JC #6: Computational Neuroscience Papers

[neurocomp2003]

M&B Journal Club: I hope this will be thread #6, if not please rename it appropriately.

Date: First week of Sept 06.
Topic: Computational Neuroscience Papers
Papers: I will attempt to present 2(maybe 3 papers) on the topic above, all models using Neural nets. The first Paper deals with Imagery: Mental Navigation. The second Paper deals with Modelling the HC-Hippocampus.
The third paper deals with language Learning in child development mainly how children begin to learn based on sensorimotor experiences first(though I have not finished reading this article)

All papers can be found on this link

http://www.science.mcmaster.ca/Psychology/sb.html

[Paper 01-Spatial Cog Section] # Byrne, P. and Becker, S. (2004), Modelling mental navigation in scenes with multiple objects. Neural Computation 16(9):1851-1872. PDF document

[Paper 02-Hippocampal Section]# Becker, S. (2005) "A computational principle for hippocampal learning and neurogenesis". Hippocampus 15(6):722-738. (link to pdf)

[Paper 03-langauge]# Howell, S. R., Jankowicz, D., and Becker, S. (2005), A Model of Grounded Langauge Acquisition: Sensorimotor Features Improve Grammar Learning. Journal of Memory and Langauge 53(2):258-276, PDF document

Note: I decided to do a review on all three because originally this would have been my Msc/Phd Research Field. Spatial Navigation & Language.
And I think its a good idea to get a good view of how a range of brain regions would be modeled.

Note: The thread that will follow, will discuss what one should look in a modelling paper.

best, NC

 

[Moonbear]

You expect us to read three papers?! :rofl:

I think this is actually JC #5, but since I'm not sure, I'll leave it as is. And, you're officially "stuck" now, so it won't drift down the page while everyone's busy reading the papers.

 

[neurocomp2003]

Things to note when reading a Modelling Paper.

NOTE- I tend to use a lot of abbrev, so if there is one that I have not defined or is unclear please post it. So that I may define it somewhere. Also note that I will use these names for the papers.
PC or EGO paper- for the byrne paper
HC paper - for the becker paper
Langauge paper - for the Howell paper.

--------------------
ALRIGHT BACK TO THE TASK AT HAND, "Things to note" when reading a Modelling Paper (henceforth I will use the term Modelling synonymously with Computational Science or specifically the neuropsych branch):

Things to look for in any generic computational science model:
[\would like to insert "code"]

[] Scientific Field of Choice - (For these papers-Neuropsych )
[] Programming Langauge -F/C/C++/Matlab/Maple (For these papers-matlab)
[] Spatial Environment -Discrete(Grid) vs Real (For the spatial papers-gridbased) vs No Spatial Environment(eg LanguagePaper)
-2D vs 3D
-How Large is the environment and what datastructures are used to maintain the environment
[] Agent/object Motion -Discrete(Grid or fixed stepsize in Real spatial env.) vs Real(stepsize varies)
- How many agents/objects exist in the environment
and what datastructures are used to maintain the environment
[] BC-Boundary Conditions -For spatial environments, how does the code handle boundary conditions
[] IC Initial Conditions - Depending on the model there are various initial conditions to take into consideration
[] Timestepping - What type, Discrete vs Real
(a) Update according to some smallest movement?
(b) Update according to some fixed global time frame?
(c) Update according to some episode/period/batch?
[] Mathematics involved -Stats(stochastics,markovmodels), DEs/PDEs/ODEs/DynSys(dynamical Systems)

[\would like to insert "\code"]

Things to look for, in a neuropsych model:

