Biology subjects

Salmelin, R.

Publications and source records attributed to Salmelin, R..

At least 19 recordsLinked to original sources

bioRxiv2026

Unstable word form representations and greater reliance on right hemisphere during word learning in adults with a history of developmental language disorder

Impairment in acquiring new vocabulary is a persistent deficit in developmental language disorder (DLD), but its cortical substrates remain unclear. We investigated learning of new words in young adults with DLD and typically developing (TD) control participants using magnetoencephalography (MEG). Participants were exposed to spoken real words and two sets of pseudowords, with shared beginnings, in two MEG sessions. One set of pseudowords was actively trained for three days, paired with fantasy animal pictures. Both groups learned new pseudowords, but adults with DLD recalled fewer items and confused similar sounding pseudowords more often than TD adults. MEG recordings revealed robust left temporal learning effects in both groups, along with left inferior frontal involvement. The DLD group additionally showed stronger right-hemispheric involvement and a significant right temporal learning effect for the newly learned words. Furthermore, only the DLD group showed changes in the left and right temporal responses for the real words after the learning phase, indicating altered lexical access for familiar words after integration of new word forms with overlapping phonology into the mental lexicon. Together, the findings suggest persisting problems in creating and accessing stable and distinctive phonological memory representations for word forms in adults with DLD, along with altered hemispheric balance in processing.

neuroscience↗

bioRxiv2026

When meaning becomes decodable: Linking the N400 evoked response to semantic representations

In non-invasive studies of the human brain, semantic processing during language comprehension has been extensively studied using the N400, a component of the electrophysiological evoked response that is strongly modulated by semantic context. More recently, a complementary approach has emerged that uses multivariate pattern analysis to perform neural decoding of semantic vectors from brain activity, operating on the basis that semantic vectors that more closely align with representations in the brain can be more accurately decoded. To consolidate these two approaches, we investigated the relationship between N400 modulation and semantic decoding performance using magnetoencephalography (MEG) in a controlled priming experiment. Twenty-five native speakers of Finnish read word triplets, for which the semantic relatedness between the two primes and the target word was manipulated based on distance in a word2vec embedding space (highly related, moderately related, or unrelated). We found that words presented after unrelated primes elicited higher N400 responses and provided the best examples for training a decoder to map distributed MEG responses to semantic vectors. Semantic information was decodable from approximately 100 to 500 ms after stimulus onset, at all three levels of contextual support. Soon after the N400 peak, neural responses no longer seemed to encode information that could be mapped to context-invariant semantic vectors. This suggests that the end of the N400 window may correspond to a turning point where the representation shifts from being word-specific to encoding the greater semantic context.

neuroscience↗

bioRxiv2026

Machine learning on magnetoencephalography data yields generalizable low-dimensional neural fingerprints that distinguish individuals across task conditions

Individual brains are unique in structure and function. Functional differences are captured by neural fingerprints, which reflect individual differences in behavior and cognition as well as group-level changes related to neurodegenerative diseases. Most research efforts so far have focused on fingerprints com-prising full functional connectomes. However, the high dimensionality of the connectomes can increase computational load and impede performance of machine learning methods in potential applications. A low-dimensional alternative that retains individual features of the full connectomes would thus be beneficial. The present study employed latent-noise Bayesian Reduced Rank Regression (lnBRRR) to learn low-dimensional latent spaces that capture individual features in functional connectivity and power spectral density data derived from MEG recordings. LnBRRR performance was assessed with low training set sizes (N=20-44), and against principal component analysis and linear discriminant analysis. Model performance was also assessed with task data, and the solutions were compared across task conditions with cosine similarity to establish whether individual features are altered by different cognitive processes. LnBRRR captured generalizable individual patterns already at N=20 but N=30-35 was needed to reach optimal test accuracies and to prevent potential overfitting. The model also achieved comparable performance to the alternative models. Latent fingerprints derived from task data attained comparable performance to resting-state latent fingerprints, and lnBRRR solutions were shown to generalize across conditions. Additionally, the model solutions for power spectral density data were discovered to be notably similar, yet differently rotated, over task conditions, suggesting that similar patterns of individual features were captured by the model regardless of the task condition. Altogether, the present results highlight lnBRRR as a potential tool for neuroimaging data analysis and demonstrate that individual differences in power spectral density are largely intrinsic and unaffected by varying cognitive processes.

