CmpE Events

Monday, December 18th

  1. CmpE Seminar: Unsupervised Feature Extraction with Competitive CNNs
    • Start time: 03:00pm, Monday, December 18th
    • End time: 04:00pm, Monday, December 18th
    • Where: AVS Conference Room, BM
    • Speaker: Taner Eskil, PhDTaner Eskil received his BSc in Mechanical Engineering and MSc in Systems and Control Engineering from Boğaziçi University. During his MSc studies he was also employed as a research assistant at Boğaziçi University Pattern Analysis and Machine Vision Laboratory (BUPAM). He completed his PhD in Computer Science at Michigan State University in 2005. His thesis subject was multiple routine design and simulation using off-the-shelf components that are distributed over the Internet. He then joined Sabancı University as a post-doctoral fellow where he lead the Vision and Pattern Analysis Laboratory with funding from EU 6th framework. In 2006, he was appointed by Işık University Computer Engineering Department as a faculty member. Taner Eskil completed one joint TÜBİTAK ARDEB and EU COST project and 2 scientific research projects internally funded by Işık University. He founded Pattern Recognition and Machine Intelligence Laboratory from which 1 PhD and 8 MSc students graduated to date. His research concentration was facial expression recognition until he received his Associate Professorship in 2015. Since then, he has been studying unsupervised feature extraction with multi layered neural networks.  Abstract:The most critical stage in machine learning with Convolutional Neural Networks (CNNs) is the algorithmic training of the filters in the hidden layers. Early approaches such as Fukushima’s Neocognitrons and Hinton’s multi layer generative models focused solely on the inputs to extract a set of representative features. This is intuitive considering the early stages of human development, when an infant tries to make sense of the environment by learning the modes of variation in his/her sensory inputs. State of the art in CNN studies however revolve around supervised training through gradient based algorithms. In gradient based learning (1) the success depends on random initialization of both the number of neurons and their weights, (2) learning stages are vulnerable to the credit assignment problem and (3) training is slow as it requires numerous epochs on samples.I propose add-vote, an unsupervised feature extraction algorithm that has its roots in adaptive resonance theory and Neocognitrons. Add-vote completely eliminates the random initialization stage, avoids propagation of updates in the network, and is extremely fast, converging in only a few epochs in most cases. I will introduce the algorithm and present our early results on 3 different tasks; MNIST handwritten digits data set, BRATS multi modal brain tumor segmentation challenge and textile defect detection.  

    • View this event in Google Calendar

Contact us

Department of Computer Engineering, Boğaziçi University,
34342 Bebek, Istanbul, Turkey

  • Phone: +90 212 359 45 23/24
  • Fax: +90 212 2872461
 

Connect with us

We're on Social Networks. Follow us & get in touch.