[\would like to insert "code"]
DEFINITION "AGENT"- in AI,ALife,Neuropsych the model or models are usually referred to as an AGENT
[] Literature - PROBABLY ONE OF THE MOST IMPORTANT
- Does the model attempt to support some other experimental paper OR predict novel ideas OR both
- Does the paper reference Lesion Studies, NI-NeuroImaging studies(eg MRI/EEG), NI with Drugs(tracking effects of NT-Neurotransmitters), Cognitive study, Cognitive Study of Brain Disorder,Cognitive Study based on some criteria like gender/race/class/educational background, Extension of another models ideas
[] Species - Human Brain, Rat Brain, Monkey Brain.
[] Age group -infancy,toddler,teen,adult
[] Neural Nets - used or not used?
- if used: connectivity type, learning type, Update Rules, Activation Rules, SIZE, Layers & Modules &
subcomponents,
input/output type, temporal rules(refractory period, theta rhythms, time elapse before firing ie. resembles
axon/dendritic length, does it store # of time fired)
Update Process(Timestepping)-Synchronous, Asynchronous,batch/episode-based learning
- {(0,1) OR (ON/OFF)}-system or phys/chem-based (lower level mechanisms like ion channels)
- error analysis

[] brain component/module - ALL(too my knowledge never been attempted, keep an eye on the IBM BlueBrain Project),
the 4 lobes or smaller components(HC,Amyg, PRC/PHC/ERC, SB, PFC,PC:7a/LIP, PVC etc. )
[] brain process - sensory type, imagery,language, navigation, object/pattern recognition etc.
[] Map(spatial environment) - # of maps: one static map, one dynamic map, many static maps, many dynamic maps
[] Input -What does the world input to the agent
Robotics(sensory detectors), Lowlevel virtual vision/audition/sensorimotor detectors(images eg
bitmaps), holistic maps(coordinate locations as input)
- 2D or 3D: in some environments, there exists a birdseye view of a map, an the visual input is thus a 1D line elongated into a 2D visual cue where each row is the same.
[] Output - Does the agent respond to the environment? Movement Only, Manipulation(grasping,kicking etc)
[] Number of agents - Do these agents interact?( eg. game Creatures or any ALife simulation)

[\would like to insert "\code"]

My next post for the discussion will occur the day before 06/09/01.

 

[neurocomp2003]

Wups, that is one long post above. Sorry about that.

Moonbear: Thanks for the Sticky

:tongue: 3 papers is a lot but considering the JC has been outta commission for a while, I thought I'd start it off with a bang
Hopefully this thread [Discussion #6], will run through all of september so that people can take there time to read and discuss the topics. If that happens then I can present the papers 1 week at a time, since I'm still working through some of the math of the first 2 papers. Boo math.

BTW there's JC#5 roaming around below without a sticky. I hope there isn't a 6th somewhere.

Is there any way to enlarge the "code" script window size?
If I insert them in the window it is half the post width. Is there a Tab option? \t?

 

[Moonbear]

I'm not sure how the code script window works. Everytime someone uses it, you seem to have to scroll through it only a few lines at a time. Can you find a way around using it?

Okay, discussing one paper at a time sounds easier. Since we don't seem to have lines of volunteers waiting to present journal clubs here, feel free to present as many in a row as you want.

 

[neurocomp2003]

Neural Net & NeuroPsych Terminology

----------------------------------
Please don't post till after 6pm EST,gives me a chance to double check errors and to complete the second paper.
----------------------------------
In this thread I will be presenting 2 Papers:
The First paper will be presented in the next post.
The Second paper will follow after the first(so two posts after this).
In Post #3, I listed things to look for in modelling paper.

If I do attempt to present the third paper listed in the first post, it'll be sometime in late sept/oct.

NOTE: I would like to take the time in this post to discuss some background info for
Neural Nets & NeuroPsych as I remember them.
-----------------------
Some Cognitive Terms(abbrev in "[]")
-----------------------
[WM/STM] Working Memory/Short Term Memory: What your brain is currently processing,or has recently processed.
[LTM] Long Term Memory: longer then short term, converted from hippocampus by LTP-long term potentiation, basically
what your able to recall from memory after several days or months without cues.

