DEC AltAc (Alternative to Academia) is a monthly seminar designed to bring together researchers and students interested in the "valorisation" of research in cognitive science, through the link research-industry-society in all its forms. Speakers are invited to share their experience and discuss with participants during an informal presentation.

Wednesday, December 18th, 2024, 12:30am – 13:30pm, ENS,  Émile BOREL (U203), 29 rue d’Ulm, 75005 Paris

Charles de Dampierre (Bunka.ai) - “Extract trends and patterns from large quantities of information. Approaches from Computational Social Science, Human Computer Interface and Artificial Intelligence

Abstract

With the rise of Big Data and Artificial Intelligence, humans need to interact with new entities: AI chatbots on one hand and an ever-growing amount of information on the other. These two aspects are interconnected: AI is fed by a growing amount of information and, in turn, it creates even more information.
  
Many challenges arise for humans to interact with and control these systems: How to memorize information? How to detect the right information in a vast amount? How to detect patterns? How to increase discoverability and serendipity? How to leverage Collective Intelligence ? In other words, how can these systems be made more transparent so that humans can make decisions better, sort information better and use information better?

Existing approaches to Big Data systems rely on simple visualizations or human heuristics. However, these methods become limited when dealing with unstructured data (such as text and images) and when the volume of data scales up: summarizing and understanding 1,000,000 documents in a limited amount of time is still a complex task. For instance, pattern detection in large-scale datasets is impossible with the human eye and decision making can then only be made on the basis of a limited amount of information. This is detrimental to Collective Intelligence where aggregation of information leads to better insights (Kameda, 2022).

Research in the field of Human Computer Interaction (HCI) and Artificial Intelligence (AI) has shown that it was possible to create systems to help humans interact better with complex informational environments. A commonly supported concept in information acquisition is that individuals learn more effectively when they can control their learning experience, meaning the information acquisition process is self-directed (Gureckis, Markant, 2012). Active Learning systems are designed to encourage user discovery, enhance personal agency (i.e., when someone feels its actions lead to outputs), and improve cognitive outcomes such as memorization and the development of higher-order-thinking skills (Dubinsky, 2024). For instance, increase in controllability of these systems lead to better decisions (Rouault, 2022) as well as a better acceptance of those systems themselves (Vantrepotte, 2022). 

To develop such systems, AI technologies have proven very useful with the concept of "embeddings" (Devlin, 2018), which enables it to map information on a two-dimensional scale. This mapping acts as an active learning system, promoting exploration and leveraging our innate ability to visualize the world and compare concepts visually. Cognitive science has demonstrated that 2D maps and diagrams are particularly effective for representing complex information dimensions—such as topic distribution and document relationships—in a way that is easily comprehensible (Harold et al., 2016; Olshannikova et al., 2015). Moreover, recent advancements in neuroscience indicate that the human brain utilizes the same neural mechanisms to encode both physical and abstract spaces (Bellmund et al., 2018), further supporting the effectiveness of this approach.

Despite this abundant literature in mapping information with tools like LDAvis (Sievert & Shirley, 2014), UMAP (McInnes et al., 2020), TSNE (Cai & Ma, 2022) or Wizmap (Wang et al., 2023) or Nomic.ai, those systems are limited in the amount of data they can process and the insights they can create.

During this presentation, I will present the tools developed by Bunka to explore large datasets and better understand the interaction between users and Artificial Intelligence.”
 

Links: 

Short bio:

Dr. Charles de Dampierre, the CEP of Bunka.ai, a company that analyzes the trends and recurrences in interactions between artificial intelligence and users. Charles de Dampierre holds a PhD in Computational Social Science from ENS Paris, a master degree in Cognitive Science from ENS/EHESS and a master degree from HEC Paris. He is is a laureate 2024 of the i-PhD competition for the Bunka project. 

 

About the DEC AltAc seminar series