A brief overview of generative AI technologies and their use for social good initiatives, including cultural training, medical image generation, drug design, and public health.
2. What is generative AI?
• Deep learning frameworks that can produce new data based on
input prompts and large training datasets
• Can have any/all of these steps in the framework:
• Encoder-decoder structure
• Training sample matches
• Random noise and blending components
• Comparison steps to ensure realism
4. Text Generators
• Massive training datasets
• Typically scraped and
possibly quality controlled
• Mostly in English
• Deep learning frameworks with
billions of parameters to train
• Can be modified by fine-tuning
• Specific examples relevant to
text generation task at hand
• LoRA as quicker way to train
5. Image Generators
• Many types
• Encoder-decoder steps in some
• Pull up related images
• Blend images
• Add random noise to back-fill
• Image generators plus comparison steps
• Two competing generators with one a
few training steps ahead of the other
• Comparison step to benchmark
against real dataset
• Some rely heavily on topology
7. Case 1: Medical
Image Generation
• Medical imaging data issues:
• Small sample sizes
• Sample imbalance (rare diseases…)
• Issues when augmenting small samples
or imbalanced samples:
• Biological structure fidelity in
generation (ex: ventricles in brain)
• Image variety in generation
8. TopoGAN
• Solution involves a generative
adversarial network with
topological awareness
• Topology
• Betti number
introduction
• Advantages:
• Preserves structures
like branching and
loops
• Generates large
number of images
close to target images
9. Case 2: Human Resource Diversity
Training
• Mindbloom
• Addresses training needs by providing synthetic people with whom to discuss
several types of conversations
• Employee reporting sexual harassment
• Addressing cultural mismatch of new employee
• Policy changes that impact employees
• Misgendering in the workplace
• Conversation and voice generation with proprietary generative algorithms
• Demo
11. Case 3: Protein
Generation
• Designing and testing new drugs takes a lot of
time and money.
• Not good for new pandemics in urgent need
of treatment
• Increased drug costs for consumers
• Many types of proteins/molecules in venom of
different animals
• Metalloproteinases, three finger toxins,
phospholipidase A2, disintigrins…
• Varies by geography and species
• Slight modifications of toxins as good
initial drug designs
12. Graph Generators
• Approach to protein/molecule-specific generative models:
• Translate protein/molecule to graph form
• Define properties of interest (solubility, for instance) or binding score
• Create generative model to work on generating similar graphs
• GAN trials generate new proteins/molecules with:
• Better target properties
• More variety
• Less time/cost to generation than other models/human generation
13. Case 4: Public Health Campaigns
• Many recent infectious diseases that can be spread from person to
person:
• Ebola
• COVID-19
• HIV
• Issues with traditional generation of video and poster messaging to
address behaviors contributing to spread
• Time to create script, image, and translations for local populations
• Lives lost in delays