Sussex AI seminar: Robyn Waller and Russell Waller

Sussex AI seminar: Robyn Waller and Russell Waller By: Aleks Kossowska Last updated: Wednesday, 12 November 2025 Abstract: In Waller & Waller (2021), assembled bias names how composing local pieces into a global picture can introduce emergent error—often worsened by discretization, where models must assemble meaning from many chunks. Transformers embrace discretization but pair (discrete) tokens with attention, which mixes representations by similarity (softmax of dot products), letting global..