Summaries

29-Jan Top-down | 5-Feb More models | 12-Feb default network | 19-Feb Clay Curtis as guest |

26-Feb Language and Orbital PFC | 5-Mar Orbital damage | 12-Mar Corey McMillan as guest | 2-Apr (Dorso)lateral PFC |

29-Jan MODELS: PFC as top down modulator
Doug Bemis

The discussion today focused upon two potential functions of the PFC: the top-down modulation of primary sensory areas resulting from prediction of upcoming stimuli and "cognitive control". Building primarily upon articles from Bar (2007), the PFC has been hypothesized to be constantly forming predictions about upcoming stimuli given the current and preceding surrounding context. These predictions are thought to arise from learned associations between temporally related stimuli. The occurrence of the one stimuli in such an association is then used by the PFC to activate relevant associations that can predict the forthcoming, associated stimulus. The mapping from current input to relevant associations in performed by an analogy mechanism that is underspecified in Bar's papers. The result of this associative activation is the top-down modulation of primary sensory areas. In some manner, these areas are primed by the PFC for future stimuli. This theory is possibly relevant for the focus of the class as prediction, in a general sense, is believed to be a crucial part of language comprehension. Also, the localization of prediction by Bar and semantic composition by our lab is similar. This raises the question of how these two processes might be related.

Cognitive control has many definitions. The general essence seems to be the selection of relevant information. Usually this is couched in a framework that enables the inhibition of competing actions. A canonical example is the ringing of a phone. It is argued that this stimulus is automatically linked to the action of answering the phone. However, this automatic response must be overridden if the phone is not yours. This overriding is cognitive control. Cognitive control is important to investigate as it, in one form or another, dominates much of the frontal cortex literature. Relative to language, it is mention often in models developed to explain parsing strategies in the face of garden path sentences. For our lab specifically, many responses to our results are along the lines of "Oh, that's just cognitive control." Therefore, we should know what that might mean.

Here are the discussion questions from the class:
1. Difference between top-down and bottom-up processing
2. Does mechanism / neural localization of prediction depend on stimulus content?
3. Relation between prediction from context and prediction on the basis of "gist"
4. What do we mean by analogy?
- How does this differ from association
5. Precise mechanics of bar-model
6. What's the format of the representations manipulated by the PFC?
- How are they manipulated, and where do they live?
7. Role of novelty in these computations?
- At what point is the PFC no longer needed? How learned does it have to be?
- What counts as uncertainty in language perception?
- Maybe stimulus is fuzzy?
- What does it mean to "predict from context"?
- Should predictions effects result from highly-constrained contexts (where prediction is possible), or highly-variable, when prediction might be more needed?
8. Does feedforward processing co-occur with top-down processing?
- related to 5.
9. Evidence that PFC represents?
- Is there a notion of "temporary" representation that's useful?
10. Directionality of effects of successful prediction
11. How does LSF-idea translate to other domains?
12. Working in advance vs. monitoring

Suzanne presented some background research in prediction and its relation to language processing. This work focused primarily upon EEG results from Federmeier demonstrating early, differential effects relative to how predictive the surrounding context is of a given sentence ending. These effects also seem to vary by hemisphere. In general, they suggest that prediction causes increased activation at earlier time points. Suzanne and Hugh's research also demonstrates an early prediction effect, showing an increased M100 amplitude when expectations are violated. These results raise the issue of what it means to predict. Is it a preactivation of lower-level neuronal networks? Does prediction in general entail increased or decreased activation as a result, and how does this vary by model. A possible extension of the prediction models onto our AMF results might be that that earlier peak, around 200msec, reflects prediction mechanisms responding to "gist" information in the stimulus, and the later peak, around 450msec, reflects prediction from the context. It remains to be explained what exactly qualifies as "gist" and "context" information for these predictions and what the resulting predictions might be.

