Human-in-the-loop Neural Networks: Human Knowledge Infusion

Issei Suzuki, Pitoyo Hartono

Keywords: human-in-the-loop;topological representations;metric learning;dimensionality reduction;transfer learning

TL;DR: A study to build a mechanism for allowing infusion of human knowledge into neural networks

Abstract:

This study proposes a method for infusing human knowledge into neural networks.
The primary objective of this study is to build a mechanism that allows neural networks to learn not only from data but also from humans. This motivation is triggered by the fact that human knowledge, experience, personal preferences, and other subjective characteristics are not necessarily easy to mathematically formulate as structured data, hindering them from being learned by neural networks. This study is made possible by a neural network model with a two-dimensional topological hidden representation, Restricted Radial Basis Function (rRBF) network. In rRBF, the hidden layer's low dimensionality allows humans to visualize the internal representation of the neural network and thus intuitively understand its characteristics. In this study, the topological layer is further utilized to allow humans to organize it considering their subjective similarities criterion for the inputs. Hence, the infusion of human knowledge occurs during this process, which initializes the rRBF. The subsequent learning process of rRBF ensures that the infused knowledge is inherited during and after the learning process, thus generating a unique neural network that benefits from human knowledge. This study contributes to the new field of human-in-the-loop (HITL) AI, which aims to allow humans to participate constructively in AI's learning process or decision-making and define a new human-AI relationship.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5416