neuroscience↗

bioRxiv2026

Estimating mutual information and Pearson correlation on neural evoked responses

In neural evoked responses, small variations in the timing or duration of responses can be observed when the same functional response is recorded in different trials, different experimental conditions or by different sensors. These variations limit the ability of correlation-based methods to detect similarities between signals. Mutual information (MI) provides an alternative similarity measure, capable of capturing both linear and non-linear dependencies, yet its practical use is hindered by lack of consensus on estimators for continuous data and the limited understanding of the behavior of the estimators on realistic signals. In this work, we investigate how to estimate the similarity of neural evoked responses by systematically comparing sample Pearson correlation with three of the most common MI estimators. We describe their behavior using both simulated signals and real magnetoencephalographic data. In the simulations, the estimators are tested against a set of transformations that depict realistic changes in neural evoked responses. Subsequently, we propose guidelines for defining adaptive lower bounds on the similarity estimates and analyzing the similarity rankings induced by the different estimators. Our findings reveal trade-offs between measures sensitivity and different signal properties. We confirm that Pearson correlation is reliable in describing linear relationships for low-noise signals, and we identify parameter settings that stabilize MI estimators, enabling them to capture complex signal dependencies. Together, these results introduce practical parameter choices and thresholding strategies for mutual information and provide guidance for selecting and interpreting similarity measures in the analysis of neural evoked time series.

neuroscience↗

bioRxiv2026

Predictive coding narrows the gap between convolutional networks and human brain function in misspelled-word reading

Humans can readily recognize words even when they are misspelled, though with slower responses. We investigated whether predictive coding could be a feasible computational mechanism to explain both the robustness and the additional processing time. By incorporating brain-inspired predictive coding dynamics into a convolutional neural network (CNN), we assessed whether the resulting interplay between feed-back predictions and feed-forward errors enhanced the models brain-likeness in misspelled-word reading. The initial CNN was trained to classify images of rendered text from a 1000-word Finnish vocabulary (supervised), and then enhanced with feedback predictive coding connections, which were trained with the learning objective of reconstructing the activity in the previous layer (unsupervised). The model, with and without the predictive coding dynamics enabled, was then evaluated using the same real and misspelled word stimuli that were presented to human participants during a magnetoencephalography (MEG) recording. The predictive coding dynamics improved model performance on misspelled words, particularly reducing the accuracy gap between real and word-like misspelled words, thereby aligning overall performance more closely with human behavioral patterns. Furthermore, representational similarity analysis (RSA) and multivariate regression showed a stronger correspondence between model activations and human MEG responses when predictive coding dynamics were enabled. These findings provide converging evidence for predictive coding dynamics as a biologically plausible computational mechanism for the brains ability to cope with misspelled words. Author summaryWhen we read, we can often understand a word even if it is misspelled, e.g, "lamguage". Hence, computational models of visual word recognition in the brain should also exhibit such flexibility. Convolutional neural networks (CNNs) are gaining popularity as models of visual processing in the brain, including reading, yet they typically fail when faced with misspelled words. In our study, we enhanced a CNN by adding feedback connections that perform predictive coding, i.e. continuously attempt to reconstruct the input from the initial output and adjust the output until it can do so. Through this predictive coding loop, the model continuously refines its internal representations, mimicking how the brain may "clean up" noisy inputs. We found that this predictive coding mechanism not only allowed the model to better identify the closest real word form of misspelled words, but also produced activity patterns the more closely resembling those measured from human brain using magnetoencephalography. These results suggest that predictive coding could be a key computational principle underlying the brains remarkable flexibility in reading and provide insights into how biological mechanisms could inspire brain-like computational models.

neuroscience↗

bioRxiv2025

Dynamic feedforward and feedback influences on left ventral occipitotemporal cortex: evidence from word and pseudoword reading