Finally, we discussed cognitive control briefly. Many theories seem to assign "cognitive control" as the general function carried out by the entire PFC, and so it is rather difficult to begin to map this in any substantive way onto our findings. One lead is that Miller (2000) suggests the ventral-tegmental area responds to rewards and signals the anticipation of rewards following a stimulus. However, it is difficult to figure out how this might be applicable to the AMF. It could be that the AMF reflects the suppression of competing semantic representations, rather than an integration process. However, the lack of a correlation between the behavioral effect of coercion and the amount of suppression necessary (McElree) argues against this.

5-Feb MODELS: Temporal integration, Adaptive coding, Structured events
Suzanne Dikker: click here for a summary.

12-Feb PFC in the brain’s “default network”
Hanna Gelfand

The discussion today explored the brain’s “default network,” a system of regions that displays more activity during rest than during some cognitive tasks. Further, we investigated the related regions in the context of our current research because the locus of our AMF effects seems to overlap with the default network region. More specifically, the medial prefrontal cortex (MPFC), which roughly includes both the ACC and OFC, are most relevant for us. Notably, activity in the default network does not decrease just from stimulation. Activity decreases when the brain is involved in a working memory task or a task with some kind of cognitive load. Given that the AMF generator seems to be part of this network, then its function should be something that could conceivably operate during “rest.” Additionally, we can start to think about how the data that we have on the role of the MPFC can restrain theories of the default network.

We then delved into several possible theories that hinged on the following important claims: A central role for the PCC is that it is involved in memory retrieval. The literature strongly tries to attribute this region to retrieval of autobiographic material. Also, the MPFC seems to take part in some sort of combinatory function. Hypothesis #1: PPC performs memory retrieval
Hypothesis #2: MPFC combines representations into more complex representations. Thus, “mind wandering,” as proposed by Mason et al. (2007), may be a process that involves composing more complex “memories.” Also, if this structure really does combinatory processing in many different domains then input would have to be something other than your autobiographical information.
Hypothesis #3: In many situations, the input to the MPFC is something other than PCC memories. In this situation the pathway away from the PCC to the MPFC is suspended

We briefly discussed the inverse correlation between the PCC and task relevant regions, concluding that the asymmetry between the PCC and prefrontal regions makes sense because the MPFC is something that gets used a lot in tasks, while the PCC does not.

We then moved on to discuss the controversial nature of theories about the default network. Gilbert et al. (2006) argues against the claim that the MPFC participates in a network of regions that support “stimulus independent thought.” Gilbert et al. argue that the data equally supports the hypothesis that in the absence of a stimulus you have heightened attention to the environment. Thus, the effect is due to increased attention to the external environment.

Discussion Questions from this class: 1. Why is the DFN suppressed during tasks? In reference to this question we discussed whether the following assumption was really “fair:” Each of the regions that participate in the DFN are involved with other processing and kick in when a stimulus is absent. Stemming from this assertion we wondered whether a network of regions that only operated during rest existed.
2. Importance of neuropsych evidence to locus of function
3. Self-projection/prospection special?
- We noted this as a uniquely capable human ability that may have a link to linguistic thought about semantics and possible worlds. 4. Does the role of the DFN reduce to retrieval?
5. Relation between language and the DFN
- DFN activity may be related to inner “speech” to the extent that semantic representations are thoughts.
6. PFC implicated for goal-directed behavior – mind wandering seems goalless
7a. How to relate the TOM findings and linguistic manipulations?
7b. Is the relevant notion of TOM to language:
1) Imagining yourself in the situation of the expression?
2) Constructing a mental model that matches the speakers?
- Uncontroversially, the latter process must occur during language comprehension, but does this involve the same TOM as something like “self-reflection?” We defined TOM as imagining yourself in someone else’s head and self-projection as imagining yourself in “you-prime” or rather another version of the reality. We concluded this idea wondering what remembering tasks drive the network where “self-reflection” is involved and asserting that it is easy to make self-projection into the future look like TOM but it is much harder to do so with remembering.

