Prediction with Short Memory
Speaker: PreetumTitle: Prediction with Short Memory
Date: 09 Apr 2018 5:30pm-7:00pm
Location: Pierce 320
Food: Chinese food
Abstract: I’ll discuss the paper “Prediction with Short Memory”, which asks the question: How much of the past do we need to remember to predict the future? (ie, how complex do predictive models really need to be). For example, for a Hidden-Markov-Model with N states, they show that remembering the past O(log N) observations is sufficient to predict the distribution of the next observation (with small average L1 error vs. the optimal estimator). This talk itself will be brief and require only short memory.