Left ventral occipitotemporal cortex (vOT) is crucial in reading, yet its functional role is viewed either as a prelexical feedforward hub or a bidirectional interface between sensory and higher-order linguistic systems. The two competing views have usually been considered mutually exclusive. To comprehensively explain the functional complexity of left vOT, we investigated the temporal and spectral dynamics of information flows involving left vOT during visual word and pseudoword reading using magnetoencephalography (MEG) and two directed connectivity metrics, i.e., phase slope index (PSI) and Granger causality (GC). Specifically, feedforward connectivity from low-level visual areas to vOT was observed for all conditions, with the strength of orthographic information flow to left superior temporal cortex (ST) modulated by stimulus word-likeness. Conversely, feedback flow from left ST to vOT appeared for pseudowords that allow top-down linguistic constraints to facilitate reading, and occurred later for word-like than complete pseudowords. Our findings suggest that left vOT during word reading functions in a hybrid manner: operating in efficient feedforward mode for familiar word recognition while flexibly recruiting bidirectional processing for unfamiliar pseudowords. By disentangling feedforward and feedback dynamics with high temporal and spectral resolution, we empirically reconcile competing theories of vOT function.

neuroscience↗

bioRxiv2025

Capturing individual variation in children's oscillatory brain activity during nREM sleep

Human brain dynamics are highly unique between individuals: functional neuroimaging studies have recently described functional features that can be used as neural fingerprints. As such studies have generally been performed on adults, it is unclear if the results are applicable during brain development. Here we applied Bayesian reduced-rank regression (BRRR), to extract low-dimensional representations of electroencephalography (EEG) power spectra measured from normally developing children at different ages. The clinical EEG recordings represented different sleep stages in a pediatric population (N=782) of ages varying from 6 weeks to 19 years. The representations learned within specific sleep stages successfully separated between the subjects. Furthermore, a model trained with data across sleep stages generalized to unseen data, suggesting that informative neural fingerprints could be extracted. Fingerprint stability increased with the age of the subjects. In comparison to correlation-based measures in differentiating individuals from each other, the BRRR model performed better, highlighting the usefulness of the dimensionality reduction for data obtained in demanding clinical settings. While further studies are needed to address the possible non-linear maturation effects over developmental periods, our results demonstrate, for the first time, the existence of stable within-session neurofunctional fingerprints in pediatric populations.

neuroscience↗

bioRxiv2025

Misspelled-word reading modulates late cortical dynamics

Literate humans can effortlessly interpret tens of thousands of words, even when the words are sometimes written incorrectly. This phenomenon suggests a flexible nature of reading that can endure a certain amount of noise. In this study, we investigated where and when brain responses diverged for conditions where misspelled words were resolved as real words or not. We used magnetoencephalography (MEG) to track the cortical activity as the participants read words with different degrees of misspelling that were perceived to range from real words to complete pseudowords, as confirmed by their behavioral responses. In particular, we were interested in how lexical information survives (or not) along the uncertainty spectrum, and how the corresponding brain activation patterns evolve spatiotemporally. We identified three brain regions that were notably modulated by misspellings: left ventral occipitotemporal cortex (vOT), superior temporal cortex (ST), and precentral cortex (pC). This suggests that resolving misspelled words into stored concepts involves an interplay between orthographic, semantic, and phonological processing. Temporally, these regions showed fairly late and sustained responses selectively to misspelled words. Specifically, an increasing level of misspelling increased the response in ST from 300 ms after stimulus onset; a functionally fairly similar but weaker effect was observed in pC. In vOT, misspelled words were sharply distinguished from real words notably later, after 700 ms. A linear mixed effects (LME) analysis further showed that pronounced and long-lasting misspelling effects appeared first in ST and then in pC, with shorter-lasting activation also observed in vOT. We conclude that reading misspelled words engages brain areas typically associated with language processing, but in a manner that cannot be interpreted merely as a rapid feedforward mechanism. Instead, feedback interactions likely contribute to the late effects observed during misspelled-word reading.

neuroscience↗

bioRxiv2023

Test-retest reliability of MEG functional brain connectivity related to language processing

