Skip to main content

Developing and delivering neurocomputational models to bridge between brain and mind.

Periodic Reporting for period 3 - BRAIN2MIND_NEUROCOMP (Developing and delivering neurocomputational models to bridge between brain and mind.)

Reporting period: 2018-09-01 to 2020-02-29

What is the problem/issue being addressed?
The promise of cognitive neuroscience is truly exciting – to link mind and brain in order to reveal the neural basis of higher cognitive functions. This is crucial, scientifically, if we are to understand the nature of mental processes and how they arise from neural machinery but also, clinically, if we are to establish the basis of neurological patients’ impairments, their clinical management and treatment. Cognitive-clinical neuroscience depends on three ingredients: (a) investigating complex mental behaviours and the underlying cognitive processes; (b) mapping neural systems and their function; and (c) methods and tools that can bridge the gap between brain and mental behaviour. Experimental psychology and behavioural neurology has delivered the first component. In vivo neuroimaging and other allied technologies allow us to probe and map neural systems, their connectivity and neurobiological responses. The principal aim of this ERC Advanced grant is to secure, for the first time, the crucial third ingredient – the methods and tools for bridging systematically between cognitive science and systems neuroscience.

Why is it important for society?
The strategic target of this ERC Advanced grant is to initiate a new drive towards the ultimate goal of cognitive neuroscience: to specify how neural machinery synthesises cognitive function and dysfunction after brain damage, at a computational-mechanistic level. The multi-disciplinary work will be focussed on the neural basis of language and its disorders (aphasia and acquired dyslexia). The results will be relevant to both basic and clinical disciplines (including psychology, linguistics, cognitive neuroscience, neuropsychology, speech pathology, behavioural neurology, etc.). The topic focus of language is socially and strategically important given that language impairments are a common, disabling feature in many types of dementia (including Alzheimer’s disease and the progressive aphasias) and following stroke (1/3 patients acutely and 1/5 in the chronic phase) – diseases which are becoming ever more common in Europe’s ageing societies. The 2012 European Brain Council study on the costs of brain disorders estimated that, as at 2010, 6.3m people in the EU suffer from dementia and 8.2m people from stroke (with estimated, conservative total health costs of €105b and €64b, respectively). All future projections indicate a worrying escalation in the rate and costs of these two diseases.

What are the overall objectives
This ERC Advanced grant contains three work packages, which run in parallel. All three workpackages (WP) run for the entire duration of the award. Each workpackage represents a different type of research rather than a different time period. The three workpackages are interlinked with information arise in WP2 and WP3 feeding into WP1.The three interlinked workpackages are the same as in the proposal; they are:

Work Package 1 – Neurocomputational modelling: building computational models of language functions; linking these to cognitive neuroscience data; linking them to neuropsychological patient data.

Work Package 2 – Understanding how language functions arise from the underlying neural machinery: using combinations of cognitive neuroscience techniques – fMRI, MR tractography, rs-fMRI connectivity, TMS, etc. to explore the neural bases of healthy language function including identification of key regions and their connectivity.

Work Package 3 – how damage and recovery in neural systems give rise to acquired language impairments (aphasia) after brain damage: building up a large scale patient database and using this to explore the lesion correlates of different language impairments, of the connectivity/disconnections, and changes in function with respect to recovery.
Excellent progress has already been made in all three workpackages. In many cases we have not only conducted the research but we have also submitted and published papers as well. This progress is summarised in the sections below.

Work Package 1– Neurocomputational modelling:
Building computational models is highly demanding, yet we have made excellent progress. Our original neurocognitive models of semantics and language did not consider alternative computational architectures and how these interact with the nature of the representations to be learned and the interaction with cognitive-executive control mechanisms. We have, therefore, completed three sets of simulations to explore (i) the impact of the depth on a model of semantics (how many hidden layers); (ii) the interaction with different densities of representation; and (iii) comparison of models with different architectures but balanced computational resources.

