Deep Learning Concepts and Logistic Regression with Theano

Summary

The article introduces deep learning, explains its layered approach to learning data representations, and outlines why it may help reduce hand-built feature engineering. It discusses possible quantitative finance applications, including time-series analysis, regime-change detection, and extracting signals from satellite imagery of retail parking lots, oil storage tanks, and crops.

The tutorial series uses Theano to demonstrate logistic regression on handwritten-digit classification and compares CPU and GPU training. The reported MNIST result is about 7.5% test error after 1,000 epochs; on the stated desktop setup, GPU training takes four seconds and is reported as 6.6 times faster than CPU training. Logistic regression serves mainly as an introduction to the software and workflow, with a multilayer perceptron planned as a later model. The article presents finance applications as promising examples, not demonstrated trading results, and emphasizes that time-series modeling requires further care and research.

Key ideas

• Deep learning builds layered representations that can reduce dependence on manually engineered features.
• The article presents satellite imagery as an alternative data source for estimating business activity and commodity supply.
• A Theano logistic regression model classifies MNIST digits and illustrates GPU training.
• The reported classification error does not establish that the method produces profitable trading signals.
• Time dependence and the need for substantial mathematical and programming background are important considerations.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.