The number of studies examining changes in functional connectivity of the human brain is increasing rapidly. In this magnetoencephalography (MEG) study, we examined the reliability of dynamic connectivity related to language processing in a picture naming test-retest paradigm, using data collected from the same participants on two separate days. We determined the connections that were reliable across both days and also examined the behavioral, functional, and structural properties underlying this reliability. A particularly salient finding among a rich set of results was a reliable pattern of beta connectivity increase in the left motor and frontal regions (0-400 ms and 400-800 ms after stimulus onset) and gamma connectivity decrease in the bilateral motor regions (800-1200 ms) which we suggest to represent the motor preparation of speech production. Furthermore, the reliable connections tended to be more frequently associated with the behavioral performance than the non-reliable ones. Finally, the reliable connections were also linked to stronger functional connectivity, as well as to stronger structural connectivity and shorter structural path length, as determined through diffusion MRI (magnetic resonance imaging). Overall, this study defines reliable language-related functional connectivity and introduces practices that may increase reliability. AUTHOR SUMMARYResearch applying connectivity metrics in neuroimaging has increased rapidly during recent years. Hence, the focus has also been to define the best methods for increasing the reliability of connectivity estimation. This study determined reliable functional connectivity from MEG data related to language processing. Moreover, we defined what makes a connection reliable by studying the behavioral, functional, and structural properties underlying the reliable connections.

neuroscience↗

bioRxiv2023

Beta- and gamma-band oscillatory connectivity support naturalistic reading of continuous text

Large-scale integration of information across cortical structures, building on neural connectivity, has been proposed to be a key element in supporting human cognitive processing. In electrophysiological neuroimaging studies of reading, quantification of neural interactions has been limited to the level of isolated words or sentences due to artefacts induced by eye movements. Here, we combined magnetoencephalography recording with advanced artefact rejection tools to investigate both cortico-cortical coherence and directed neural interactions during naturalistic reading of full-page texts. Our results show that reading vs. visual scanning of text was associated with wide-spread increases of cortico-cortical coherence in the beta- and gamma-bands. We further show that the reading task was linked with increased directed neural interactions compared to the scanning task across a sparse set of connections within a wide range of frequencies. Together, the results demonstrate that neural connectivity flexibly builds on different frequency bands to support continuous natural reading.

neuroscience↗

bioRxiv2023

Human sensorimotor beta event characteristics and aperiodic signal are highly heritable

Individuals phenotypes, including the brains structure and function, are largely determined by genes and their interplay. The resting brain generates salient rhythmic patterns which can be characterized non-invasively using functional neuroimaging such as magnetoencephalography (MEG). One of these rhythms, the somatomotor ( rolandic) beta rhythm, shows intermittent high amplitude events which predict behavior across tasks and species. Beta rhythm is altered in neurological disease. The aperiodic ( 1/f) signal present in electrophysiological recordings is also modulated by some neurological conditions and aging. Both sensorimotor beta and aperiodic signal could thus serve as biomarkers of sensorimotor function. Knowledge about the extent to which these brain functional measures are heritable could shed light on the mechanisms underlying their generation. We investigated the heritability and variability of human spontaneous sensorimotor beta rhythm and aperiodic activity in 210 healthy adult siblings spontaneous MEG activity. Both the overall beta spectral power as well as time-resolved beta event amplitude parameters were highly heritable, whereas the heritabilities for peak frequency and measures of event duration remained nonsignificant. Interestingly, the most heritable trait was the aperiodic 1/f signal, with a heritability of 0.94 in the right hemisphere. Human sensorimotor neural activity can thus be dissected into different components with variable heritability. We postulate that differences in heritability in part reflect different underlying signal generating mechanisms. The 1/f signal and beta event amplitude measures may depend more on fixed, anatomical parameters, whereas beta event duration and its modulation reflect dynamic characteristics, guiding their use as potential disease biomarkers.

neuroscience↗

bioRxiv2022

Cortical time-course of evidence accumulation during semantic processing

Our understanding of the surrounding world and communication with other people are tied to mental representations of concepts. In order for the brain to recognize an object, it must determine which concept to access based on information available from sensory inputs. In this study, we combine magnetoencephalography and machine learning to investigate how concepts are represented and accessed in the brain over time. Using brain responses from a silent picture naming task, we track the dynamics of semantic processing, and show that the brain gradually accumulates semantic information before eventually reaching a plateau. The timing of this plateau point varies across individuals, indicating differences in the temporal domain of semantic processing, a new dimension of individual variability in the faculty of language.

neuroscience↗

bioRxiv2022

A large-scale computational model to accurately predict early brain activity in response to written words

Neuroimaging studies have provided a wealth of information about when and where changes in brain activity might be expected during reading. We sought to better understand the computational steps that give rise to such task-related modulations of neural activity by using a convolutional neural network to model the macro-scale computations necessary to perform single-word recognition. We presented the model with stimuli that had been shown to human volunteers in an earlier magnetoencephalography (MEG) experiment and evaluated whether the same experimental effects could be observed in both brain activity and model. In a direct comparison between model and MEG recordings, the model accurately predicted the amplitude changes of three evoked MEG response components commonly observed during single-word reading. In contrast to traditional models of reading, our model directly operates on the pixel values of an image containing text. This allowed us to simulate the whole gamut of processing from the detection and segmentation of letter shapes to word-form identification, with the deep learning architecture facilitating inclusion of a large vocabulary of 10k Finnish words. Interestingly, the key to achieving the desired behavior was to use a noisy activation function for the units in the model as well as to obey word frequency statistics when repeating stimuli during training. We conclude that the deep learning techniques that revolutionized models of object recognition can also create models of reading that can be straightforwardly compared to neuroimaging data, which will greatly facilitate testing and refining theories on language processing in the brain.

neuroscience↗

bioRxiv2020

Picture naming yields highly consistent cortical activation patterns: test-retest reliability of magnetoencephalography recordings

Reliable paradigms and imaging measures of individual-level brain activity are paramount when reaching from group-level research studies to clinical assessment of individual patients. Magnetoencephalography (MEG) provides a direct, non-invasive measure of cortical processing with high spatiotemporal accuracy, and is thus well suited for assessment of functional brain damage in patients with language difficulties. This MEG study aimed to identify, in a picture naming paradigm, source-localized evoked activity and modulations of cortical oscillations that show high test-retest reliability across measurement days in healthy individuals, demonstrating their applicability in clinical settings. For patients with a severe language disorder picture naming can be a challenging task. Therefore, we also determined whether a semantic judgement task ( is this item living) with a spoken response ( yes/no) would suffice to induce comparably consistent activity within brain regions related to language production. The MEG data was collected from 19 healthy participants on two separate days. In picture naming, evoked activity was consistent across measurement days (intraclass correlation coefficient (ICC)>0.4) in the left frontal (400-800 ms after image onset), sensorimotor (200-800 ms), parietal (200-600 ms), temporal (200-800 ms), occipital (400-800 ms) and cingulate (600-800 ms) regions, as well as the right temporal (600-800 ms) region. In the semantic judgement task, consistent evoked activity was spatially more limited, occurring in the left temporal (200-800 ms), sensorimotor (400-800 ms), occipital (400-600 ms) and subparietal (600-800 ms) regions, and the right supramarginal cortex (600-800 ms). The naming task showed typical beta oscillatory suppression in premotor and sensorimotor regions (800-1200 ms) but other consistent modulations of oscillatory activity were mostly observed in posterior cortical regions that have not typically been associated with language processing. The high test-retest consistency of MEG evoked activity in the picture naming task testifies to its applicability in clinical evaluations of language function, as well as in longitudinal MEG studies of language production in clinical and healthy populations.

neuroscience↗

bioRxiv2019

Dynamic time-locking mechanism in the cortical representation of spoken words

Human speech has a unique capacity to carry and communicate rich meanings. However, it is not known how the highly dynamic and variable perceptual signal is mapped to existing linguistic and semantic representations. In this novel approach, we utilized the natural acoustic variability of sounds and mapped them to magnetoencephalography (MEG) data using physiologically-inspired machine-learning models. We aimed at determining how well the models, differing in their representation of temporal information, serve to decode and reconstruct spoken words from MEG recordings in 16 healthy volunteers. We discovered that time-locking of the cortical activation to the unfolding speech input is crucial for the encoding of the acoustic-phonetic features. In contrast, time-locking was not highlighted in cortical processing of non-speech environmental sounds that conveyed the same meanings as the spoken words, including human-made sounds with temporal modulation content similar to speech. The amplitude envelope of the spoken words was particularly well reconstructed based on cortical evoked responses. Our results indicate that speech is encoded cortically with especially high temporal fidelity. This mechanism may contribute to the frequently reported entrainment of the cortical oscillations to the amplitude envelope of speech. Furthermore, the phoneme content was reflected in cortical evoked responses simultaneously with the spectrotemporal features, pointing to an instantaneous transformation of the unfolding acoustic features into linguistic representations during speech processing.\n\nSummary: Computational modeling of cortical responses to spoken words highlights the relevance of temporal tracking of spectrotemporal features, which is likely pivotal for transforming the acoustic-phonetic features into linguistic representations.