Secondly, we have generated a bilateral model of language and recovery. Specifically, we have generated the first-ever computational model of a bilateral ‘dorsal’ language pathway model (for left and right hemispheres) which is trained to repeat words. The model generates behavioural responses and can also a proxy for fMRI BOLD imaging and analyses. The model successfully learns to repeat and can generalise this to novel items (nonwords) - which is a key ability of human phonological language systems. Using this framework we have explored how varying the level of computational resources across the pathways leads to a division of labour in the contribution of each pathway to spoken output and how this influences performance after left vs. right pathway damage. Like real patient, the model shows chronic aphasia after left hemisphere but not right hemisphere damage.

1.2.2 Work package 2 – understanding how language functions arise from the underlying neural machinery:
We have made excellent progress on developing analyses of white-matter tractography in order to explore anatomical connectivity maps and how these change when brain regions are removed. Our initial study is a proof of concept study in which different parts of the anterior temporal lobe are removed from the brain images of healthy participants in order to check the resultant pattern of disconnections. We have also continued to work on exploring the relationship between structural (white matter) vs. functional connectivity and relating these to cognitive functions. We also had the objective to explore individual differences and the impact this has on the resultant effect of TMS. This has now been down with respect to the varying contributions that word meaning has when reading aloud.


1.2.3 Work package 3 – how damage and recovery in neural systems give rise to acquired language impairments (aphasia) after brain damage:
A major and continuing target is to build up the number of patients in our chronic post-stroke aphasia database and to increase the density of the behavioural data. We continue to make excellent progress on both. The database has risen from 31 cases to 85, and we have added a whole new set of behavioural data on verb production and comprehension. This makes it one of the most detailed databases on aphasia on the world. We are now adding to this database through new recruitment in Cambridge focussed on extending both the chronic cases and tracking recovery from the acute stage. This database has supported a range of studies and a collection of published papers.
As well as tackling all stated targets in the ERC funded project, we have also taken the chance to go beyond these when exciting opportunities have arisen. These split into two sets – further basic cognitive neuroscience and clinical translation.

Further basic cognitive neuroscience:
(A) As well as contemporary tractography based studies, we have also explored the historical accounts of human white matter neuroanatomy from post mortem studies. This includes the seminal work of the Dejerines who produced the first formal investigations at the end of the nineteenth century. This work was published in French and we provided the first English translation with an accompanying opinion article.
(B) As a part of exploring the semantic function of the anterior temporal lobe, we have explored the region’s response to various social and non-social concepts. We utilised these data in the explorations of patients with temporal lobe epilepsy (see WP 3) but also to examine how these types of concept map into the anterior temporal region.

Clinical translation:
The proposal did not list specific impact targets. However, we have three broad areas of impact activity. The first is through academic and clinical dissemination of results and materials. Our excellent publication output is listed above which provides the major, worldwide dissemination of results. We supplement this with presentation to national and international conferences and clinical workshops (e.g. Lambon Ralph was the invited keynote speaker to the recent Clinical Aphasiology Conference in Salt Lake City, Utah – where he presented work from WP3). We also make available all new materials, etc. such as the new verb semantic battery to clinicians and researchers.
Secondly, with a view to clinical impact, we are current progressing with using our basic science findings to generate new clinical prediction models (see Section 1.2.3.a).
Thirdly, we have been able to utilise our large patient database and the findings from their detailed behavioural and neuroimaging assessments, to explore novel speech therapy interventions. The first of these explored a new approached to treatments for aphasic word-finding difficulties. Traditionally, therapies focus on improving and restoring item accuracy which can be achieved but patients’ rarely exhibit carry-over of the therapy effect to their own spontaneous connected speech. We tackled this problem in a new way by focussing not only on naming accuracy but also speed. We were able to show that across the patient group (N=20) that there was improved naming accuracy and remarkable levels of carry over to connected speech if the therapy focussed on accuracy followed by increased naming speeds.