CECN2
From Computational Cognitive Neuroscience Wiki
2nd Edition of CECN
With an expanded team of authors, we are working on a second edition of CECN. Please enter any suggestions below for things you'd like to see in it! Thanks!
Randy O'Reilly, Yuko Munakata, Michael Frank, Seth Herd, David Noelle, Ken Norman
Plans So Far
- Overall: Shorter, sweeter, better graphics, more accessible, more concrete examples of cognitive models earlier, math & nitty-gritty stuff moved to appendix at end of chapter, more build-from-scratch models so people really see what model building is about, and of course updates/new models in each of the domains.
- Procedural details will be offloaded to the simulator entirely (as is the case now with the new Emergent project CECN1 Projects, with a summary of the exploration provided in the text.
- Better questions: clearer distinctions between basic and advanced questions
- Separate chapter on reinforcement learning, "Combo" learning integrated into task learning chapter.
- Separate memory chapters for weight-based vs. activation-based (latter then takes over more of the temporal stuff from chapter 6)
Suggestions for New Features
We'd love to hear what you have to suggest! Enter general comments here, and comments by chapter in the sections below.
- Jude Allred: end each chapter (or begin) with a set of learning objectives. You have tons of content -- a checklist of concepts that should be understood upon completing the chapter would be very helpful
Chapter 1: Introduction
Chapter 2: Neurons
- Randy: Introduce Bucket analogy (and potentially tug-of-war) (see Randy's units slides) early and try to work through all the main properties of the neuron using that. Need a spigot for the top, and a toilet-tank like float thing that turns off the water as the level rises (to simulate equilibrium of excitatory input) -- is it possible to include Na/K pump??
- Randy: start out chapter with a quick first-pass exploration showing overall behavior of the unit, before getting into all the details -- gives people the core essentials. Maybe even put the detector model first, without any parameter exploration or anything -- just a simple demo of what a unit does and how it gets more excited with better-fitting inputs, and the key feature of the threshold. The hebbian learning chapter does this and it seems like a good idea.
- Michael: student suggestion: "One thing I would like to see in the next edition of this text is a table which has the variables described in the equations listed across from the "written out word" e.g. Vm = Membrane Potential, Ec = Equilibrium potential. It's kinda tough to look back in the body of the text for what the symbols mean."
- Michael: Some brief discussion is needed on relationship between model neuron equations with simple integrate and fire equations (typically written as differential eq'ns), in an optional section. I think it's important to relate to other basic computational neuroscience models so that interested readers can put them in context. A good short paper on different types of spiking models is Izhikevich (2004), IEEE Neural Nets). As it is, the models invoke conductances (which simple I&F models do not), but don't try to capture details of action potential (which some spiking models do, even without conductances, cf Izhikevich 2003). A short presentation of the diff eq version would be helpful. (V' = g_e(E_e-Vm) +... ; g_e' = ... etc)
Chapter 3: Networks
- Randy: simplify "transformation" language by just focusing on categorization. Use stereotyping and other salient examples of simplifying categories to illustrate ubiquity of this cognitive process. Use "loose" vs "precise" thinkers as analogy for g_bar.l manipulations? Inhibition analogy: feedforward is "strategic planning" whereas feedback is "tactical" ability to deal with situations as they arise. Appeal to individual differences in personality (strictly as analogy..). Maybe introduce attractor concept earlier instead of later? haven't read this in a while so some of these might be redundant with current text.. ;)
Chapter 4: Hebbian Model Learning
Chapter 5: Error-driven Task Learning
Chapter 6: Reinforcement Learning
- Combination of Error-driven and Hebbian (Leabra) will be moved to end of chapter 5.
- Question: should reinforcement learning come before error-driven task learning, or after??
- More elaborate temporal stuff & SRN's moved to 2nd Memory chapter on Activation-based memory?
Chapter 7: Large-Scale, Areas
Chapter 8: Perception
- Move attn_simple model to the start of the chapter (after biology) and use other sims as detailed elaboration of this overall big picture -- same structure as in language chapter. This model has a lot of cool data etc and is easy enough to really understand, but by the time people get to it they are perhaps overloaded..
- In general it seems that this chapter needs some serious triming and focusing -- lots of complaints about length. The model treatment has already been tightened considerably in the new sims.
Chapter 9: Learning and Memory
Idea: split into 2 separate chapters, one on weight-based memory and the other on activation-based memory and temporal integration.
Chapter 10: Language
- Hannah Snyder re Broca's area:
- not saying that Broca's area is responsible for speech output (while some people still argue for a specific role in syntax, no one argues for a general role in speech production any more).
- not saying that damage to Broca's area causes Broca's aphasia (it is neither necessary nor sufficient-- damage to Broca's area alone causes cognitive control problems, including with language, but not aphasia, while damage to more posterior premotor, insula and white matter can cause Broca's aphasia without any damage to Broca's area).
- I think you could keep much of the same content, but just ascribe the speech output motor sequencing functions you describe to "anterior motor association areas" instead of Broca's area.
