Periodic Reporting for period 2 - SIAM (Scaling Up Innovation through Analogy Mining)
Periodo di rendicontazione: 2021-07-01 al 2022-12-31
Many world-changing breakthroughs in science and technology were enabled by analogical transfer, as ideas from one domain were used to solve a problem in another. Observing water led the Greek philosopher Chrysippus to speculate that sound was a wave phenomenon; an analogy to twisting a cardboard box allowed the Wright brothers to design a steerable aircraft. Despite its value for innovation, very little progress has been made towards automating the process of analogy-finding in real-world settings, and the problem has maintained a longstanding status as a “holy grail" in artificial intelligence.
The goal of this project is to tackle head-on this important problem and develop principled tools for automatically discovering analogies in large, unstructured, natural-language datasets such as patents and scientific papers. Such tools could revolutionize a variety of fields, allowing scientists and inventors to retrieve useful content based on deep structural similarity rather than simple keywords.
This project explores the multiple roles AI and machine learning can play in the analogical innovation pipeline. We focus on the three core components of the pipeline:
(1) Developing representations and similarity metrics to facilitate comparison between potential analogs
(2) Imbuing the algorithms with commonsense knowledge and abstraction capabilities
(3) Guiding people in adapting the discovered analogies to solve the original problem
The goal of this project is to tackle head-on this important problem and develop principled tools for automatically discovering analogies in large, unstructured, natural-language datasets such as patents and scientific papers. Such tools could revolutionize a variety of fields, allowing scientists and inventors to retrieve useful content based on deep structural similarity rather than simple keywords.
This project explores the multiple roles AI and machine learning can play in the analogical innovation pipeline. We focus on the three core components of the pipeline:
(1) Developing representations and similarity metrics to facilitate comparison between potential analogs
(2) Imbuing the algorithms with commonsense knowledge and abstraction capabilities
(3) Guiding people in adapting the discovered analogies to solve the original problem
We have made substantial initial progress along several fronts:
* We have developed new methods to extract structure from texts and augment it with commonsense facts.
* We have explored new domains to test our ideas on. We have extended our work to natural-language texts describing processes, to biomimicry (bio-inspired design), and to images.
* We have suggested new approaches to improve commonsense abilities of natural-language understanding algorithms, based on recent advances in cognitive psychology.
* We have identified a set of clues that could potentially be useful for helping people in the adaptation phase, and devised and coded an experiment to test their effectiveness.
* We improved on our initial prototype analogical search engine to create a working interactive version that could be used by researchers in real time. This necessitated major improvements in scalability, accuracy, and enabling new ways for researchers to filter and explore the results. Under realistic usage with PhD-level researchers we validated that our system was able to spur more creative ideas than baselines.
* We have developed new methods to extract structure from texts and augment it with commonsense facts.
* We have explored new domains to test our ideas on. We have extended our work to natural-language texts describing processes, to biomimicry (bio-inspired design), and to images.
* We have suggested new approaches to improve commonsense abilities of natural-language understanding algorithms, based on recent advances in cognitive psychology.
* We have identified a set of clues that could potentially be useful for helping people in the adaptation phase, and devised and coded an experiment to test their effectiveness.
* We improved on our initial prototype analogical search engine to create a working interactive version that could be used by researchers in real time. This necessitated major improvements in scalability, accuracy, and enabling new ways for researchers to filter and explore the results. Under realistic usage with PhD-level researchers we validated that our system was able to spur more creative ideas than baselines.
We expect to continue and make progress towards our goal, ultimately resulting in a prototype inspiration engine that can
(1) Identify analogies between distant domains based on abstract similarity,
(2) Take advantage of commonsense knowledge not appearing in the text,
(3) Highlight the commonalities and differences between the source and target in a way that will help the user quickly adapt the solution to the problem at hand.
We will conduct a series of ideation studies, testing the engine’s ability to improve users’ creativity and innovation.
(1) Identify analogies between distant domains based on abstract similarity,
(2) Take advantage of commonsense knowledge not appearing in the text,
(3) Highlight the commonalities and differences between the source and target in a way that will help the user quickly adapt the solution to the problem at hand.
We will conduct a series of ideation studies, testing the engine’s ability to improve users’ creativity and innovation.