neuroscience↗

bioRxiv2019

Post-hoc modification of linear models: combining machine learning with domain information to make solid inferences from noisy data

Linear machine learning models \"learn\" a data transformation by being exposed to examples of input with the desired output, forming the basis for a variety of powerful techniques for analyzing neuroimaging data. However, their ability to learn the desired transformation is limited by the quality and size of the example dataset, which in neuroimaging studies is often notoriously noisy and small. In these cases, it is desirable to fine-tune the learned linear model using domain information beyond the example dataset. To this end, we present a framework that decomposes the weight matrix of a fitted linear model into three subcomponents: the data covariance, the identified signal of interest, and a normalizer. Inspecting these subcomponents in isolation provides an intuitive way to inspect the inner workings of the model and assess its strengths and weaknesses. Furthermore, the three subcomponents may be altered, which provides a straightforward way to inject prior information and impose additional constraints. We refer to this process as \"post-hoc modification\" of a model and demonstrate how it can be used to achieve precise control over which aspects of the model are fitted to the data through machine learning and which are determined through domain information. As an example use case, we decode the associative strength between words from electroencephalography (EEG) reading data. Our results show how the decoding accuracy of two example linear models (ridge regression and logistic regression) can be boosted by incorporating information about the spatio-temporal nature of the data, domain information about the N400 evoked potential and data from other participants.\n\nHighlightsO_LIWe present a framework to decompose any linear model into three subcomponents that are straightforward to interpret.\nC_LIO_LIBy modifying the subcomponents before re-assembling them into a linear model, prior information and further constraints may be injected into the model.\nC_LIO_LIAs an example, we boost the performance of a linear regressor and classifier by injecting knowledge about the spatio-temporal nature of the data, the N400 evoked potential and data from other participants.\nC_LI

neuroscience↗

bioRxiv2018

Reconstructing meaning from bits of information

We can easily identify a dog merely by the sound of barking or an orange by its citrus scent. In this work, we study the neural underpinnings of how the brain combines bits of information into meaningful object representations. Modern theories of semantics posit that the meaning of words can be decomposed into a unique combination of individual semantic features (e.g., \"barks\", \"has citrus scent\"). Here, participants received clues of individual objects in form of three isolated semantic features, given as verbal descriptions. We used machine-learning-based neural decoding to learn a mapping between individual semantic features and BOLD activation patterns. We discovered that the recorded brain patterns were best decoded using a combination of not only the three semantic features that were presented as clues, but a far richer set of semantic features typically linked to the target object. We conclude that our experimental protocol allowed us to observe how fragmented information is combined into a complete semantic representation of an object and suggest neuroanatomical underpinnings for this process.

neuroscience↗

bioRxiv2018

Cracking the problem of neural representations of abstract words: grounding word meanings in language itself

In order to describe how humans represent meaning in the brain, one must be able to account for not just concrete words but, critically, also abstract words which lack a physical referent. Hebbian formalism and optimization are basic principles of brain function, and they provide an appealing approach for modeling word meanings based on word co-occurrences. Here, we built a model of the semantic space based on word statistics in a large text corpus, which was able to decode items from brain signals. In the model, word abstractness emerged from the statistical regularities of the language environment. This salient property of the model co-varied, at 280-420 ms after word presentation, with activity in the left-hemisphere frontal, anterior temporal and superior parietal cortex that have been linked with processing of abstract words. In light of these results, we propose that the neural encoding of word meanings is importantly grounded in language through statistical regularities.

neuroscience↗