Stephen R. Welbourne

Why bilateral damage is worse than unilateral damage to the brain

Human and animal lesion studies have shown that behavior can be catastrophically impaired after bilateral lesions but that unilateral damage often produces little or no effect, even controlling for lesion extent. This pattern is found across many different sensory, motor, and memory domains. Despite these findings, there has been no systematic, computational explanation. We found that the same striking difference between unilateral and bilateral damage emerged in a distributed, recurrent attractor neural network. The difference persists in simple feedforward networks, where it can be understood in explicit quantitative terms. In essence, damage both distorts and reduces the magnitude of relevant activity in each hemisphere. Unilateral damage reduces the relative magnitude of the contribution to performance of the damaged side, allowing the intact side to dominate performance. In contrast, balanced bilateral damage distorts representations on both sides, which contribute equally, resulting in degraded performance. The models ability to account for relevant patient data suggests that mechanisms similar to those in the model may operate in the brain.

Using Parallel Distributed Processing Models to Simulate Phonological Dyslexia: The Key Role of Plasticity-related Recovery

PMSP96 [Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson, K. Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56115, 1996, Simulation 4] is an implementation of the triangle model of reading, which was able to simulate effects found in normal and surface dyslexic readers. This study replicated the original findings and explored the possibility that damage to the phonological portion of the model might produce symptoms of phonological dyslexia. The first simulation demonstrated that this implementation of PMSP96 was able to reproduce the standard effects of reading, and that when damaged by removal of the semantic input to phonology, it produced the kind of frequency/consistency interactions and regularization errors typical of surface dyslexia. The second simulation explored the effect of phonological damage. Phonological damage alone did not result in a convincing simulation of phonological dyslexia. However, when the damage was followed by a period of recovery, the network was able to simulate large lexicality and imageability effects characteristic of phonological dyslexia-the first time that both surface and phonological dyslexia have been simulated in the same parallel distributed processing network. This result supports the view that plasticity-related changes should be a significant factor in our understanding of chronic behavioral dissociations.

A comparison of dual gradient-echo and spin-echo fMRI of the inferior temporal lobe

Magnetic susceptibility differences at tissue interfaces lead to signal loss in conventional gradient‐echo (GE) EPI. This poses a problem for fMRI in language and memory paradigms, which activate the most affected regions. Two methods proposed to overcome this are spin‐echo EPI and dual GE EPI, where two EPI read‐outs are serially collected at a short and longer echo time. The spin‐echo method applies a refocusing pulse to recover dephased MR signal due to static field inhomogeneities, but is known to have a relatively low blood oxygenation level dependant (BOLD) sensitivity. In comparison, GE has superior BOLD sensitivity, and by employing an additional shorter echo, in a dual GE sequence, it can reduce signal loss due to spin dephasing. We directly compared dual GE and spin‐echo fMRI during a semantic categorization task, which has been shown to activate the inferior temporal region—a region known to be affected by magnetic susceptibility. A whole brain analysis showed that the dual GE resulted in significantly higher activation within the left inferior temporal fusiform (ITF) cortex, compared to spin‐echo. The inferior frontal gyrus (IFG) was activated for dual GE, but not spin‐echo. Regions of interest analysis was carried out on the anterior and posterior ITF, left and right IFG, and part of the cerebellum. Dual GE outperformed spin‐echo in the anterior and posterior ITF and bilateral IFG regions, whilst being equal in the cerebellum. Hence, dual GE should be the method of choice for fMRI studies of inferior temporal regions. Hum Brain Mapp 35:4118–4128, 2014.

The role of plasticity-related functional reorganization in the explanation of central dyslexias

This investigation explored the hypothesis that patterns of acquired dyslexia may reflect, in part, plasticity-driven relearning that dynamically alters the division of labour (DOL) between the direct, orthography → phonology (O → P) pathway and the semantically mediated, orthography → semantics → phonology (O → S → P) pathway. Three simulations were conducted using a variant of the triangle model of reading. The model demonstrated core characteristics of normal reading behaviour in its undamaged state. When damage was followed by reoptimization (mimicking spontaneous recovery), the model reproduced the deficits observed in the central dyslexias—acute phonological damage combined with recovery matched data taken from a series of 12 phonological dyslexic patients—whilst progressive semantic damage interspersed with recovery reproduced data taken from 100 observations of semantic dementia patients. The severely phonologically damaged model also produced symptoms of deep dyslexia (imageability effects, production of semantic and mixed semantic/visual errors). In all cases, the DOL changed significantly in the recovery period, suggesting that postmorbid functional reorganization is important in understanding behaviour in chronic-stage patients.

Serial effects in parallel models of reading.

Highlights ► We develop a parallel reading model that includes a visual processing component. ► We examine whether the model can account for “serial” effects found in normal reading. ► The length by lexicality interaction is an emergent property of parallel models with visual processing. ► “Serial” effects in reading can be produced by parallel models. ► Visual processing may be the key to understanding “serial” effects in normal reading.

