preparing scaling Bayes.Entrep with Prob.Comp

[hyunjimoon]

target audience: BE evaluators
for PC evaluators, i summarized bayes.entrep in sister thread #234

this thread's purpose is two:

1. to put scalable auto-modeling using Bayesian synthesis and domain specific language on my Bayes in Business supporters' radar 📡

2. to share my willingness to bet to prob.comp with Vikash and seek path to officially join prob.comp project

format: Bob Horn's information mural

papers

• Bayesian Synthesis of Probabilistic Programs for Automatic Data Modeling

Gen Tutorial

angie's action

• show Vikash this demo which is stan-based startup pivot simulation that tests existing hypothesis in entrepreneurship literature and align with prob.comp's vision to seek path in proceeding the project to develop startup education and prediction tool under computer human interaction expert's umbrella

• plan developing education material with Academy of Management conference Professional Development Workshops team (led by Andreas Schwab) e.g. gen-finance gen-finance #177

• model entrepreneurial growth IAI. innovative augmented intelligence 🤖 #174 (comment)

• predicting startup growth with hierarchical gaussian process #200

[hyunjimoon]

Fact

• George shared challenges in inference controller

Angie's Belief

• "Kidnap the robot" from my friend (George)'s notebook is the key to understand, explain, predict, synthesize startup pivot.

Angie's action

sharing modularized prior and simulated re-action, asking for verification + further action

modularized prior

simulate re-action to prompt

similar need-solution dynamics in punctured equilibrium from evolutionary psychology?

baseline controller might represent a standard approach to robot navigation, using observed data to infer the robot's position and move toward the goal. In contrast, the robust controller introduces particle filtering, a method that uses a set of "particles" to represent the distribution of possible states, to improve the agent's beliefs about its location, particularly when uncertainties or changes in the environment occur.

concept of punctuated equilibrium from evolutionary biology—and by extension, evolutionary psychology—can offer intriguing insights into the processes of adaptation and innovation, such as those found in startup pivots or the development of robust controllers for localization algorithms.

Punctuated equilibrium posits that evolutionary development is characterized by long periods of stability (equilibrium) interrupted by short, sudden changes (punctuations). In the context of startups, this can parallel the periods of steady growth or consistent business models that are periodically disrupted by significant shifts or 'pivots' in response to market changes, technological advancements, or new information.

Applying this to your friend George's situation with his localization algorithm, the 'Kidnap the robot' animation might metaphorically represent a startup's unexpected shift in the market or a sudden realization that the current business model isn't viable—a 'kidnapping' of the startup's trajectory. George's difficulty with developing an inference controller could be akin to a startup struggling to find a new direction after such a disruption.

The process of developing a robust controller could benefit from evolutionary principles in the following ways:

1. Adaptation: Just as species must adapt to survive after a punctuation event, startups need to pivot effectively to adapt to new market conditions. A localization algorithm must similarly adjust to unexpected changes in the environment.

2. Diversity of Approaches: In evolutionary terms, genetic diversity can increase a species' resilience. For George, incorporating a variety of inference strategies could provide a robustness that a single approach lacks. A startup might explore different business models or product offerings to find the most viable path forward.

3. Selective Retention: Evolution operates on the principle that successful adaptations are retained over generations. Similarly, a robust localization algorithm should retain successful strategies (or 'particles' in a particle filter context) while discarding those that are less effective. For startups, this means learning from what works and what doesn't, maintaining flexibility without losing sight of the core vision.

4. Rapid Iteration: Punctuations in evolutionary terms are relatively rapid bursts of change. For George's algorithm, this suggests that rapid prototyping and iterative testing could quickly identify the most effective strategies post-'kidnap'. In the startup world, swift iteration allows for rapid learning and adjustment to the new market conditions.

By drawing parallels between these evolutionary mechanisms and the challenges George faces, one can suggest that a more dynamic, adaptive, and resilient approach to algorithm development may emerge. This perspective might not only aid in refining his inference controller but also serve as an inspiration for startups navigating their pivots.

