AI Models for Structured Data | neuralk-ai

Understanding Tabular Data and AI Advancements

Tabular data refers to information organized into rows and columns, a common structure found in sources like SQL databases, spreadsheets, and .CSV files.

Significant advancements have been made in artificial intelligence, particularly concerning unstructured and sequential data. However, current large language models (LLMs) are inherently flexible.

Challenges with Current LLMs

These models are designed to generate coherent outputs by manipulating input tokens, without strict adherence to predefined structures. Accessing or running these top-performing LLMs can be costly, either through API usage or cloud infrastructure expenses.

Despite these challenges, many organizations have already established data strategies, utilizing data warehouses or data lakes to consolidate crucial information. They also employ data scientists to analyze this data and refine business strategies.

Neuralk-AI: A Focus on Structured Data

Neuralk-AI, a French startup, is dedicated to developing AI models specifically tailored for tabular data. The company recently secured $4 million in funding to further its research and development.

According to Neuralk-AI co-founder and chief scientific officer Alexandre Pasquiou, data that holds substantial value for companies is often long-established, structured in tables, and utilized by data scientists for machine learning algorithm creation.

Targeting Commerce Companies

Neuralk-AI recognizes an opportunity to refocus AI model development on structured data. Initially, the company intends to offer its model as an API to data scientists within commerce businesses.

These companies heavily rely on data – including product catalogs, customer databases, and shopping cart trends – making them ideal early adopters.

Limitations of LLMs for Traditional Machine Learning

“LLMs excel at search, natural language interaction, and answering questions based on unstructured documents,” Pasquiou explained. “However, they encounter limitations when applied to classic machine learning tasks, which are fundamentally rooted in tabular data.”

Applications for Retailers

Neuralk-AI’s technology enables retailers to automate intricate data processes, including intelligent deduplication and data enrichment.

Furthermore, the company’s models can be used to identify fraudulent activities, refine product recommendations, and generate accurate sales forecasts for improved inventory management and pricing strategies.

Funding and Partnerships

Fly Ventures spearheaded the $4 million funding round, with Steam AI also contributing. The investment also included participation from several business angels.

• Thomas Wolf from Hugging Face
• Charles Gorintin from Alan
• Philippe Corrot and Nagi Letaifa from Mirakl

Testing and Future Development

The Neuralk-AI team is currently refining its models. They plan to conduct testing with prominent French retailers and commerce startups.

These include E.Leclerc, Auchan, Mirakl, and Lucky Cart.

“We anticipate releasing the initial version of our model, along with a public benchmark for performance comparison, within three to four months,” Pasquiou stated.

“Our goal is to establish ourselves as the leading tabular foundation model in representation learning by September.”