Demystifying the secrets of pre-training

: PSD versus CAE on an industrial use case

Abstract

In industry, printed circuit boards with many small components are assembled using pick and place robots, which require component descriptions for component verification and placement. Descriptions are created manually, which is error prone and time-consuming. To automate the description process, the Automatic Chip Classification and Parameterisation Tool (ACCPaT) was created, which relies on heuristics and neural networks to construct component models. In this paper, the unsupervised pre-training methods Predictive Sparse Decomposition (PSD) and Convolutional Auto Encoders (CAE) are applied to the convolutional neural networks in the tool, to reduce training time and improve classification accuracy of the networks. The effect of both methods in conjunction with supervised training are thoroughly studied. Most significantly, is the factor 2 reduction of training time while achieving even better classification accuracy

Date of Award	28 Feb 2014
Original language	English
Supervisor	B. Mesman (Supervisor 1), M.C.J. Peemen (Supervisor 2), J.L. Horijon (External coach) & Alexander Bock (External coach)