Introducing ... Broca's area

So far I’ve looked at general cortical circuits, not specifically at language regions. Today I’ll take a closer look at one of more specifically language related areas: the Broca’s area. Broca’s area deserves a special attention not only because of historical reasons but also because of the clear left-right asymmetry. The Broca’s area expands Brodmann’s cytoarchitectural Areas (BA) 45 (anterior) and 44 (posterior). This 45/44 division roughly corresponds to Economo’s cytoarchitectural division FDG/FCBm and neuroanatomical regions PTr/POp (Pars Triangularis / Pars Opercularis).

Both BA45 and BA44 have leftward volumetric asymmetry, and the asymmetry is affected by handedness (Foundas et al. 1998). The volume fraction of cell bodies in areas 44 and 45 (Amunts et al. 2003) and the size of layer III pyramidal neurons in area 45 (Hayes and Lewis 1995) are shown to be greater on the left side. There are extensive horizontal connections in supragranular layers (I-III) and to a lesser extent in infragranular layers (Tardif et al. 2007). Compared to the visual cortex, there are less connections from infragranular layers to supragranular layers (Tardif et al. 2007).  These two facts (horizontal and inter-layer) may suggest extensive topological processing in the feature space [according to my imagination].

BA45 is characterized by a larger and more uniformly granular layer IV compared to BA44, which is characterized by a dysgranular layer IV invaded by numerous pyramidal neurons (Brodmann 1909). Since BA45 is more granular than BA44, it is predicted that feed forward connection tends to originate more from supragranular neurons in BA45 and terminates in infragranular layers in BA44. It is also predicted that feedback connection tends to originate more from infragranular neurons in BA44 and terminates in supragranular layers in BA45.

Adams RA, Shipp S, Friston KJ. Predictions not commands: active inference in the motor system. Brain Structure and Function. 2013 May 1;218(3):611-43.

Amunts K, Schleicher A, Ditterich A, Zilles K. Broca's region: cytoarchitectonic asymmetry and developmental changes. Journal of Comparative Neurology. 2003 Oct 6;465(1):72-89.

Barrett LF, Simmons WK. Interoceptive predictions in the brain. Nature Reviews Neuroscience. 2015 Jul 1;16(7):419-29. -> Barbas et al. diagram “cortical layer infragranular supragranular M2 M1 connection flow”

Brodmann K. Vergleichende Lokalisationslehre der Grosshirnrinde in ihren Prinzipien dargestellt auf Grund des Zellenbaues. Barth; 1909.

Foundas AL, Eure KF, Luevano LF, Weinberger DR. MRI asymmetries of Broca's area: the pars triangularis and pars opercularis. Brain and language. 1998 Oct 1;64(3):282-96.

Hayes TL, Lewis DA. Anatomical specialization of the anterior motor speech area: hemispheric differences in magnopyramidal neurons. Brain and language. 1995 Jun 30;49(3):289-308.

Tardif E, Probst A, Clarke S. Laminar specificity of intrinsic connections in Broca's area. Cerebral Cortex. 2007 Dec 1;17(12):2949-60.

  

Brain Grain

In the last blog I described the canonical six layer cortical structure. That was intentionally a very simplified description. For one thing, the cortical layers include multitudes of cell types the distribution of which differ among different areas. There certainly should be non-systematic variances among brain regions that are shaped by evolution. However, one systematic citoarchitectonic variation in the brain is granularity, ranging from granular (clear layer IV and dense neurons) to dysgranular to agranular (lacking layer IV). The first figure below shows the granularity distribution in the human brain (Beul and Hilgetag 2015, based on Economo 2009). In the sensory modality, granularity seems to decrease as the paths move from the primary to association to higher order association cortices. In the motor modality, it may seem the opposite, as granularity increases from primary to association to higher association cortices. This however is also natural, since in the sensory modality the primary neural pulse propagates from sensory organs, whereas in the motor modality the primary neural pulse propagates to motor organs.