[Nav] Navigation: the ability to move ones body around a spatial environment using visual/sensorimotor cues.
[ObjRec] Object Recognition: ability to recognize shape,color,patterns/textures by visual cues.
[MI] Mental Imagery: the ability to recall from memory without visual/sensorimotor input or cues.
[MI-nav] Mental Navigation: the ability to navigate a spatial environemnt by memory(without visual/sensorimotor cues)
[Allo] Allocentric Reference Frame: Coordinate Frame w.r.t. World Geometry
[Ego] Egocentric Reference Frame: Coordinate Frame w.r.t. to teh self/agent
THere are a coupld Ego frames (Head-Centered,Body/trunk-Centered)
[] Landmarks: Visual Cues such as objects,room geometries that allows one to map out an environment
Global Landmarks(eg CN tower) Local Landmark(eg chair,picture)

-----------------------
Brain Modules/Components:
-----------------------
[PVC] -Primary Visual Cortex, low level processing.
[V2] -visual association areas 2
[V3] -visual association areas 3
[V4] -visual association areas 4
[iT/TE/TEO] -Inferior Temporal Lobe-object characteristics
[mT] -medial temporal Lobe-object location
[msT] -medial/middle superior temporal Lobe-object location[CC/aCC/pCC] -cingulate Cortex: anterior, posterior
[PHC] -parahippocampus, for storage of object location
[PRC] -perirhinal Cortex - object recognition info like texture(not location),
[ERC] -Entorhinal Cortex - component that gives access to the HC below receives info from PHC/PRC
[HC] -Hippocampus, Hippocampal Region(ERC->DG-dentate gyrus,CA3,CA1,SB-subiculum->ERC)[PC] -Parietal Cortex: Areas 7i & LIP(lateral-intraparietal)-mental imagery
[dlPFC]-dorsolateral PreFrontal Cortex- working memory.
[FS/FG] - frontal lobe sulcus/gyrus

-----------------------
Brain Cell Patterns
-----------------------
[] Place Cells- fires for a location in a world geometry
[HDC]Head Direction Cells- fires for a specific direction w.r.t world geometry, visual cues, Earth's magnetic field.

----------------------------------
NNET-Neural Net Terminology
----------------------------------
[] Neuron: The Fundamental unit, in most models a neuron is considered either on/off (0 or 1).Thus we neglect the innerworking of the neural cell physiology.
Also Time delays like refractory period are not considered.
[] Layer: Many Neurons, grouped together for similar behaviour or connections
[] Module: Many Layers, work together to work on a process
[] Nnet: the overall structure
[] Layer Connection Type: Feed Forward(early processing layer to later processing layer), Feed Back(later processing layer to early processing layer), Recurrent(Self-Connecting)
Cyclic-if loops are created much like the Limbic System.
[] Learning Rules: Many kinds, the most primitive is Hebbian Learning. [w+=n*dw,n is the learning rate]
[] Layer Learning Type : Competitive, Cooperative
[] NNet Learning Type : Supervised, Unsupervised, Reinforced

 

[neurocomp2003]

[Paper 01-Spatial Cog Section] # Byrne, P. and Becker, S. (2004), Modelling mental navigation in scenes with multiple objects. Neural Computation 16(9):1851-1872. PDF document
-------------------------------------------------------------------------

Topic: Modelling Mental Navigation in an Environment with Multiple Objects(as seen in the title)

-------------------------------------------------
-------------------------------------------------
Questions you should ask before reading further:
[] How are World Geometries(rooms,boundaries), Global Landmarks(eg CN Tower,Sun), Local Landmarks(objects in the local environment) represented in the brain?
[] How are the above stored in relation to each other?
[] How is one capable of remembering configurations?
[] How does the brain convert between Allo to Ego Coordinates(ie Cognitive BirdsEye View Maps to Relative Mapping) and vice-versa
[] How does the brain convert from Retinal to headcentered to bodycentered Coordinates?
-------------------------------------------------
-------------------------------------------------

The post is outlined as follows:
-------------------------------------------------
[] The Paper: Brief Description
[] Things to look for, as discussed in post #3
[] The Agent & Environment
[] The NNet Model
[] Model Predictions of Navigation
[] Possible Discussion Questions-------------------------------------------------
The Paper:
-------------------------------------------------
The paper explores how the brain stores spatial configurations of numerous objects in an environment. It first discusses the concepts and brain modules involved such as mental-navigation, ego-representation(PC brain regions) vs allo-representation, World Geometries vs Objects, dlPFC for working memory of object location, place cells in HC & head-direction cells in PC.

According to the authors, the process of storing spatial configurations of local objects involves not allo-representation but ego-representation and involves the brain components dlPFC(WM) and PC(7a,IPS:LIP,imagery). The dlPFC or WM stores allo- or ego- locations of the objects whereas the PC regions perform ego-updating as the individual navigates the environment(ie head-direction cells). They suggest through references(GoldmanRakic) that locations are stored in WM if they have been previously processed by the PC regions(7a,LIP) which border the visual cortex areas. Thus an individual must be attentive to an object inorder for that object to move into WM processing. This illustrates the "WHAT vs WHERE" visual dichotomy for the module dlPFC proposed by GoldmanRakic.

The model presented by the authors, extends the Model ideas presented by Droulez and Berthoz(1991) for navigation with the addition of Head-Direction Cells for Spatial-Updating via Ego-rotations about an environment. They illustrate how teh PC computes primitive EGO-motion to store spatial configurations of local objects. Notice that they ignored incorporating WM processing and suggest that it would be rather simple to add this component later. The authors also cite 2 other papers: Shelton &McNamara (2001); Spelke&Wang(2000); and state that any model attempting to perform spatial navigation should at the very least support the results from these papers

[1] Shelton&McNamara show that subject recall spatial configurations more easily when asked to view or positioned at a previously explored location rather than a novel one.

[2] Spelke&Wang illustrate that ego-representation is more important in object configuration whereas allo-representation is more important in world geometries and global landmarks. In the Spelke&wang results, they showed that disoriented subjects performed less accurate in spatial configuration recall. Their Tests were either to remove the individual from their location blindfolded, or to disorient them by making them dizzy(spinning).

As per the papers discussion, their model was able to support the results from both Shelton&McNamara(2001) and Spelke&Wang(2000).

------------------
Things to look for
------------------
[] LITERATURE: Cognitive & neuro-imaging
[] BRAIN MODULE/BRAIN PROCESS: PC, for ego-representation and ego-Updating
[] SPECIES/AGE GROUP: Human & Rat, Adult
[] MATH: Simple 3D Transformations for ego-motion, some stats(gradient descent)-------------------
Agent & Environment
-------------------
[] 2D Gridded Cartesian Map(Discrete not Real)

[] agent parameters: angular & linear velocities, time for learning. Agent is always located at Origin in ego-coordinates with head direction aligned with +y-axis.
[] agent input from environment: Object location, i believe based on either a 2D allocentric map and headdirection, or (x,y). May also be a line representing visual cues. NOTE-no 3D cues(3D Projected to 2D images).
[] contains HDC- Head-Direction cells
[] agent output: activity from main neural net layer, represents recall of object location.
[] agent update method- "Serial Updating": the agent sequentially visits all Objects,rather then using parallel cues.
[] AGENT TEST: Novel viewpoints vs previously viewed or explored Viewpoints. -------------------
Neural Net Architecture
-------------------
[] includes neurons that fire for Head direction .
[] There was no diagram of the neural net architecture
[] learning rule-gradient descent
[] Input Layer(Environment Map?)-Feed Forward to main layer, input layer represents an allocentric map.
[] Main Layer(represents PC): Recurrent, Competitive, 31x31 2D neuronal Layer. The main layer is a competitive layer with bump map activity or winner-take-all firing based
on location or direction. The layer is also topologically organized based .
[] Connections: the weighted connections between layers acts as a ego-transformation. Hence updating Ego-representation in memory.
[] Error source: optimum velocity range for neuronal activity,"random" noise(activity from other parts of the brain),errors
in internal allo-map stored in memory.-------------------
Predictions
-------------------
[1] The reaction time to determine whether a viewpoint was "novel vs explored" is dependent on the # of objects in the configuration. In fact it should be monotonically increasing in Plot of "#object vs Time"
[2] The longer an individual is exposed to an environment, the more accurate they should perform on the tasks seen in
McNamara and Spelke & Wang. An example would be a child's first weeks of exposure to their bedroom.-------------------
Questions
-------------------
[1] How much of parallel cues is stored in memory for object location and world geometry? Or must the individual be attentive to each object/geometric "corner" in the environment(hence serial updating)?
[2] Is it possible to create a cognitive map(allo-map OR birds eye view) through egomotion? Or must the individual view a real allo-map in order to create a mental one?
[3] Does the model suggest that the brain(through child growth) creates layers of Neurons with generic Activity for an Ego-Map and Allo-Map, which rewires for different surroundings based on visual/auditory/sensorimotor cues? That is to say that (a) there exist layers of neurons that fire for a specific angle and distance w.r.t the agents head or body regardless of environment. (b) there exist layers of neurons that fire for a generic NxN(2D) or NxNxN(3D) allo-map, which fire for activity based on cues/memory.
[4] Are these maps topologically organized in the brain(similar to the notion of retinotopic)?

 

[neurocomp2003]

Computational Neurosci Paper #2

[Paper 02-Hippocampal Section]# Becker, S. (2005) "A computational principle for hippocampal learning and neurogenesis". Hippocampus 15(6):722-738.
------------------------------------------------------------------

Topic: Novel Hippocampal Learning Principle & Hippocampal Neurogensis

-------------------------------------------------
-------------------------------------------------
Questions you should ask before reading further:
[] How is spatio-temporal memory stored?
[] What are the difference/similarities between Learning, Recall & Recognition
[] In Child development is there a fundamental mechanism/rules for (a) Neurogenesis, the placement of new [nrn]neuronal cells in the brain
(b) synaptic connections between all cells eg hebbian learning.
[] How does Adult Neurogenesis occur, and can we manipulate it some how.

-------------------------------------------------
-------------------------------------------------

The post is outlined as follows:
-------------------------------------------------
[] The Paper: Brief Description
[] Things to look for, as discussed in post #3
[] The Agent & Environment
[] The NNet Model
[] SimulationMethods
[] Model Predictions of Navigation
[] Possible Discussion Questions

-------------------------------------------------
The Paper:
-------------------------------------------------
The paper has 3 goals
[1] a novel Computing Learning Principle for Learning & Recall in the Hippocampal Region.
[2] a functional-role hypothesis for the neurogenesis that occurs in the Hippocampal:[DG]Dentate Gyrus
[3] a functional-role hypothesis for the recurrent connections of HC:CA3

-----------------
History: Marr's Hippocampal Computing Principles
(a) Sparse Representations-eg winner take all/competitive learning. Few neurons firing for a pattern.
(b) rapid hebbian learning
(c) associative recall-associating one set to another set not necessarily distinct.
(d) consolidation

Marrs Primary Goal of the Hippocampus is Pattern Completion
-----------------

Hippocampal Structure: (note the ERC is the input zone to HC,arrows imply direct connections or info flow)
[] ERC->DG->CA3->CA1->ERC
[] ERC->CA3
[] ERC<->CA1
[] CA3->CA3(Recurrent)
[] CA1->SB->ERC

-----------------
The paper first discusses the importance of the Hippocampus to Neuropsychology, its effect on anterograde and retrograde amnesia of episodic memory and not
other forms of "Learning & Memory(eg. semantic,perceptual, proceduarl,simple conditioning)". It also discusses previous models from 1990-2001, that decompose the
HC Region such that different pathways have different rules for connectivity. Next it describes Hippocampal Activity that leads to her Novel Learning Principle.

