Knowledge Vault 5 /86 - CVPR 2023
Scientific Discovery and the Environment
Sara Beery, Elizabeth Barnes, Josh Bloom, Kyle Cranmer
< Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4o | Llama 3:

graph LR classDef scientific fill:#f9d4d4, font-weight:bold, font-size:14px classDef models fill:#d4f9d4, font-weight:bold, font-size:14px classDef advances fill:#d4d4f9, font-weight:bold, font-size:14px classDef collaborations fill:#f9f9d4, font-weight:bold, font-size:14px A[Scientific Discovery and
the Environment] --> B[Experts discuss CV, ML
for scientific discovery. 1] A --> C[Leverage mechanistic models in
physical sciences. 2] A --> D[Understanding mechanisms, not just
optimizing metrics. 3] D --> E[Explainability, uncertainty quantification needed
for black box acceptance. 4] A --> F[Theories transfer knowledge, isolation
misses bigger picture. 5] A --> G[AI drives down costs,
barriers, enables applications. 6] A --> H[Generative AI: fast surrogate
modeling, inverse problems, discoveries. 7] H --> I[Future: statistical to causal
models of mechanisms. 8] A --> J[AI integrates data for
climate models with humans. 9] A --> K[Limited labels + community
science: global biodiversity monitoring. 10] A --> L[Astronomy: ML for discovery,
classification, intractable inverse problems. 11] A --> M[Equivariant, graph neural networks:
encode inductive biases, improvements. 12] A --> N[Sciences: structured ground truth
for inductive biases, architectures. 13] A --> O[Earth observation: huge datasets,
ML helps analysis. 14] A --> P[Domain expert - AI
researcher collaborations key. 15] P --> Q[Sciences: distinct datasets, problems,
sandbox for algorithms. 16] P --> R[Real-world deployment confronts limitations,
inspires research. 17] P --> S[Reduced ethics considerations with
natural world data. 18] P --> T[Common challenges across domains,
ideas transfer. 19] P --> U[Engage experts, find non-obvious
impactful problems. 20] P --> V[Learn minimum domain knowledge,
establish common language. 21] P --> W[Persevere finding collaborators bought
into ML, CV. 22] A --> X[Sciences adopting AI, engaging
through workshops, conferences. 23] P --> Y[Incorporate collaborators expertise, mechanisms,
math into models. 24] P --> Z[Learn some science, intuition
for meaningful progress. 25] P --> AA[Approach with curiosity about
data, problems uniqueness, challenges. 26] P --> AB[Identify essential vs abstractable
domain science parts. 27] P --> AC[Focus where ML is
impactful, perhaps only way. 28] P --> AD[Find collaborators already engaging
with CV, benchmarking. 29] P --> AE[Expect interesting, challenging, unexplored
problems, not just off-the-shelf. 30] class A,B,C,D,E,F scientific class H,I,J models class G,K,L,M,N,O advances class P,Q,R,S,T,U,V,W,X,Y,Z,AA,AB,AC,AD,AE collaborations


1.- The panel consists of experts in scientific fields discussing how computer vision and machine learning can advance scientific discovery.

2.- In physical sciences, there are often strong mechanistic models of the data generating process that can be leveraged.

3.- Scientific progress is measured by understanding underlying mechanisms that give rise to data, not just optimizing a single metric.

4.- Black box models are not readily accepted in science without some level of explainability and uncertainty quantification.

5.- Scientific theories transfer knowledge across domains, so isolating problems misses the bigger picture.

6.- Advances in AI are driving down costs and barriers to entry, enabling more scientific applications.

7.- Generative AI is allowing fast surrogate modeling of complex simulators, accelerating inverse problems and enabling new discoveries.

8.- Future progress may move from statistical generative models to causal models that capture underlying mechanisms.

9.- Climate models incorporating human systems and behavior would be a major advance, enabled by AI's ability to integrate disparate data.

10.- Leveraging limited expert labels with community science has enabled global-scale biodiversity monitoring using computer vision.

11.- Astronomy unlocked bottlenecks by using machine learning for discovery, classification and inference on intractable inverse problems.

12.- Equivariant neural networks and graph neural networks allow encoding physical inductive biases for exponential improvements.

13.- Sciences provide richly structured ground truth for studying effectiveness of inductive biases and neural architectures.

14.- Earth observation from satellites generates huge datasets of complex images that machine learning can help scientists analyze.

15.- Collaborations between domain experts and AI researchers are key, with both sides learning from each other.

16.- Sciences offer interesting datasets and problems distinct from typical computer vision applications, providing a sandbox for new algorithms.

17.- Real-world deployment in sciences directly confronts limitations of current methods and inspires new research directions.

18.- Ethical considerations are often reduced when working with scientific data about the natural world rather than human subjects.

19.- Many common challenges exist across AI applications in different scientific domains, allowing transfer of ideas.

20.- Engage with domain experts to find the interesting, non-obvious problems where AI can have an impact, not just low-hanging fruit.

21.- Learn the minimum amount of domain knowledge needed and establish common language, without needing to become a domain expert yourself.

22.- Persevere in finding the right collaborators who are bought into using machine learning and computer vision methods.

23.- Sciences are increasingly adopting AI and engaging with the computer vision community through workshops and conferences.

24.- Incorporate your collaborators' expertise to build the model using the mechanisms and math of their field.

25.- Be open-minded to learn some of the science and intuition behind the data in order to make meaningful progress.

26.- Approach collaborations with curiosity about what makes the data and problem unusual or challenging compared to previous experience.

27.- Identify parts of the domain science that are essential to understand versus those that can be abstracted away.

28.- Focus on problems where machine learning is not just impactful but perhaps the only way to make progress.

29.- Find collaborators who have already started engaging with computer vision and benchmarking methods themselves.

30.- Expect intellectually interesting, challenging, and unexplored problems, not just applications of off-the-shelf methods.

Knowledge Vault built byDavid Vivancos 2024