There have been some important observations made in terms of inter-layer connection patterns. One is that the cortico-cortical connectivity pattern in terms of the source and target layers differs depending on the difference of granularity between the source and target areas in non-human primates (Barbas 1986, Barbas and Rempel-Clower 1997). When the granularity difference is large, the connections go either from neurons in the shallow layers of the higher granularity area to shallow layers in the lower granularity area, or to the opposite direction (Figure 2). When granularity difference is small, the source and destination layers are more distributed along the depths. Also, it is observed in the rodent cortex that there are less inter-laminar inhibitory connections in less granular regions (Beul and  Hilgetag 2015). More specifically, in the most granular area (striate), there is inhibition V/VI->(IV and II/III), IV->II/II. In a less granular area (somatosensory), less long-range inhibition to the shallow layer, thus V/VI->IV, IV->(II/III and V/VI). In agranular area (primary motor), no clear inhibitory interlaminar connections were found.

So this sounds simple enough, but just in case you’re itching to hear more about neural modeling at this point, Beul and  Hilgetag (2015) presented neural models based on literature survey on inter- and intralaminar connections as well as prior modeling efforts. In an agranular, rodent frontal cortex model (right column in Figure 3), there are recurrent interlaminar connections V/VI<->II/III and intralaminar inhibitory connections. In a granular, cat striate cortex model (left column in Figure 3), there are linked interlaminar loops VI->IV->V->VI  and VI->IV->II/III->V->VI, intralaminar inhibitions, plus interlaminar inhibitions V->II and IV->II. These recurrent connections may serve amplification, gain control, and normalization (Beul and  Hilgetag 2015).

Fig 1. Granularity gradient

Fig 2. Between-area connectivity patterns

Fig 3. Within-area connectivity models (left:granular area, right: granular area)

References

Barbas H. Pattern in the laminar origin of corticocortical connections. Journal of Comparative Neurology. 1986 Oct 15;252(3):415-22.

Barbas H, Rempel-Clower N. Cortical structure predicts the pattern of corticocortical connections. Cerebral Cortex. 1997 Oct 1;7(7):635-46.

Beul SF, Hilgetag CC. Towards a “canonical” agranular cortical microcircuit. Frontiers in neuroanatomy. 2015 Jan 14;8:165.

von Economo C. Cellular structure of the human cerebral cortex. Karger Medical and Scientific Publishers; 2009.

Hello world

This blog is about language related brain regions and neural network modeling of natural language processing. First I want to know more about the brain before going into the neural modeling business. I’ve lived long enough to see at least two, if not three, waves of neural networks “booms”. First, the Perceptron (Rosenblatt). Second, the Back-Propagation (Rumelhart, Hinton, Williams and possibly more preceding them), and now Deep Learning. The third wave will probably fade soon, unless we address the basic problems with current neural network models (eg. Marcus 2003 – more on this in a later blog) and learn more deeply from the biological brain. And most importantly, I don’t know about you but I want to know how the brain works, and the language may be a window to the human mind, granting a limited logical access to the dark, huge space of human psyche.

Ok, so the linguistic brain.  I want to start more specifically at the architecture of the cortical layers in general, since I find it’s a bit boring to start at Brodman’s area and sort of block diagram level, although we’ll eventually get there. The cerebral cortex consists of up to six layers, numbered from the cortical surface towards the deeper layers towards the brain stem. The naming varies but it’s easy to remember in four groups as I (molecular layer), II-III (external granular and external pyramidal layers), IV-V (internal granular and internal pyramidal layers), and VI (multiform layer).

The figure below is a diagram of cortical cells and connections taken from a review paper by Harris et al. (2013) and grossly simplified by picking up only prominent connections. According to the paper, the primary input feeds mainly to layer IV. The higher order input reaches layers I, V, VI.

Cells in the layer II/III projects to layer V and outputs to higher order and contralateral cortices.  The cells in layer IV projects to all layers but most strongly to layer II/III. The cells in layer V can be classified broadly into two.  The first class cells are located relatively shallow and project to layers II/II and ipsilateral and contralateral cortex and striatum. The cells that belong to he second cell type in layer V are the primary output cells. The layer VI includes at least two cell types. One has long horizontal axons. The other has slow corticothalamic projection that reaches reticular and sensory thalamus, and also projects to interneurons in layer IV.

The diagram might still look complicated, but it really isn’t if you pick up a path from one external input to one external output, for instance, higher order input to higher order output, which I would be interested if I want to model higher language related cortices. It at least looks like a good start, I hope.

References

Marcus GF. The algebraic mind: Integrating connectionism and cognitive science. MIT press; 2003.

Harris KD, Mrsic-Flogel TD. Cortical connectivity and sensory coding. Nature. 2013 Nov 7;503(7474):51-8.