CS 294-4: Connectionist and Neural Computation

Term Projects

One half of your final grade will be based on your term project. There are many exciting projects available on campus and at ICSI and scientists eager to have your help. We also encourage you to seek a project that furthers your own interests and fits the course. It is fine to work in teams of two and it is often interesting to team with someone with a different background.

• Extensions of the Tlearn system - mainly to support structured nets
  
• Simplifying large trained networks Prof. Morgan
  
  
• We spent a full lecture on Regier's system for learning spatial terms. The program is still around and a good project would be to resurrect it and extend it any one of a number of ways.
  
• Training Bailey's verb-learner on additional languages
  
  
• Implementing and testing Bailey's recruitment model
  
  
• Augment a search engine using metaphorical matching
  
  
• Link the human-body simulator "Jack" to Bailey's verb-learner
  
  
• Redo and update Regier's system for learning spatial terms
  
  
• Add animation to the Bailey or Narayan system
  
  
  
• Learning first-order context-sensitive rules from examples using structured connectionist nets.
  
• The following five term-projects relate to SHRUTI --- a connectionist model of temporal binding and (rapid) reflexive reasoning
  
• Develop an X-Schema for posing wh-questions to SHRUTI and extracting the answer computed by SHRUTI.
  
• Refinement of domain knowledge via backpropagation and hebbian learning.
  
• Integration of reflective (or deliberate) processes with reflexive processing.
  
• Demonstrate that binding via temporal synchrony/desynchrony can be an emergent property of suitably interconnected connectionist models.
  
• Modeling how attention can overcome certain limitations of reflexive reasoning by (i) "focusing" the spreading of activation into certain directions and (ii) having improving the tuning of synchronous activity.
  
• Simulating errors in decision making due to stress.
  
• The following three term-projects relate to SMRITI --- a model of episodic memory formation in the hippocampal system.
  
• Modeling of de'ja` vu.
  
• Modeling "forgetting" --- both because of "decay" and because of earlier memories being "overwritten" by subsequent memories. The goals are to explain: (i) primacy and recency effects, (ii) why irrelevant memories are more likely to be forgotten than likely ones (though, highly relevant memories may sometimes be forgotten).
  
• Study the effect on episodic memory of diffuse cell loss in the entorhinal cortex. One possibility would be the modeling of memory loss characteristic of early stages of Alzheimer's Disease.
  
• Detailed quantitative analysis of episodic memory capacity and the formation of spurious memories.
  
  

Pointers to some on going projects at ICSI:

SHRUTI: Connectionist models of reflexive reasoning. This project attempts to understand how networks of simple and slow neuron-like elements can encode a large body of knowledge --- containing several million items --- and perform a wide range of interesting inferences within hundreds of milliseconds.

SMRITI: A Model of Episodic Memory Formation in the Hippocampal System. Our ability to remember events and situations in our daily life demonstrates our ability to rapidly acquire new memories. This project attempts to demonstrate how a transient pattern of activity representing an event may be rapidly transformed into a persistent structural encoding via long-term potentiation (LTP) and depression (LTD) of synapses.

Speech research at ICSI has primarily been focused on speech recognition and auditory processing, with three main specializations: auditory-inspired signal processing, statistical modeling, and models of phonology and language for speech understanding.