𝗚𝗿𝗮𝗱𝗶𝗲𝗻𝘁 𝗱𝗲𝘀𝗰𝗲𝗻𝘁 𝗶𝘀 𝗮𝗹𝗹 𝘆𝗼𝘂 𝗻𝗲𝗲𝗱 𝗳𝗼𝗿 𝗶𝗻𝘁𝗲𝗿𝗽𝗿𝗲𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆? In our ICML paper, we show that it might just be. ExPLAIND is a method that unifies data attribution, model component attribution, and training dynamics by computing an exact decomposition of model behavior based on gradient products.
