Creativity enhancement by artificial intelligence

GhiblyCalimero / Pixabay

Could AI enhance human creativity? In this article, we’ll explore the way neural networks could propose amazing association of words related to a problem. We trained a simple neural network with 5 million patents between 1976 and 2013, allowing to combine concepts and find new ideas.


Creativity methods like TRIZ, SIT, CK propose to structure the way humans can generate new ideas from a previous knowledge [1][2]. TRIZ proposes to solve the creativity problem by performing an abstraction -based on rules manually deduced from patents- that could inspire creative people for a specific problem.

In 2014, we proposed to use a NLP approach to generate inspiring graphs [3]. In 2013, a brilliant solution was published, allowing to combine words to fit a new sense [4]. For instance, king – man + woman = queen is a coherent equation solved by a deep learning model (word2vec). However, word2vec main pretrained models are based in Wikipedia or news, thus are not relevant for scientific creativity. So could we train a neural network based on a relevant scientific dataset: the US patents.


Word2vec is a method to represent words as vectors having a correlation with other words around. These vectors, as any vectors, can be added or subtracted to generate new vectors representing a new word. For this experiment, we trained a neural network on 5 million patent abstracts, representing a 3 Gb dataset. We only considered words that appear at least ten times. Two models can be used for word2vec (CBOW and skip gram), the calculation has been performed with skip-gram. The final result is a set of 100-dimension vectors for 203543 words. We used GENSIM library implementation to generate the pretrained model.

Let’s Create

What can we do that could enhance creativity? Let’s try some stories.

Non toxic Chemistry

R&D Mediation recently contributed to the deeptech startup Alysophil. The idea behind Alysophil is producing chemicals by an environmentally friendly process called microreactors or flow chemistry. This idea didn’t come from AI but what would word2vec say about that.

We can compute : REACTION  ➕CHEMICAL ➕ QUICKTOXIC that gives: MICROREACTOR but also suggest interesting reactions (susuki, solid-solid, enzyme catalysed)


If we ask to the neural network what would be a detonation with less speed, we also get an interesting result:


we also get ‘pyrotechnics’ and ‘gas generation’ that is indeed lower speed reactions.


Other creative equations







Human creativity is a mix of desire and knowledge. If, at the present time, the desire is not covered by AI, the knowledge’s access can be strongly improved by AI. Will AI become a partner for creativity sessions? Word2vec is basically a first step in NLU (Natural language understanding), but a very near future NLU will propose a complete sense understanding of complex sentences. At this moment, AI will likely be the best friend for creators.



[1] L. E. Brunet, “TRIZ, SIT, CK Connections and Disconnections between Three Major Theoretical Frameworks on Creativity,” Current opinion in creativity, innovation and entrepreneurship, vol. 1, no. 2, 2012.
[2] L. E. Brunet, “Créativité en ingénierie,” in ref. article : ag5210, vol. base documentaire : TIB127DUO, T. de l’ingénieur S. de conception pour l’innovation, Ed. 2015.
[3] L. E. Brunet and K. Le Meur, “A big data and Darwinian approach of scientific creativity,” in R&D Management Conference proceedings, Stuttgart, 2014, vol. 1, pp. 401–407.
[4] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient Estimation of Word Representations in Vector Space,” arXiv:1301.3781 [cs], Jan. 2013.


Télécharger cet article au format PDF ou ePub

Laisser un commentaire

Votre adresse de messagerie ne sera pas publiée. Les champs obligatoires sont indiqués avec *

Ce site utilise Akismet pour réduire les indésirables. En savoir plus sur comment les données de vos commentaires sont utilisées.