Abstract

Machine learning models (e.g., generative models) are useful for automating many tasks, including generating content (e.g., text, image, video, etc.). However, outputs generated by these models may include undesirable content that reflects biases and stereotypes. Thus, machine unlearning approaches have been leveraged to remove bias in models. Conventional techniques for removing bias from models typically rely on contrastive learning to learn unbiased models, but these techniques generate full-precision models that require significant resources for implementation and deployment. To address these issues, techniques proposed herein provide mixed-precision quantization for machine unlearning through contrastive learning, thus facilitating the implementation of a quantized unlearned model that is efficient and accurate.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS