The End Of Knowledge - Vault 2 - ICLR (2014-2023)

graph LR classDef crespo fill:#f9d4d4, font-weight:bold, font-size:14px; classDef ai fill:#d4f9d4, font-weight:bold, font-size:14px; classDef art fill:#d4d4f9, font-weight:bold, font-size:14px; classDef data fill:#f9f9d4, font-weight:bold, font-size:14px; classDef debate fill:#f9d4f9, font-weight:bold, font-size:14px; A[Sofia Crespo
ICLR 2023] --> B[Sofia Crespo explores
biology-AI interactions. 1] B --> C[Crespo fascinated by
microstructures, life forms. 2] B --> D[AI tools reveal
new perspectives. 3] D --> E[Neural Zoo: AI-generated
creature textures. 4] E --> F[Artificial life inspires
Crespo's AI creatures. 5] E --> G[GANs generate 3D
insect models. 6] D --> H[Cyanotype combines AI,
analog photography. 7] A --> I[AI art rapidly
changed since 2018. 8] B --> J[Crespo prefers custom
AI models, datasets. 9] J --> K[Crespo collects aquatic
data, recordings. 10] J --> L[Project trains models
on organic photos. 11] J --> M[Algorithmic simulation generates
coral reef models. 12] A --> N[Endangered species lack
online data, images. 13] N --> O[Charismatic species have
more online interest. 14] N --> P[GAN maps outputs
to endangered species. 15] P --> Q[Limited data affects
model accuracy. 16] P --> R[Times Square raises
endangered species awareness. 17] J --> S[Marine sensor data
drives generative art. 18] A --> T[Visual indeterminacy resonates
with Crespo. 19] T --> U[Codex Seraphinianus inspires
fantastic creatures. 20] T --> V[Medieval chimeras similar
to GAN creatures. 21] A --> W[Crespo prefers constructive
AI tool dialog. 22] W --> X[Polarized AI debate
in arts community. 23] W --> Y[Opt-out options important
for training data. 24] A --> Z[Quirks, imperfections appeal
in AI art. 25] Z --> AA[Artistic intent differs
from technical objectives. 26] Z --> AB[AI-generated artifacts as
art debated. 27] A --> AC[AI tools won't
stagnate creative art. 28] AC --> AD[AI normalized, but
self-expression distinguishes art. 29] AC --> AE[Engineers should respect
artistic imperfections, processes. 30] class A,B,C crespo; class D,E,F,G,H,I,J,K,L,M,P,Q,S,W,X,Y,Z,AA,AB,AC,AD,AE ai; class N,O,R data; class T,U,V art; class W debate;

Resume:

1.-Sofia Crespo, an artist exploring interactions between biology and AI, gave the opening keynote at ICLR 2023, an AI conference in Rwanda.

2.-As a child, Crespo was fascinated by using microscopes to see microstructures and life forms invisible to the naked eye.

3.-In 2018, Crespo was introduced to AI tools like neural style transfer which had a similar impact of revealing new perspectives.

4.-Crespo experimented with using AI to generate textures and images of jellyfish, amphibians, reptiles, birds and insects in a series called Neural Zoo.

5.-The field of artificial life, popular in the 80s/90s, inspired Crespo to explore generating biomorphic shapes and evolved virtual creatures using AI.

6.-Collaborating with engineers, Crespo generated 3D models of insects by training GANs on "MRI-like" image slices of 3D insect meshes.

7.-Crespo printed AI-generated images using 19th century cyanotype techniques to combine modern AI with analog photographic printing hundreds of years old.

8.-AI art has changed rapidly from 2018 to now; back then it required much explanation, now tools like DALL-E, Midjourney are well-known.

9.-Crespo believes in training custom AI models on self-collected datasets vs just using pre-trained models in order to connect with her subjects.

10.-Crespo collects data through methods like diving expeditions with aquatic drones, marine sensors, hydrophones to record across frequencies inaudible to humans.

11.-In one project, Crespo trained models on a handmade dataset of camera-less analog photos using organic materials and plant chlorophyll as emulsions.

12.-To generate 3D models of coral reefs, Crespo used an algorithmic simulation as training data since real 3D scans of seafloors are difficult.

13.-Looking through the Red List of endangered species, Crespo noted many critically endangered species have little available data or images online.

14.-Species that are cute/charismatic like red pandas have much more online data and interest than obscure species like lichen, though both are important.

15.-To highlight this disparity, Crespo scraped iNaturalist to get data on 10,000 species, trained a GAN, and mapped outputs to critically endangered species.

16.-Due to limited training data, the model sometimes generated creatures that looked nothing like the intended species, highlighting the representation problem.

17.-The resulting images were displayed on giant screens in Times Square to raise awareness about the lack of data on many endangered species.

18.-Crespo is exploring using marine sensor data to drive meaningful interactions and generative art, and working with sounds captured via hydrophones.

19.-The paper "Visual Indeterminacy in GAN Art" resonated with Crespo in how GANs can generate images that suggest meaning but defy easy identification.

20.-Crespo is inspired by works like the Codex Seraphinianus which appears to be an encyclopedia of fantastic creatures rendered in a realistic style.

21.-Accidental chimeras from inaccurate illustrations based on explorers' secondhand descriptions in medieval times have a similar appeal as GAN-generated creatures.

22.-Over-hyping new AI tools leads to lots of noise and things dying down faster; Crespo prefers constructive dialog about their actual capabilities.

23.-There is polarized debate in the arts community about the impacts of AI; some proudly declare their work "AI-free", others embrace the tools.

24.-Crespo believes providing opt-out options for artists who don't want their work used for training AI models is important moving forward.

25.-AI achieving photorealism doesn't necessarily make it more interesting to artists; the quirks, imperfections and "indeterminacy" are part of the appeal.

26.-Artistic intent can't be cleanly reduced to technical objectives like classification accuracy; artists may even prefer "overfitting" for certain goals.

27.-There's ongoing philosophical debate on whether AI-generated artifacts can be considered art, and the roles of human intent vs cultural consensus.

28.-The ubiquity of AI tools likely won't make all art converge or stagnate, just as ubiquitous cameras on phones didn't kill creative photography.

29.-AI will become normalized as a tool but dedication and having something meaningful to express will still distinguish art as a means of self-expression.

30.-When collaborating with artists, engineers should respect that optimizing everything isn't always the goal; imperfections and manual processes can be desirable.

Knowledge Vault built byDavid Vivancos 2024