We ended class discussion on the following note: literature on the activity of the default regions correlates to performance on tasks. In other words, more activity in the default network results in better performance.

19-Feb Guest speaker: Clay Curtis
Madelaine Krehm

In today's class, we talked about general theories of prefrontal function and we discussed how they might relate to language. Professor Curtis discussed the motivation for Curtis and Esposito (2004). The debate surrounding the functional organisation of the prefrontal cortex was heavily influenced by early animal lesions studies, but hadn't included a full review of the literature. An extensive review of the older animal lesion data provided no support for the dorsal and ventral pathways specialising in spatial/nonspatial information (Goldman-Rakic) or qualitatively different operations (Petrides). Although this data could have been acquired through neuro-imaging techniques, this paper underlines the value of a review of older animal studies. The theories that were prominent at the time have fallen out of favour. We talked about how to unpack tasks into computations. What are the representations being maintained during the delay in match to sample task? Does the initial stage in this task activate a preexisting representation, or does it create a new one? What is impaired is the maintenance of a location. There is a coordinate system, possibly with gross retinotopic organisation in the central sulcus. Space is mapped effectively despite the gross level of representation because one point maximally activates the receptive field. The closer to this point, the more activation there is.

Does the principal sulcus store spatial location? Just behind principle sulcus is the arcuate, which is where eye movements originate . It is organised by motor topography (nearby space = nearby eye movement). It is also possible that the arcuate encodes the plan to move eyes from x to y instead of the movement itself.

This organisation has implications for the spatial vs. object tasks. In the spatial delay task, animal knows what action he'll perform as soon as he sees the stimuli. But in object task the animal can't plan ahead of time. The tasks are meant to measure spatial and non-spatial abilities, but also may tap into action and non action.

Representation
Bistable cells have 2 stable states- the downstate, with low, stochastic levels of firing, or upstate where there is no mechanism to keep the cell active, but the firing remains very high until some trigger to end the state. There is a relative increase in density of bistable cells in PFC. In computer models, the on switch to activate an upstate is fed forward from posterior locations. The off switch could be the organism's response. In object tasks, if the cell has a preference for the object, it will activate. But objects seem to be less integral than space. How similar are the working memory demands in each type of task? Spatial tasks are much easier to train in monkeys. But within object tasks, non match to sample tasks easier to train. This may be because of an exploratory/novelty preference?

Are there duplicates of object representations in the frontal lobe? Is this really feasible for everything, eg. words? Do cells of PFC store information? Clay thinks that this is unlikely, and that the frontal representations are more akin to top down tags. Parietal areas have more precise representations. The PFC is a poor place to store information because the receptive fields are so gross. But there needs to be some mechanism to keep parietal cells active. Is it possible that bistable cells from the PFC could maintain the activation in the parietal cells? Parietal cells would not be continue to fire, but the cell membrane would become more excitable. It's possible that the connectivity between the parietal and frontal cortex is part of the mechanism that maintains the activation in the parietal lobe. This theory is generally thought about in reference to space.

Inna asked about cross modal integration and the cells in the PFC. Professor Curtis explained that audition is craniotopic, and is sensed in terms of the head not the eyes. The parietal cortex is involved in the translation from audio space into a retinotopic organisation. There aren't separate cells for audition and vision- some cells in the parietal cortex respond to both. Lizabeth Ramonski is researching a monkey model of communication. She shown monkeys videos of monkeys communicating while recording from superior temporal sulcus or ventrolateral PFC. Some cells in these areas respond to the auditory stimuli, some respond to the visual, and some respond differentially if audio and visual tracks matches. Professor Pylkkanen asked how these frontal cortex theories could be applied to semantic complexity and integration and how representations get built. Professor Curtis responded that this feed-forward model for spatial representations might not work for other things. There is no feedforward topography to the PFC. The PFC has a multifunctional spatial map that may not exist for other stimuli. We know that a dynamic map of where we represent things in the environnment exists in the parietal cortex There have been similar results studying the frontal cortex, but they are always messier. The Miller and Coen paper we read a few weeks ago was not a great example of this: why would monkeys have cells for shapes or stamps? This could be a result of overtraining? The monkeys may have begun with representations in the parietal cortex, but over training created representations in the PFC. Would the experiment work with novel objects, without establshed connectivity? What happens over the course of training? This may work as an analogy for language, because it may be overtrained. But there are some problems with this idea- How do cells know where to feedback to in posterior cortex? How does connectivity exist? How is info combined in terms of connectivity in PFC? Another theory is that Broca's area is an anterior-posterior gradient of abstraction, ranging from phonemes to words to syntax to discourse

Humans can do both spatial and object tasks with lesions (exhibited in both TMS studies and stroke patients). One of the first PET studies tried to get dorsolateral activation, but instead got activity in parietal and orbitofrontal cortex. Could it be that the areas in the human brain have been pushed back as compared to their monkey brain analogues as the PFC developed? This is possible, but would not explain how posterior the relevant areas are as indicated by both activation and lesion studies indicate. It is also possible that the majority of the recordings in monkeys actually were actually further back than the articles indicate.

The discussion also touched on other possible candidates for maps in the PFC, including colour and syntactic categories. Professor Curtis responded that he thought there were simply too many things to map, and that visual variations in colour do not correspond to a map-like activation.

Professor Curtis brought up another theory of PFC functional organisation. What if the ventral/dorsal pathways don't correspond to spatial/object processing, but to action/perception? This seems to fit the data better. One part selects actions, one percepts. Patients with damage to ventral pathways can do action, but have trouble judging orientation, indicating a spatial deficit, but representing the action. This theory is competition for the where/what dichotomy

We moved onto the articles focusing on human data
Doug asked about the differences in monkey and human brains. Professor - no one has tested higher order stuff with monkeys. Professor Pylkkanen wondered whar the starting point behind this idea of hierarchy in action was. The idea comes from the apraxia and planning literature. Badre's pyramid of abstraction does not collapse if a middle area is lesioned- deficits are shown at that level of abstraction only. Badre was able to predict where the lesions were using behavioural data.

Inna and Professor Pylkkanen pointed out the similarity between the frontal cortex theories of the 2 human papers. Do more abstract rules always exist for a longer amount of time? The main difference seems to be the temporal aspect. Where are bistable neurons found? They are more frequent near arcuate, when near principle sulcus- but even at their most dense they make up only 5% of neurons. Suzanne noted that the time based theory was similar to Fuester's model of temporal integration.
Professor Curtis talked about planning in terms of cell activation- neurons from 2 possible futures get activated. But how many possible futures can be planned for?

To sum up, Professor Curtis said that he thought that theories of the PFC are more plausible if they emphasize the action hierarchy. Parsing frontal cortex is very important- until we can subdivide into it into areas of study, it will be hard to look at. A lot of detail is lost if brains are averaged together instead of compared by area within individual frontal cortices. Researchers either focus on a very specific or a very global level when studying the PFC because so little is known. It hard to identify subdivided areas aside from central sulcus, because the anatomy is so variable between individuals. The Principle sulcus in non-human primates doesn't really map onto a single area in human brain.

5-Mar Consequences of Orbital PFC damage
Cristina Rabaglia: click here for a summary.

• Corey's slides

The primary topic for the class was to determine what the role of the dorsolateral prefrontal cortex (DLPFC) might be in language, if any.

The main non-linguistic proposals are that the DLPFC enacts working memory (i.e. maintenance and not manipulation of representations) or it performs cognitive control (e.g. the selection of actions, representations, responses, etc.)

The primary support for these non-linguistic hypotheses is derived from single-cell primate research. However, Rowe et al. (2000) conducted an fMRI study in which subjects first viewed a dot display for 15 sec. Then, they had to keep the display in mind during a blank interval before moving a target to a position previously occupied by a dot. (The appropriate position was indicated by a line across the screen - the target position was then the place at which the line intersected a previous dot position). The results of the study were that maintenance (i.e. the interval between the display and the response) showed increased activation in the left parietal lobe, while selection (the response phase) showed increased BOLD response in the DLPFC.

Our question: Can we explain the DLPFC findings for language in terms of selection (or maintenance)? One concern is that previous findings seem to indicate that Broca's Area (BA) performs selection during language comprehension. Thompson-Schill hypothesizes that BA performs selection among competing representations, initially among semantic alternatives. Perhaps these results can be assimilated under Badre's topology, in which the lateral PFC deals with increasingly abstract representations along the posterior-anterior gradient. Alternatively, the posterior-anterior dimension might access different time scales, with longer time scales appearing more anteriorly (c.f. Rowe et al.) and shorter time scales appearing more posteriorly (cf. Thompson-Schill).

A quick selection of DLPFC effects in language revealed the following. Balaguer et al. found that irregularly inflected verbs produce more activity in the DLPFC compared to regularly inflected verbs. This result is consistent with a selection hypothesis, if irregular verbs require selection among alternatives (i.e. the incorrect, rule-generated regular form). Novais-Santos et al. showed increased activity for subordinate argument structures, as compared to dominant argument structures. For example, the verb 'asserted' can have as its argument either a DO ('asserted the belief readily') or a subordinate clause ('asserted the belief would be justified'). However, it is more often found with DOs, and so, in this case, the DO is the dominant structure, while the SC argument is the subordinate structure. This finding is also consistent with a selection hypothesis, as selecting a subordinate structure plausibly involves suppressing the more dominant structure. On a more general note, Thompson-Schill pitted a 'high-selection' verb generation task against a 'low-selection' version and found more activity in the DLPFC. In the 'high-selection' condition, subjects were asked to generate a verb in conjunction with a noun not strongly associated with a particular action, such as 'wheel', whereas in the 'low-selection' condition, the prompting noun was associated strongly with a single action (i.e. 'scissors'). This result was then argued to generalize to more abstract selectional tasks such as assessing whether an apple was heavy. There was additionally something involving a seagull.

However, not all studies seemed to be consistent with the selectional hypothesis. In particular, Caplan et al. tested the comprehension of subject extracted clauses versus object extracted clauses, and independently varied the amount of 'constraint' present in the relationship between the two arguments of the sentence. For example, 'The policeman that arrested the thief' is an example of a constrained subject extracted clause, while 'The lawyer that irritated the banker', is an example of an unconstrained subject extracted clause. Presumably this division is intended to capture a measure of the cloze probability of the second argument. Caplan et al. report increased DLPFC activation for unconstrained object extraction vs. unconstrained subject extraction, supporting a working memory hypothesis for the DLPFC. However, the other contrast to drive the DLPFC was constrained subject extraction minus unconstrained subject extraction. This result is difficult to interpret, as it appears to be exactly opposite of the effect predicted by a selectional hypothesis. Another difficult to interpret study involved TMS. This study, by Manenti et al., zapped subjects in the frontal lobe and found that they performed semantic tasks slower after left zaps and syntactic tasks slower following right zaps.

The interim conclusion then seems to be that the majority of the considered evidence is consistent with a selectional hypothesis, and several pieces of evidence seem to argue against the working memory hypothesis. For example, Rowe et al.'s previous localization of working memory in the parietal lobe was replicated by Novais-Santos et al. However, Caplan's subject vs. object extraction result does provide some support for the working memory hypothesis.

The last point of discussion was to attempt to refine 'selection' and 'maintenance'. At what level of representation do they operate? Do they create / embody representations themselves, or do they simply manipulate them from a distance. If so, what would that mean, functionally, and in a representational sense? I believe the question was left somewhat open.