Minimally Invasive Machine Unlearning via Monosemantic Neural Activation Identification

In the age of artificial intelligence, machines learn from vast amounts of data to make predictions and decisions. But what happens when we need them to “unlearn” something—whether to protect privacy, correct biases, or comply with regulations? Existing approaches to machine unlearning can be effective at removing specific data from a model’s memory, but they often cause unintended changes that make the model less reliable overall. Minimal Invasive Machine Unlearning (MIMU) is a breakthrough approach designed to minimize disruption to a model while still effectively forgetting the necessary information. Instead of making broad adjustments to the entire system, MIMU takes a precise and targeted approach. By identifying how specific data points influence individual neurons in the AI model, it carefully removes only the necessary traces, leaving the rest of the model intact. This innovation helps maintain the accuracy and reliability of AI systems while ensuring that unwanted data can be safely and efficiently erased. We hope MIMU can make a significant step forward in making AI systems more adaptable, ethical, and trustworthy in a world where data privacy is more important than ever.

Faculty Supervisor:

Ga Wu

Student:

Partner:

Layer 6 AI

Discipline:

Computer science

Sector:

Finance and Insurance; Professional, scientific and technical services

University:

Dalhousie University

Program: