Notes initially from The Hippocampal and Parietal Foundations of Spatial Cognition
- 1 Chapter 1: Introduction
- 2 Random Questions
- 3 Chapter 2:Spatial Frames
- 4 The Organization of Spatial coding in the Hippocampus
- 5 Memory for places: A Navigational Model
- 6 Spatial Learning and Localization in Rodents
- 7 Spatial View Cells and the Representation of Place
- 8 Hippocampal place cells connected by hebbian synapses
- 9 A Model of Hippocampally dependent Navigation, using the tD rule
- 10 A model of spatial map formation in the hippocampus of the rat
- 11 The hippocampus as a cognitive graph
- 12 The Effects of Changes in the environment on the spatial firing
- 13 Place Cells, Head Direction Cells, and the Learning Landmark Stability
- 14 Cues that Hippocampal Place Cells Encode
- 15 Cognitive Maps Beyond the Hippocampus
- 16 A functional hypothesis for adult hippocampal neurogenesis
- 17 A computational Principle for Hippocampal Learning
- 18 Neurogenesis may relate to some but not all types...
- 19 Multiple representations in the hippocampus
Chapter 1: Introduction
- There are three main areas involved in spatial processing: the prefontal cortex, the parietal cortex, and the hippocampus
- The Parietal Cortex and the hippocampus both receive highly processed data, mostly from the visual pathway.
- evidence for the hippocampal role in spatial processing comes from either single unit recordings in movign rats, or the effects of hippocampal lesions in rats.
- The place cells fire as a result of sensory input (vision, smell, whiskers) and when the rat/primate sees certain landmarks a group of place cells fire. Those place cells alone can not pinpoint a location, the individual place cells cover a far too large area to be that percise. but a collection of a few hundred place cells could determine a percise location. because the PFs are so large, ppl claim that PCs are active in different environments. this is misleading b/c it makes the reader think that their PFs are very percise and active in different regions, whereas PCs are probably tuned into a certain area and whenever a stimulus enters their PF they go off. The question becomes what form of input do they recieve. Clearly the fact that only certain visual cues affect PCs indicates that the highly processed input to the Hippocampus is filtered into essential elements.
- The location of PFs has nothing to do with the animals attention to a particular location.
- The PFs can change shape and size dynamically
- complex spikes are actually pyramidal cells
- Fast theta cells are actually interneurons
- Things like a single white card dominating PF organization indicates that the higher brain regions break the visual input down in a systematic fashion.
- When a rat's PCs have been set up, turning the lights on or off has little to no affect.
- Perhaps CA3 has two different types of neurons to ensure that input goes to pyramidal cells, and not interneurons which play some other role
- The place cell theory was derived b/c no other simple theory really holds.
- almost all the models involve using sensory input into a multi layer neural network. The more ambitious ones have tried to incorporate head direction cells and theta EEG.
- place cells only exist in CA1 and CA3
- What benefit would there be in having a distributed encoding scheme?
- Perhaps it's not distributed, but represents a different form of data
Chapter 2:Spatial Frames
The Parietal cortex is composed of two parts, the anterior and the posterior, which are further subdivided into two regions. The anterior part is divided into the postcentral gyrus and the caudal bank. The posterior part is divided into and upper and a lower lobule by the intraparietal sulcus. For more in depth jargon see page 5 of The Hippocampal and Parietal Foundations of Spatial Cognition
models of parietal and hippocampal interaction
Mnemonic processing on Two Time Scales In this case, both regions store the same type of spatial information but at different time constants. Essentially the hippocampus takes in all the spatial information but retains it for a short period of time, whereas the parietal cortex abstracts the information and stores it over a long period of time. Ideally the hippocampus would store the information then pass it onto the parietal cortex to abstract out, however, there is very little evidence to support this.
Spatial Processing in Different Frames of Reference' In this model, the two regions of the brain are considered to code for two independent streams. The Parietal cortex deals with the dorsal stream by forming an egocentric view by determing what objects are present. This fits nicely with the evience tha thte parietal cortex is involved with orienting limbs. Conversely, the hippocampus deals with the ventral stream by determing where objects are located, and thus forming an allocentric spatial map.
A Coplimentary Spatial and Mnemonic Role for the Parieto-hippocampal system According to this model, egocentric views would not be ideal for long term memory since activation of that memory would depend on the body being in that exact same position. Therefore, an allocentric view would be ideal for long term memory. Therefore the hippocampus and the ventral stream would be best suited for long term memory.
The Organization of Spatial coding in the Hippocampus
Not a Model
- direction, speed and turning angle all influence the firing rate of rat PCs
- PCs have no isomorphism with environment they are mapping
- The hippocampus could be massively parallel where each part of the environment is simoltaneously calculated
- The hippocampus has PCs which have overlapping PFs
- As it turns out, the PCs have much more overlap than had they randomly been assigned to a location in space
Three results were discovered about local ensembles of PCs
- A small set of PCs cover a substantial portion of the environment
- PCs overlap more so than if by chance
- Partial overlap between each cluster.
- PCs with overlap had similar preferences for movement speed, direction and turning angles
- This is in direct contradicition with O'Keefes theory of non-overlapping PCs
- Overlapping PCs increases spatial resolution, and makes it easer to locate oneself in allocentric space
- assumes there is an egocentric map located in the neocortex (my guess is the parietal cortex)
- The hippocampus stores snapshots of this egocentric map and acts as a database. When a new environement is encountered, the hippocampus is queried to see if it has been encountered before
- The hippocampus uses auto-associative memory to recall previously visited environments
- Proves that the auto-associative model of Marr is also consistent with the spatial properties of the hippocampus
- Neurons in the EC have bigger PFs
- The egocentric map is stored in the neocortex (parietal cortex) and the hippocampus acts as the autoassociative memory store.
Spatial Learning and Localization in Rodents
- An allocentric spatial map is basically where other objects are relative to me
- Location based spatial allocation is just metric or graph based where you have an origin, and everything is portrayed in relation to that.
- Relation based captures the relationships amongst any two points...and is more or less useless
- rats tend to learn places as opposed to stimulii
- short landmarks are generally not remembered because they become obscured by tall objects
- The locale system makes the spatial map and I believe it is carred out by the hippocampus
- The taxon system carries out path finding and I believe that it is not carried out by the hippocampus
- The Dg has granuale cells
- The CA regions have pyramidal and interneuron cells.
- PCs are active in multiple parts of an environment
- PCs are active in multiple environments
- PCs use ensemble encoding; a set of PCs code for one place.
Spatial View Cells and the Representation of Place
- Primate Spatial View cells are better for episodic memory because you remember what you see, not where you are in 3D space
- the primate hippocampus recieves input from the Entorhinal cortex
- the EC recieves input from the parahippocampal gyrus and the perirhinal cortex
- the PG and PC recieve input from the Parietal, Prefontal and temporal cortex
- The hippocampus also projects back to these cortical areas via the subbiculum, the EC and the Parahippocampal gyrus
- There is convergence in the forward direction, and divergence in the backpropogation direction
- There are both alocentric and egocentric neurons in the primate hippocampus
- Some cells respond to body motion, for example some cells respond when their body is rotated along the verticle axis
- Rats may develop place cells b/c of the weaker visual system and the reliance on distal visual cues and olfactory reliance.
- Representation of space could be useful in memory formation since it could also be used in path finding.
Hippocampal place cells connected by hebbian synapses
- Bases distance on connection strength between neurons
- very crude model, it even states that there is no evidence that the hippocampus actually stores distances in this way.
- provides navigation as well, which IMO is silly b/c there is no evidence that the hippocampus carries out navigation
- Presents a paradox where the PCs have enough info to form a spatial map, but not enough to guide navigation
A model of spatial map formation in the hippocampus of the rat
- Uses LTP in the hippocampus to code for a location
- this shifts the location coded by there ensemble activity away from the actual location.
- These shifted maps provide navigation
- this model speculates future navigation.
- PCs only found in CA1 and CA3
- Location determined by averaging an ensemble of place cells
- They claim that the shifted representation arises in the CA3 since it has recurrent connections
- The CA1 receives the shifted rep. as well as the original signal from EC therefore it has the current (true) location, and future (shifted) location.
The hippocampus as a cognitive graph
- EC to DG displays LTP
- EC to CA3 displays LTP
- EC to CA1 displays LTP
- DG to CA3 displays LTP
- CA3 to CA3 displays LTP
- CA3 to CA1 displays LTP
- They believe that the recurrent connections play a huge part b/c a class with recurrent connections has more in common with 2D space than an omnidirectional circuit.
- The arguement they give is that in a two layer omnidirectional graph, only paths of length one can be found.
The Effects of Changes in the environment on the spatial firing
- Looks at the responses of place cells when animal's environment changes
- Recordings were made of 160 CA1 cells
- rotating the cue card had an equal rotation of the PFs
- rotating the wall(s) also had an equal rotation of the PFs
- The exact details of the card (color, texture, thickness) were irrelevent
- Changed size of cue card by 50% ellicited no change in the PFs
- removing the cue card just rotated the PFs by an arbitrary angle but DID NOT change the PFs. This implies that the card is sufficient but not necessary to anchor the angular coordinate
- Increasing the area by four fold only increased PFs by a factor less than four. they concluded that more research is necessary
- It is impossible to guess the firing fields within a rectangle by looking at the firing fields within a cylinder
- inserting barriers only affects cells in the immediate vicinity. this is also true of a transparent barrier
- Knowing the firing patterns in one situation does not help to guess them in another situation
Place Cells, Head Direction Cells, and the Learning Landmark Stability
- Kind of a bizarre experiment, it looks at the stability of place cells amongst two groups of rats. those that have recieved training under non dissorienting conditions, and those that haven't
- One observation was that a cue could only be used as a reference point if it proved to be stabla, otherwise the hippocampus tends to neglect it.
- Head direction cells can also be controlled by cues in the environment, but also operate in the absence of cues
- This ability of head direction cells to orient themeselves in the absence of local cues leads the authors to conclude that it could form a form of internal compass
- The tetrodes made contact with CA1 and the thalamus
- When the rat had no prior dissoriented training, the RFs remained stable and referenced on the cue
- When the rat recieved prior dissorientation training, the cues had weaker control over the PCs and the head direction cells
- Theta cells also turned out to be spatially biased.
- Most of the time, the rotation of the head direction cells, was matched by a rotation of the PCs
- it is unknown whether these two are dependent or just correlated.
Cues that Hippocampal Place Cells Encode
- The main premise is that hippocampal neurons encode a hierarchical representation of the environment
- Evidence suggested that PFs responded to remembered spatial relationships when important cues were eliminated.
- As distal stimuli were removed or impaired, local stimuli took over
- As local stimuli were removed or impaired, distal stimuli took over
- Double rotation trials led to a completely new remapping totally unrelated to either the local or distal cues, which lead them to believe that the encoding was based on both local and distal cues.
- Distal cues were often sufficient but not necessary for PF activity
- If local cues are not stable, PFs will shift towards distal cues.
- PCs controlled by local cues were generally unaffected by double rotations
- PCs controlled by distal cues generally followed one single distal cue, or switched to local cues after double rotation trials.
- According to them, precedence is given such that distal and local cues (together) are favored first, then distal cues, finally local cues.
- It seems that the hippocampus stores redundency, and in the presence of noise/fucker moving a cue, switches to a different representaiton.
- The hierarchical encoding of stimuli is proposed because the stimulus or set of stimuli that controlled a place field change from a preffered to an alternative set when the organization or the content of the environment changed.
Cognitive Maps Beyond the Hippocampus
- a very comprehensive literary review
- The rodent hippocampus can locate the rat to within 1 cm
- author believes that Path integration in the rodent H is accomplished through a loop including the H proper, the subiculum, the parasubiculum, and the superficial layers of the EC
- Current head direction is maintained in the postsubiculum and updated via a loop including the anterior thalamic nuclei
- The rodent determines whether an envioronment is new or not through pattern separation in the DG and CA3
- PCs are also found in the superior layers of the EC, the DG, the subiculum, and the parasubiculum
- The H recieves input from high level sensory cortices
- Head direction cells have been located in the postsubiculum, anterior thalamic nuclei, and the lateral dorsal nucleus of the thalamus
- Reference frames in rodents are a view that the H encodes goals, and each goal has a different reference frame which is described through place codes.
- Rodents only operate in one reference frame at a time
- Author believes that the nucleus accumbens drives locomotion
A functional hypothesis for adult hippocampal neurogenesis
- Claim that new neurons help prevent catastrophic interference (two environments being coded to the same code). The old neurons are thought to be stable, it's the new ones which are extrememly plastic and respond to new environments
- if the hippocampus is lesioned no new facts or episodes can be stored in long term memory
- The EC sends output (to H) via layers II/III and recieves input (from H) via layers V/VI
- The fact that the DG has more neurons than the EC and a lower firing rate implies that the DG implements a sparse representation
- New neurons form synapses more rapidly
- Most new neurons die after neurogenesis, the ones that survive can last for up to a year
- U can deal with catastrophic interference by interleaving the different inputs, but in a real environment this is impractical
- The way the hoppfield network deals with catastrophic interference is by having a really sparse representation such that for any environment, only a few neurons code for it
- Even with sparse representation, eventually u will get so many different environments that u get collisions. u can deal with this by gradually getting rid of old memories, or u can assume that the capacity is large enough (as is the case with the brain)
- Ppl generally think that CA3 is the storage site of the H b/c it has reccurent collaterals so it resembles the hopfield networks
- catastrophic interference here comes in two forms: too many environments and too much resemblence amongst the environments
- To reduce the resemblece you code just recode and sparsify the input. this is thought to be carried out by the DG since it is the input to CA3 and it has a sparse representation. It is assumed that the DG implements lossy compression by discarding the information that is useless.
- The CA3 is thought to encode the input, and the CA1 is thought to decode it.
- Therefore, the DG takes the input from EC, sparsifies or compresses it, then gives it to the CA3
- they propose that neurogenesis prevents catastrophic interference by adding new neurons for newly encountered events.
- CA3 does not need new neurons since hoppfield networks are immune to catastrophic intereference. the CA1 also need no new neurons sinc eit only adapts to the DG, therefore, only the DG need new neurons
- in their model, the DG is the encoder, the CA3 is the memory, and the CA1/Subiculum is the decoder
A computational Principle for Hippocampal Learning
- claim that neurogenesis in the DG is for creating memory traces for highly similar items.
- they believe that all regions of the hippocampus use the same learning principle
- Since DG cells far outnumber EC cells, we get a very sparse representation in the DG
- They believe that the trisynaptic path?? is used for encoding (probably DG-CA3-CA1-Sb-EC) whereas the preforant path that bypasses the DG is used for decoding...which makes sense since the DG is used to present a sparse encoding to the CA3.
- Unlike Marr who proposes the H is an autoassociation net, these people say that the H links spatiotemporal representations.
- They believe that the H implements some form of greedy algorithm
- The H is thought to be a convergence zone from virtually everywhere else in the brain, and projects back to pretty much all of them
Neurogenesis may relate to some but not all types...
- They checked to see if reduction in neurogenesis affect hippocampal functions (spatial navigation, fear response, spooning)
- The DG with only 1-2 million cells adds 250,000 new cells every month!!
- Fear conditioning went down
- Spatial navigation didn't go down for some fucked up reason!! One possibility is that spatial learning involoves new neurons, but it can still function w/o them...in other words the experiment was a POS
- The fact that new neurons die very quickly might be because of their time limited role in learning
Hi Paskari, I'm also doing research in this field, but on the behavioural side. Sorry for spamming your page, but just thought you might want to know that there's some contention about the methods used in this paper. For some reason, rodents treated with MAM/irradiation to ablate neurogenesis have impairments in fear conditioning but NOT spatial navigation (this applies to papers by Shors, Saxe, etc). On the other hand, genetically modified rats/mice seem to show impairments in spatial navigation just fine. Bloody weird if you ask me. Anyway, interesting stuff you have here, have to admit I lol'd at this particular note
Multiple representations in the hippocampus
- Talks about the H representing multiple different environments, but I just interpret this as sparse coding.
- Claim that the firing of place fields can determine not only the rats location but the unique environment it is in