives me a basic insight on how machine learning runs and what its production process should look like.

I did a couple of upgrades on the FNN system and added a physical sensor. And I created these generative cross symbols to investigate the concept of phase 1 the process from observation to creation. What I learned in this phase is that machine is capable of play as an actor. It can contribute to my work and add an extra layer of interpretation.

The purpose of this phase was to investigate a more complex neural work that is repurposed by me to generates symbols. This project’s outcome is an under-trained sketch Recurrent Neural Network symbol design model that can produce seemingly irrational but inspiring house symbols. I also created experimental wayfinding signage that can continuously playback my ML house symbols. With proper training, this technology can be directly used on signage or logo and plays as a part of a branding system.