Page 64 - Flytxt
P. 64
which are one of the most valuable within the data stream seen so far.
assets of enterprises focusing on AI. AutoML systems should leverage this
In the context of big data, data often feature to bring in more completeness
arrives in streams in real time. This and actionability. Another area AutoML
brings in the possibility of data being could be made more effective is by
deviating from the normal i.i.d case storing its previously learnt knowledge
and hence exhibit concept drift, which (ML pipelines or Model Configurations)
could make the previously built model pertaining to different tasks, data sets,
ineffective to be used anymore. In such and domains, and applying it intelligently
a scenario, model needs to be capable to discover an optimal initial pipeline to
enough to detect concept drift and start with, given a new task and data set.
adapt to it automatically without manual In other words, AutoML systems should
intervention. Bayesian online change be able to learn from its own historical
point detection algorithm is one such experience in a lifelong learning setting.
prominent algorithm which can detect
change points (which characterize One of the active research areas pursued
the concept drift phenomena) by Flytxt is to build a generic and scalable
by probabilistically modelling the AutoML framework which can provide
distributional variances of features the above-mentioned capabilities.
64 INSIGHTZ - VOLUME 03, 2018