[jeanbaptiste]

Punctuated equilibrium posits that evolutionary development is characterized by long periods of stability (equilibrium) interrupted by short, sudden changes (punctuations). In the context of startups, this can parallel the periods of steady growth or consistent business models that are periodically disrupted by significant shifts or 'pivots' in response to market changes, technological advancements, or new information.

Yes! Exactly, i think this is a good mapping of exaptation in the startup evolution world.

[jeanbaptiste]

About Exaptation and Pivot, these two articles (on biotech/pharma innovation!) might interest you:

• https://pubsonline.informs.org/doi/10.1287/orsc.2017.1116 (. In this paper, we measure the frequency of exaptation in the pharmaceutical industry. We find that about 42% of new functions derived from existing drugs have an exaptive nature)
• https://www.semanticscholar.org/paper/Functional-diversification-and-exaptation%3A-the-of-Andriani-Cattani/b303cd47c9f7c19d8e83dc06f05e8f7df355b446 (Functional diversification and exaptation: the emergence of new drug uses in the pharma industry)

[hyunjimoon]

Jaao, Angie
Jaao's (5th year phd) thesis is three components as below:

crosscat: https://jmlr.org/papers/volume17/11-392/11-392.pdf, http://probcomp.csail.mit.edu/software/crosscat/ he seems to be most actively working on the third project with google (large population model)

[hyunjimoon]

Project	Application	Benefit
Probabilistic Programs that are Multi-task Learners for Data Science	Enhance data analysis and predictive modeling for entrepreneurship.	Efficient handling of large-scale, uncertain data, improving decision-making processes.
Probabilistic Programming for Uncertainty-aware Parsing of Natural Language into Domain-specific Languages	Improve parsing and understanding of natural language data relevant to entrepreneurship.	Enhance the accuracy of data interpretation and hypothesis testing, aiding strategic decision-making.
Probabilistic Programming System for Learning Models of Relational Databases	Develop relational data models to support complex entrepreneurial decisions.	Facilitate robust data management and analysis, enabling better resource allocation and risk management.

for relational db:

[hyunjimoon]

to answer charlie's question in halfway persuasion on PC(BE).txt, i prepared three summaries of vikash's talk:

ted talk on AI That Understands the World, Using Probabilistic Programming

summary: Probabilistic programs provide a new symbolic language for expressing uncertain knowledge about possible worlds and the processes to infer them. It's a new medium for knowledge representation. Examples demonstrate probabilistic programming outperforming machine learning systems on tasks such as perceiving 3D structure, reducing errors in perception (compared to Tesla's neural networks), cleaning millions of database records from US Medicare, and forecasting econometric time series. These probabilistic programs seem to understand the world more like humans do, in terms of symbolic representations.

answer to your questions on its novelty + success case:
examples of AI domain experts using probabilistic programming in areas like cancer clinical trials, salary surveys, genetic engineering experiments, and US Census data. In each case, the probabilistic programs learned were able to generate synthetic data very similar to the real data, indicating they captured key patterns and relationships (here).
symbolic aspect of probabilistic programming that sets it apart from neural networks. Symbolic representations, dating back to ancient times, have been crucial for increasing individual and collective human intelligence. Probabilistic programming combines this symbolic approach with probability to handle uncertainty (here).

Probabilistic Programming Tutorial Part 1, 2

Part I

1. An MIT view on probabilistic programming and intelligent systems
2. Example probabilistic programs in Gen

• Online learning of probabilistic program structure from structured data
• Real-time inference of 3D scenes from video

Part II
3. Inference (meta-)programming in Gen
4. Automating and scaling inference
5. OpenGen: from Gen.jl to PyGen and the GenTL C++ Template Library
6. Implications for large-scale AI systems

[hyunjimoon]

🙋‍♀️ query

Q1. among three query language bayesianQL, inferenceQL, genSQL which technology can best support startup pivoting decision? Maturity, Feature, support from developers affect my use. i'd expect older technology would have lower stability but higher versatility and support from developers.

Q2. what is the family tree of the three language? if it is BQL to IQL to GSQL what "need" for feature prompted birth of the latter ones? how would this be relevant to startup pivoting decisions?

👨🏽‍🏫 predicted answer

below is short comparison of three languages from supply and demand's side (historic order)

Supply Side Table

Feature	Form	Function	Goal
Bayesian Query Language (BQL)	- Extends SQL with probabilistic constructs - in: SQL-like queries with probabilistic extensions - out: Probabilistic estimates, synthetic data, posterior distributions, relevance scores	- ESTIMATE: Estimate probabilistic expressions - SIMULATE: Generate synthetic data - PREDICT: Compute posterior distributions - INFER: Perform inference over missing data - RELEVANCE PROBABILITY: Compute relevance scores	- Democratize Bayesian data analysis - Enhance decision-making with probabilistic reasoning - Provide scalable and flexible probabilistic analysis - Improve data quality through imputation and anomaly detection
Inference Query Language (IQL)	- Extends SQL to support probabilistic models - Specific input/output format not detailed	- Automated Bayesian inference - Generate synthetic data - Compute probabilities and mutual information - Override learned models with custom probabilistic programs	- Make Bayesian data analysis accessible to broad audiences - Assist experts in auditing and improving data, models, and inferences
Gen Structured Query Language (GenSQL)	- Declarative extension of SQL - in: SQL queries augmented with primitives for querying probabilistic models - out: Probabilities, synthetic data	- GENERATE UNDER: Generate synthetic data - PROBABILITY OF: Compute probabilities - GENERATIVE JOIN: Join tables with synthetic rows generated from a probabilistic model - Model conditioning with GIVEN	- Enable complex Bayesian inference workflows in SQL - Provide a declarative query language for interleaving SQL with generative modeling and inference

Demand Side Table

Feature	User	Query Usecase	Maturity	Features	Support from Developers	Suitability for Startup Pivoting
Bayesian Query Language (BQL)	- Database users and administrators - Data scientists and analysts - Business analysts - Researchers and academics - Developers and engineers	ESTIMATE PROBABILITY OF FDA_approval = 'Yes' GIVEN cash_needed = 5000000 AND scenario = 3;	- Oldest and most mature technology among the three - Well-documented features - Robust community support	- Comprehensive support for probabilistic constructs - Powerful tools for predictive analytics and uncertainty modeling	- Strong community and developer support - Extensive documentation - Wide range of use cases	- High suitability due to its maturity, extensive features, and strong support - Capabilities in handling complex probabilistic queries and generating synthetic data are essential for making informed pivot decisions in a startup environment
Inference Query Language (IQL)	- Typical SQL database users aiming to apply Bayesian inference and probabilistic programming	GENERATE synthetic_results FROM meditech_model GIVEN cash_needed = 5000000 AND scenario = 3;	- Newer than BQL - Less documented - May not have as wide adoption or support	- Focuses on automating Bayesian inference - Capable of handling complex probabilistic models - Customizable to fit specific analysis needs	- Limited compared to BQL - Potential challenges in adoption and troubleshooting	- Powerful in automating Bayesian inference, but the lack of maturity and comprehensive documentation could be a hindrance for startups needing immediate, reliable support for pivot decisions
Gen Structured Query Language (GenSQL)	- Users familiar with SQL who want to leverage probabilistic models in data analysis workflows	SELECT * FROM meditech_scenarios GENERATIVE JOIN meditech_model GIVEN cash_needed = 5000000 AND scenario = 3;	- Similar to IQL in terms of maturity - Newer and less mature compared to BQL - Community support and documentation might be limited	- Integrates seamlessly with SQL - Declarative approach simplifies complex Bayesian modeling - Effective for interleaving SQL operations with generative modeling	- Less comprehensive compared to BQL - Limited documentation and examples available for advanced use cases	- Effective for users familiar with SQL - Maturity level might be a concern for startups needing stable and well-supported technologies

process

[hyunjimoon]

[hyunjimoon]

casestudy building with prob.comp(matin and mattieu) on medtech company (medical device with >10times accurate imaging) that aims to find win win strategy that satisfies both ceo, vc, employee

[hyunjimoon]