With 13 years of experience in building high impact data science products such as customer segmentation, chatbots, recommendation system, time series modeling, anomaly detection and leading technical sales for machine learning in Europe for a large consulting organization, he evolved into the role of a senior digital strategist. Now, he assists businesses in improving their advanced analytics practice to get actionable insights from data and discover monetization opportunities.
We need to learn the representation of statistical objects hidden in data so that we could do classification and aggregation tasks on them. Mostly these representations are learnt by providing constraints in forms of labels. How could we learn these representations without labels? What should be the property of these representations? Existing methods are Autoencoders which are non-linear PCA, clustering, Autoregressive models
But can we be creative?