06-03, 16:30–17:10 (Europe/London), Warwick
AutoEncoders (AEs) are among the most popular techniques in modern machine learning. Thanks to their strong representation learning capability, they can be used not only to generate data, but also for many other tasks, e.g. clustering, dimensionality reduction and transfer learning.
Despite their popularity, their application is usually advertised mostly for applications with static tabular (e.g. for recommender systems) and image data (e.g. for computer vision tasks). With this talk we will try to shed some light on a less well-known area of application, namely the use of AEs with time series data. After a brief introduction on AEs we will highlight challenges to their application in the time-series domain with a particular focus on clustering, features extraction and transfer learning.
The talk is for everyone with an interest in deep learning, time series and their intersection. Despite some working knowledge of applied machine learning (deep learning in particular) and time series analysis would be beneficial, the talk will be delivered in a format accessible to all data science practitioners.
The talk has one main goals: compare different families of AutoEncoders to time series data to show some possible data science applications.
In particular, we will consider two tasks: unsupervised time series clustering, and transfer learning for time series prediction. The AEs results will be compared with the ones obtained using standard techniques.
For this talk, we will rely on an open-source dataset: such as the one from the M5 Forecasting competition (https://mofc.unic.ac.cy/m5-competition/). After the talk, you will gather general knowledge about AutoEncoders, along with their strengths and weaknesses, and their application to Time Series data.
No previous knowledge expected
Data Scientist at Tesco