HC Activity, described how each region is active in encoding & retrieval(based on input)
xx Encoding(Learning) Retrieval(recall&recognition)
DG ERC->DG ***** Silent,not very active if at all
CA3 DG->CA3 ***** ERC&CA3->CA3
CA1 ERC->CA1 ***** ERC&CA3->CA1

Her Novel Learning Principle is based on the fact that Backprop(requires analysis over many layers) is too slow for associative learning(mappings) in brain processing.
The principle is thus built on the idea that all layers of the HC should somewhat reconstruct the ERC activity, and she provides a Greedy Optimization Technique of Hebbian Learning
for each set of connections between layers that occur in the HC. She also proposes that all other models' Learning rules can be acheived from this principle of Greediness.
The parameters of the Greedy Fashion learning principle are Objective Functions for Learning(that will be maximized on learning trials), size constraints,activation levels, and connectivity constraints.
Finally, because of this attempted remapping of the ERC activity by all layers, this principle can be known as the "Learning Principle of Invertibility". -----------------
The model supports the following:
(a)idea that any multilayer NNET should outperform a single layer
(b)Hippocampal lesion studies in recognition and recall.
------------------
Things to look for
------------------
[] LITERATURE REFERENCED: Cognitive & neuro-imaging & Lesion Studies & drug studies
[] BRAIN MODULE/BRAIN PROCESS: Hippocampal Region(ERC,DG,CA3,CA1)
[] SPECIES/AGE GROUP: Human & Rat, Adult
[] MATH: Matrices, Statistics-------------------
Agent & Environment
-------------------
[] Environment: Binary input(similar to information theory)
[] Agent Parameters: rate of DG turnover; rate of DG Neurogenesis(size);-------------------
Neural Net Architecture
-------------------
[] SIZE: ERC 200; DG 1000 nrns; CA3 300; CA1-400;
[] Multilayered
[] Greedy Hebbian Learning.
[] Many Training Pattern sets

-------------------
Simulation Methods
-------------------
All methods ran the same model with different tests. Note: to create a lesion the wweights of the lesioned portion were set to 0.

[1] Memory Capacity: Recognition & Recall, Lesioned & Nonlesioned.
Result: her model performed to experimental results, that is in lesioned studies, recognition is somewhat impaired and recall is severely impaired.
[2] Memory Capacity: DG size capacity
Result: her model illustrated that small DG sizes really impaired recall and recognition. These small DG sized models sometimes even performed worse than
lesion studies of Method #1.
[3] Memory Capacity: DG neurogenesis
Result: her model showed that neurogenesis can help differentiate between highly similar inputs rather than be used for coding novel inputs. -------------------
Predictions
-------------------
[1] The DG Neurogenesis occurs to store or differentiate highly similar events/states/inputs rather than Kemperman's 2002 proposal for Storing novelty information.
This prediction would help explain the "spaced learning vs massed learning" effect. I'm guessing this is in reference to Time, or temporal learning. That is how frequently
learning trials occur.
[2] CA3 Recurrent Connections are used for temporal associative learning creating an continuous attractor network. This new prediction resulted because
CA3->CA1 pathways seems to suffice for pattern completion(Marr's original hypothesis for these recurrent connections).
[3] Learning Principle based on each layer reconstructing the activation pattern of the ERC.-------------------
Questions
-------------------
[1] Are NNets of small size magnitudes feasible for models of the brain or are they only fun "Information Theory" Projects? will this principle hold up
if we attempted a 3D environmental simulation? It should after all its only information Theory.
[2] How are teh connections between ERC/CA3/CA1 formed in child development if DG is required in encoding but not in retrieval? These connections can be viewed
as invertible connections to the ERC/DG path

 

[neurocomp2003]

Hopefully the above posts are readable and coherent.
Enjoy, NC

 

[Cincinnatus]

I just saw this thread. I'm interested in computational neuroscience and believe I have the background to understand these papers, I'll take a look when I have some more time...

 

[Moonbear]

Neurocomp, sorry for the lack of participation so far...I've wound up quite busy this month and haven't had spare time for reading papers outside my own area. Hopefully I'll have a little time soon to pick up and get more involved in this thread here.

 

[neurocomp2003]

Moonbear: no worries, I've been busy myself, trying to figure out how obtainable a Dual Research Career in N-body Simulations(secondary) and Computational Neuroscience(primary) is since a lot of the coding overlaps.

Forgot I had written this even though its only been a month. I'm guessing that the online journal club has died, or that my point-form writing style is to ugly hehe.