Juyang Weng (翁巨扬), Professor
Department of Computer Science and Engineering
Mental development is an autonomous, open-ended process during which an agent (human or machine) interacts with its environment (including humans) through its life span, guided by what is called the developmental program (i.e., the genome or an equivalent). A major difference between traditional programs and a developmental program is that the latter is not task-specific, i.e., the tasks that the agent will learn are unknown or not fully predictable during the programming (or birth) time. There is no lack of computational models that characterize the computational architecture of the brain at different degrees of detail. However, there is a lack of architecture models that incorporate perceptual signal processing, cognitive (logic) processing, and the developer of the processors. Understanding properties of mental architectures is of fundamental importance for understanding neural processing systems, their adaptation and developmental learning.
This lecture systematically reviews key properties of mental architecture using existing major studies and models as examples, including perceptual architectures (e.g., Neisser’s two stage visual processing scheme; Feldman & Ballard’s 100-step rule; John Tsotsos' model of immediate vision, HMM models), cognitive architectures (e.g., Soar proposed by Laird, Newell & Rosenbloom, ACT-R by Anderson, and the architecture outline by Albus), motor architectures (e.g., the subsumption architecture by Rodney Brooks and others), value system architectures (e.g., reinforcement learning, Q-learning, and other recent more complete models). Based on recent results from neural science, psychology and computational intelligence, this tutorial further explain a series of properties for higher biological mental architectures along with the corresponding architecture components. Architecture examples are used to illustrate such architecture properties. The series of architectural theory explains how a neural system that does not contain any pre-defined symbolic internal representation can be autonomously developed (through programming, prenatal growth and postnatal experience from environment) to deal with not only perceptual tasks such as recognition and classification, but also sensor-driven higher cognitive tasks, such as abstraction, logical reasoning, thinking, planning, and language acquisition and understanding.
1. Biological body development and mental development
2. The history of biologically motivated architectures
3. Review of animal learning theories, nonassociative and associative learning, classical conditioning, instrumental conditioning, time sequence learning, cognitive learning
4. The brain, the cortex, and the cortical laminar architecture
5. Supervised, reinforcement, communicative learning, and the refined 8 learning types
6. Perceptual processing: retina, LGN, visual cortex as visual processing examples
7. Cognitive processing: recognition, classification, and invariance
8. Motor processing: rehearsal and coupling of effectors
9. Sensorimotor pathways and their development
10. Motivational system: limbic system, intention, value, and their development
11. A hierarchy of mental architectures: non-observation-driven, observation-driven, attention selective, rehearsable, self-aware and self-effecting, developmental, multi-level
12. Example architectures and experimental studies
General programming experience, basic
knowledge about vector and matrix operations.
Audience: researchers in biological neural systems, signal processing, image processing, computer vision, pattern recognition, speech recognition, autonomous navigation, autonomous control, language processing, robotics, human-machine interface, and artificial intelligence.
Handout: Tutorial material will be provided by the host institution.
Biographical sketch of the lecturer:
Juyang (John) Weng
received his BS degree from