Using computational, parallel distributed processing networks to model rehabilitation in patients with acquired dyslexia: An initial investigation

Background : Traditional cognitive neuropsychological models are good at diagnosing deficits but are limited when it comes to studying recovery and rehabilitation. Parallel distributed processing (PDP) models have more potential in this regard as they are dynamic and can actually learn. However, to date very little work has been done in using PDP models to study recovery and rehabilitation. Aims : This study seeks to demonstrate how a PDP model of acquired dyslexia can be extended to provide a computational framework that is capable of making predictions about the relative effectiveness of therapeutic interventions. Methods & Procedures : A replication of Plaut, McClelland, Seidenberg, and Pattersons (1996, simulation 2) model of word reading was trained and then damaged. This damaged network was then retrained in a number of different ways designed to model both natural (spontaneous) recovery and recovery that can be attributed to a specific therapeutic intervention. Outcomes & Results : Interventions that used regular words were more effective than interventions based on inconsistent words. Early intervention (during the period of spontaneous recovery) was more effective than late intervention. Conclusions : These results suggest that this technique has the potential to provide a useful input to the therapeutic arena. The potential opportunities for further work are discussed.

Exploring the impact of plasticity-related recovery after brain damage in a connectionist model of single-word reading

The effect of retraining a damaged connectionist model of single-word reading was investigated with the aim of establishing whether plasticity-related changes occurring during the recovery process can contribute to our understanding of the pattern of dissociations found in brain-damaged patients. In particular, we sought to reproduce the strong frequency × consistency interactions found in surface dyslexia. A replication of Plaut, McClelland, Seidenberg, and Patterson’s (1996) model of word reading was damaged and then retrained, using a standard backpropagation algorithm. Immediately after damage, there was only a small frequency × consistency interaction. Retraining the damaged model crystallized out these small differences into a strong dissociation, very similar to the pattern found in surface dyslexic patients. What is more, the percentage of regularization errors, always high in surface dyslexics, increased greatly over the retraining period, moving from under 10% to over 80% in some simulations. These results suggest that the performance patterns of brain-damaged patients can owe as much to the substantial changes in the pattern of connectivity occurring during recovery as to the original premorbid structure. This finding is discussed in relation to the traditional cognitive neuropsychological assumptions of subtractivity and transparency.

Reconstructing coherent networks from electroencephalography and magnetoencephalography with reduced contamination from volume conduction or magnetic field spread.

Volume conduction (VC) and magnetic field spread (MFS) induce spurious correlations between EEG/MEG sensors, such that the estimation of functional networks from scalp recordings is inaccurate. Imaginary coherency [1] reduces VC/MFS artefacts between sensors by assuming that instantaneous interactions are caused predominantly by VC/MFS and do not contribute to the imaginary part of the cross-spectral densities (CSDs). We propose an adaptation of the dynamic imaging of coherent sources (DICS) [2] - a method for reconstructing the CSDs between sources, and subsequently inferring functional connectivity based on coherences between those sources. Firstly, we reformulate the principle of imaginary coherency by performing an eigenvector decomposition of the imaginary part of the CSD to estimate the power that only contributes to the non-zero phase-lagged (NZPL) interactions. Secondly, we construct an NZPL-optimised spatial filter with two a priori assumptions: (1) that only NZPL interactions exist at the source level and (2) the NZPL CSD at the sensor level is a good approximation of the projected source NZPL CSDs. We compare the performance of the NZPL method to the standard method by reconstructing a coherent network from simulated EEG/MEG recordings. We demonstrate that, as long as there are phase differences between the sources, the NZPL method reliably detects the underlying networks from EEG and MEG. We show that the method is also robust to very small phase lags, noise from phase jitter, and is less sensitive to regularisation parameters. The method is applied to a human dataset to infer parts of a coherent network underpinning face recognition.

Modelling normal and impaired letter recognition: Implications for understanding pure alexic reading

Letter recognition is the foundation of the human reading system. Despite this, it tends to receive little attention in computational modelling of single word reading. Here we present a model that can be trained to recognise letters in various spatial transformations. When presented with degraded stimuli the model makes letter confusion errors that correlate with human confusability data. Analyses of the internal representations of the model suggest that a small set of learned visual feature detectors support the recognition of both upper case and lower case letters in various fonts and transformations. We postulated that a damaged version of the model might be expected to act in a similar manner to patients suffering from pure alexia. Summed error score generated from the model was found to be a very good predictor of the reading times of pure alexic patients, outperforming simple word length, and accounting for 47% of the variance. These findings are consistent with a hypothesis suggesting that impaired visual processing is a key to understanding the strong word-length effects found in pure alexic patients.

Optimal connectivity in hardware-targetted MLP networks

In large neural networks, partial connectivity is both biologically plausible and a matter of necessity when targetting a hardware implementation. We are using the SpiNNaker neural chip multiprocessor to model such networks as a drop-in replacement for the Lens network simulator. For the popular MLP network, a theoretical model of the relation between connectivity, network size and gain in the activation function provides a method to set these parameters to near-optimal values. Using the model, we run a series of network simulations in Lens, permuting the parameters to explore the effects in 2 networks of different size and application. Initial test results show a clear connectivity-gain relation and a benefit to partial connectivity in both networks, with optimal hidden-output connectivity values ranging from ∼10%-∼30% depending on the network type. We show that optimal connectivity-gain settings reduce training time, minimising error oscillations during learning. Preliminary analysis also suggests that while very low connectivities may improve error they may also result in decreased adaptivity to new inputs or component failure. These results in combination with the theoretical relation give a method for determining reasonable initial connectivity and gain values at design time for an MLP network, allowing more efficient use of hardware resources such as SpiNNaker and faster simulations in any software environment. They also suggest a different way of considering the problem of MLP network design: rather than specify a fixed number of neurons, specify a fixed number of connections and vary the number of neurons to reach optimal connectivity.