<style> .slides{ background:inherit; } code { overflow: hidden; } </style> # The state-space forumulation for time-series data Mike Ricos Slides at: https://hackmd.io/@mricos/HJdJrDdnu --- ## Summary - Probability is about counting. - Conditional probability is about context. - State-space formalization is about change. - Linear regression example is about trend. - When you don't know the likelihood.. - Relationship to LSTM / Transformer --- ## Background - Probability - Frequentist v/ Bayesian - Electrical engineering and control theory - Kyrre Lekve's original lecture --- ## Conditional probability - Venn diagram of Bayes Rule - Conditional probability - Conjugate variables --- ## State-space formalization --- ## Linear regression example --- ## When you don't know the likelihood --- ## Relation to Long Short Term Memory models --- ## Relation to Transformers --- ## MCMC Code ```python= docs = [] # a doc is a tuple of (string, president) for file_name, president in zip(files, presidents): with open(f'./inaugural/{file_name}') as f: text = f.read() docs.append((text, president)) ``` ``` In [6]: type(files) Out[6]: list In [7]: type(docs) Out[7]: list In [8]: type(docs[0]) Out[8]: tuple In [9]: type(docs[0][0]) Out[9]: str In [10]: type(docs[0][1]) Out[10]: str ``` --- ## References - [Bayesian modelling of time series](https://www.uio.no/studier/emner/matnat/ibv/BIO4040/h03/undervisningsmateriale/Lectures/lecture10.pdf) - lecture slides by Kyrre Lekve. - study-groups - twitter - nodeholder ---