Preventing Toxicity and Unconscious Biases Using Large Language and Deep Learning Models
Learn how to predict and prevent unconscious biases in AI models through a 40-minute conference talk presented by Armstrong Foundjem at the OpenInfra Foundation. Explore the critical importance of developing fair, interpretable models for high-stakes decision-making in healthcare, finance, and justice systems. Discover techniques for early bias detection in multi-class and multi-label problems using large language models to classify diverse data across languages, cultures, religions, ages, and genders. Examine the implementation of fine-tuned BERT transformers for complex NLP tasks, achieving 98.7% accuracy in bias prediction through contextual text embedding and task-specific classification. Address the challenges of identifying biases in distributed online communities with complex data sources while learning practical solutions for building more trustworthy AI systems.