PROBLEMS WE WORK ON
Making use of humanity’s vast collective knowledge is
hard, and the tools we have are insufficient.
Information is spread across websites, databases, scientific papers, algorithms, statistical models, and more. This makes it hard to access, combine, and use information effectively.
Existing methods of structuring information require significant manual effort to deal with information uncertainty.
Current scalable machine learning methods are intransparent and inexact in their reasoning. This limits their reliability and viable applications.
Our system addresses these problems. It aims to provide services across different domains, for example, acting as analysts, research assistants, or data scientists.
We start with a seed
Initially, we are focusing on building a system that can receive queries in natural or domain-specific language, provide good answers, and an insight into its reasoning.
Answers can be of many types: text, numbers, images, colours, etc. A probabilistic distribution over several possible answers is given, to account for uncertainty.
The procedure that was followed to obtain the answer is shown, providing transparency into the reasoning of the system. This includes reasoning steps as well as a display of data sources, statistical methods and algorithms, machine learning models used, and more.
Our approach in a nutshell
COMPONENTS
A universal language for knowledge
TO REPRESENT
HETEROGENEOUS INFORMATION
A PROBABILISTIC FRAMEWORK FOR REASONING AND LEARNING
TO REDUCE MANUAL EFFORT
NEEDED TO CURATE THE DATA
A VISUAL AND NATURAL LANGUAGE REPRESENTATION
TO MAKE REASONING
UNDERSTANDABLE
METHODS
CATEGORY THEORY
TO CAPTURE
STRUCTURE AND COMPOSITION
BAYESIAN STATISTICS
TO REPRESENT
UNCERTAIN KNOWLEDGE
DEEP LEARNING
TO RECOGNIZE
PATTERNS
COGNITIVE SCIENCE
TO UNDERSTAND
REASONING AND LEARNING
News feed
PlantingSpace