Beatrice van Eden is a part-time PhD student and full-time CSIR employee researching scene identification based on concept learning for robots. She aims to give mobile robots the ability to continuously and autonomously form concepts of their environment based on previous experiences. Her approach involves identifying items in a scene, determining the most feasible scene, and recognizing or contributing to a scene concept. Her related work focuses on papers about object-driven context searching, semantic mapping using object-class segmentation, unsupervised feature learning for scene labeling, and extracting editable objects from photos. The most promising approaches for robotic concept formation and learning from experience involve recurrent neural networks, policy gradients, and optimal ordered problem solvers.
ECG Graph Monitoring with AD8232 ECG Sensor & Arduino.pptx
Scene Identification based on Concept Learning
1. Scene Identification based on
Concept Learning
Beatrice van Eden
- Part time PhD Student at the University of the Witwatersrand.
- Fulltime employee of the Council for Scientific and Industrial Research.
2. Index
• Broad Problem Statement
Give an overview of the research
I am busy with.
• Steering Towards a Solution
Outline the abstract hierarchy that
I consider to bring me closer to the
Solution.
• Related work
Highlights of some of the reading
I did.
iCub baby robot as used in JS' EU project IM-
CLEVER on developmental robotics and on
implementing the theory of creativity & curiosity
& novelty & interestingness on robots.
3. Broad Problem Statement
• How can we give a mobile robot the capability to
continuously and autonomously form concepts
of its environment?
• Humans can identify a kitchen area by recalling the concept formed though previous
experience.
• Certain actions and visible items contributes the concept of a environment being formed and
remembered.
• Machines/robots can exhibit a range of autonomous behaviour but autonomous concept
forming is researched in this study.
4. Steering Towards a Solution
Image of scene
Identify items in
scene
Determine most
feasible scene
Recognise the
scene / contribute
to a concept
5. Related Work - Papers
Searching for objects
driven by context
(Bogdan Alexe and Heess, Nicolas and
Yee W. Teh and Vittorio Ferrari)
Scemantic Mapping Using
Object-Class Segmentation
of RGB-D Images
(Stuckler, J.; Biresev, N.; Behnke, S)
Unsupervised feature learning for 3D scene labeling
(Lai, K.; Liefeng Bo; Fox, D)
3-Sweep: extracting editable
objects from a single photo
(Tao Chen, Zhe Zhu, Ariel Shamir, Shi-Min
Hu, Daniel Cohen)
6. Related Work - Highlights
• How to make robots learn from experience?
• Traditional reinforcement learning (Schmidhuber, CoTeSys)
• Limited to simple reactive behaviour
• Not work well for realistic robots
• Why? Dimensional explosion, continuous spaces
• Robot learning in realistic environments
• Novel algorithms
• Learning to identify important events in sensory stream
• Temporarily memorize them in adaptive, dynamic, internal states
• Most promising approaches:
• Recurrent neural networks, policy gradients, and the optimal
ordered problem solver.
7. Terms
• Recurrent neural networks - is a class of artificial neural network
where connections between units form a directed cycle.
• Policy gradients - are a type of reinforcement learning techniques
that rely upon optimizing parametrized policies with respect to the
expected return (long-term cumulative reward) by gradient descent.
They do not suffer from many of the problems that have been
marring traditional reinforcement learning approaches such as the
lack of guarantees of a value function, the intractability problem
resulting from uncertain state information and the complexity
arising from continuous states & actions
• Optimal Ordered Problem Solver - a general and in a certain sense
time-optimal way of solving one problem after another, efficiently
searching the space of programs that compute solution candidates,
including those programs that organize and manage and adapt and
reuse earlier acquired knowledge.