Graduate Student
Department of Cognitive Science
Univ. of California at San Diego

Exploring Intelligence since 2004

Jump to: Software • Publications • Research Interests • Curriculum Vitae • Personal

Software & APIs

• "Nick's Machine Perception Toolbox (NMPT)," [Website]
  
  Cross-platform C++ API and Software for Fast Visual Saliency (ICRA2008) and MIPOMDP Visual Search (CVPR2009). Easy to install and run: links to OpenCV, but no other libraries. Tested on Mac OSX, Linux, and Windows with Cygwin. Includes full API documentation.
  
  

Publications, Tech Reports, & Talks

• Butko & Movellan, 2009, "Optimal Scanning for Faster Object Detection," IEEE Conference on Computer Vision and Pattern Recognition. [PDF]
  
  We modified the POMDP model from our ICDL 2008 paper to be suitable for searching real images, rather than just abstract simulations and toy stimuli. The modified model takes any object detector as a building block. It constructs a reduced-resolution, foveated representation of the image, and applies the object detector in a limited but effective way to do faster object detection. In this, our initial investigation, we wrapped our algorithm around the Open CV Viola-Jones style face detector, and achieved a two-fold increment in speed with little loss in accuracy.
  
  
• Theocharous, Beckwith, Butko, & Philipose, 2009, "Tractable POMDP Planning Algorithms for Optimal Teaching in SPAIS," IJCAI Workshop on Plan, Activity, and Intent Recognition (PAIR).
  
  We described the problem of teaching as a POMDP with unobserveable physical and social variables. Specifically, social variables affect how a student will behave, and so they alter the transition dynamics of a POMDP. The problem of acting optimally in a POMDP with changing dynamics is very difficult, but we develop a tractable approach for learning near optimal policies quickly by using policies that have been trained for one specific social state (bored, attentive, etc.) as building blocks.
  
  
• Butko, & Movellan, 2008, "I-POMDP: An Infomax Model of Eye-Movement," International Conference on Development and Learning. [PDF]
  
  We adapted a model of human visual search to the POMDP framework, and developed a tractable class of policies that can be learned quickly with simulated experience via policy gradient. These policies yield better simulated visual search performance than optimal one-step lookahead greedy policies.
  
  
• Butko, Zhang, Cottrel, & Movellan, 2008, "Visual Saliency Model for Robot Cameras," International Conference on Robotics and Automation. [PDF] - Accepted Draft
  
  We adapted a class of visual salience algorithms to run very effeciently (much faster than real time -- 10ms per frame), and showed that these visual saliency algorithms were useful in orienting a pan-tilt robotic camera toward people. The algorithm is light-weight enough to allow for considerable post-processing on targeted regions of the image (such as facial expression recognition, etc.), and gives users interacting with the robot the impression that it is an intelligent agent, because it is looking places that are interesting to the users themselves.
  
  
• Butko & Movellan, 2007. "Learning to Learn," Sixth International Conference on Development and Learning (ICDL2007). [PDF]
  
  We used reinforcement learning to frame a "proof of concept" for "learning to learn," i.e. learning to discover the states of hidden nodes in POMDPs as quickly as possible. We showed that in a biologically plausible timeframe, an infant could learn to learn very quickly whether or not a new entity in its environment was a social agent and therefore a potential caregiver.
  
  
• Wiratanaya, Lyons, Butko, & Abe, 2007. "iMime: An Interactive Character Animation System for use in Dementia Care." Intelligent User Interfaces (IUI2007), Hawaii. [PDF]
  
  My role in this project was to use machine learning, computer vision, and reinforcement learning techniques to adapt a stimulus to be more "interesting" to an observer, by automatically observing and analyzing the user's behavior watching the stimulus.
  
  
• Butko 2007, "Generative & Discriminative Naive Bayes." Technical Report, MPLAB.UCSD-1, April, 2007. [PDF]
  
  A reference for the Discriminative Naive Bayes approach to density estimation (also known as Maximum Entropy) as well as empirical evaluations of two optimization packages using the DNB implementation as a test example.
  
  
• Butko & Triesch, 2007. "Exploring the Role of Intrinsic Plasticity for the Learning of Sensory Representations," Neurocomputing, 70(7-9):1130-1138. [PDF]
  
  A journal extension of our 2006 conference paper: In this paper, we explore a mechanism by which the brain might automatically tune itself to represent sensory input in ways that are most conducive to learning.
  
  
• Bajramovic, Mattern, Butko & Denzler, 2006. "A Comparison of Nearest Neighbor Search Algorithms for Generic Object Recognition." Advanced Concepts for Intelligent Vision Systems (Acivs 2006), Antwerp, Belgium. [PDF]
  
  In this paper, we compare different techniques for facillitating efficient learning and inference in model-free methods - classes of algorithms that learn by observation and not by assumption.
  
  
• Butko, Fasel & Movellan, 2006. "Learning about Humans During the First 6 Minutes of Life," Fifth International Conference on Development and Learning (ICDL5), Bloomington, Indiana, USA. [Talk Only]
  
  In this work, we explore possible mechanisms by which human infants (and embodied robotic agents) could learn complicated visual object categories (such as faces) with no prior experience or knowledge, and without any "teacher."
  
  
• Butko & Triesch, 2006. "Exploring the Role of Intrinsic Plasticity for the Learning of Sensory Representations," European Symposeum on Artificial Neural Networks (ESANN'06), Bruges, Belgium. [PDF]
  
  In this paper, we explore a mechanism by which the brain might automatically tune itself to represent sensory input in ways that are most conducive to learning.