Understanding Double Descent Using VC-Theoretical Framework.

In spite of many successful applications of deep learning (DL) networks, theoretical understanding of their generalization capabilities and limitations remains limited. We present analysis of generalization performance of DL networks for classification under VC-theoretical framework. In particular, we analyze the so-called "double descent" phenomenon, when large overparameterized networks can generalize well, even when they perfectly memorize all available training data. This appears to contradict conventional statistical view that optimal model complexity should reflect an optimal balance between underfitting and overfitting, i.e., the bias-variance trade-off. We present VC-theoretical explanation of double descent phenomenon, under classification setting. Our theoretical explanation is supported by empirical modeling of double descent curves, using analytic VC-bounds, for several learning methods, such as support vector machine (SVM), least squares (LS), and multilayer perceptron classifiers. The proposed VC-theoretical approach enables better understanding of overparameterized estimators during second descent.

To understand double descent, we need to understand VC theory.

We analyze generalization performance of over-parameterized learning methods for classification, under VC-theoretical framework. Recently, practitioners in Deep Learning discovered 'double descent' phenomenon, when large networks can fit perfectly available training data, and at the same time, achieve good generalization for future (test) data. The current consensus view is that VC-theoretical results cannot account for good generalization performance of Deep Learning networks. In contrast, this paper shows that double descent can be explained by VC-theoretical concepts, such as VC-dimension and Structural Risk Minimization. We also present empirical results showing that double descent generalization curves can be accurately modeled using classical VC-generalization bounds. Proposed VC-theoretical analysis enables better understanding of generalization curves for data sets with different statistical characteristics, such as low vs high-dimensional data and noisy data. In addition, we analyze generalization performance of transfer learning using pre-trained Deep Learning networks.

Online Prediction of Lead Seizures from iEEG Data.

We describe a novel system for online prediction of lead seizures from long-term intracranial electroencephalogram (iEEG) recordings for canines with naturally occurring epilepsy. This study adopts new specification of lead seizures, reflecting strong clustering of seizures in observed data. This clustering results in fewer lead seizures (~7 lead seizures per dog), and hence new challenges for online seizure prediction, that are addressed in the proposed system. In particular, the machine learning part of the system is implemented using the group learning method suitable for modeling sparse and noisy seizure data. In addition, several modifications for the proposed system are introduced to cope with the non-stationarity of a noisy iEEG signal. They include: (1) periodic retraining of the SVM classifier using most recent training data; (2) removing samples with noisy labels from training data; and (3) introducing a new adaptive post-processing technique for combining many predictions made for 20 s windows into a single prediction for a 4 h segment. Application of the proposed system requires only two lead seizures for training the initial model, and results in high prediction performance for all four dogs (with mean 0.84 sensitivity, 0.27 time-in-warning, and 0.78 false-positive rate per day). The proposed system achieves accurate prediction of lead seizures during long-term test periods, 3-16 lead seizures during a 169-364 day test period, whereas earlier studies did not differentiate between lead vs. non-lead seizures and used much shorter test periods (~few days long).

Performance metrics for online seizure prediction.

Many recent studies on online seizure prediction from iEEG signal describe various prediction algorithms and their prediction performance. In contrast, this paper focuses on proper specification of system parameters, such as prediction period, prediction horizon and data-driven characterization of lead seizures. Whereas prediction performance clearly depends on these system parameters many researchers simply set the values of these parameters in an ad hoc manner. Our paper investigates the effect of these system parameters on online prediction performance, using both synthetic and real-life data sets. Therefore, meaningful comparison of methods/algorithms (for online seizure prediction) should consider proper specification of system parameters.

Machine Learning Approach to Predicting Stem Cell Donor Availability.

The success of unrelated donor stem cell transplants depends on not only finding genetically matched donors, but also donor availability. On average 50% of potential donors in the National Marrow Donor Program database are unavailable for a variety of reasons, after initially matching a patient, with significant variations in availability among subgroups (eg, by race or age). Several studies have established univariate donor characteristics associated with availability. Individual consideration of each applicable characteristic is laborious. Extrapolating group averages to the individual-donor level tends to be highly inaccurate. In the current environment with enhanced donor data collection, we can make better estimates of individual donor availability. We propose a machine learning based approach to predict availability of every registered donor, and evaluate the predictive power on a test cohort of 44,544 requests to be .77 based on the area under the receiver-operating characteristic curve. We propose that this predictor should be used during donor selection to reduce the time to transplant.

Commonalities and differences in the neural representations of English, Portuguese, and Mandarin sentences: When knowledge of the brain-language mappings for two languages is better than one.

This study extended cross-language semantic decoding (based on a concept's fMRI signature) to the decoding of sentences across three different languages (English, Portuguese and Mandarin). A classifier was trained on either the mapping between words and activation patterns in one language or the mappings in two languages (using an equivalent amount of training data), and then tested on its ability to decode the semantic content of a third language. The model trained on two languages was reliably more accurate than a classifier trained on one language for all three pairs of languages. This two-language advantage was selective to abstract concept domains such as social interactions and mental activity. Representational Similarity Analyses (RSA) of the inter-sentence neural similarities resulted in similar clustering of sentences in all the three languages, indicating a shared neural concept space among languages. These findings identify semantic domains that are common across these three languages versus those that are more language or culture-specific.

Predicting the brain activation pattern associated with the propositional content of a sentence: Modeling neural representations of events and states.

Even though much has recently been learned about the neural representation of individual concepts and categories, neuroimaging research is only beginning to reveal how more complex thoughts, such as event and state descriptions, are neurally represented. We present a predictive computational theory of the neural representations of individual events and states as they are described in 240 sentences. Regression models were trained to determine the mapping between 42 neurally plausible semantic features (NPSFs) and thematic roles of the concepts of a proposition and the fMRI activation patterns of various cortical regions that process different types of information. Given a semantic characterization of the content of a sentence that is new to the model, the model can reliably predict the resulting neural signature, or, given an observed neural signature of a new sentence, the model can predict its semantic content. The models were also reliably generalizable across participants. This computational model provides an account of the brain representation of a complex yet fundamental unit of thought, namely, the conceptual content of a proposition. In addition to characterizing a sentence representation at the level of the semantic and thematic features of its component concepts, factor analysis was used to develop a higher level characterization of a sentence, specifying the general type of event representation that the sentence evokes (e.g., a social interaction versus a change of physical state) and the voxel locations most strongly associated with each of the factors. Hum Brain Mapp 38:4865-4881, 2017. © 2017 Wiley Periodicals, Inc.

Neural representations of the concepts in simple sentences: Concept activation prediction and context effects.

Although it has been possible to identify individual concepts from a concept's brain activation pattern, there have been significant obstacles to identifying a proposition from its fMRI signature. Here we demonstrate the ability to decode individual prototype sentences from readers' brain activation patterns, by using theory-driven regions of interest and semantic properties. It is possible to predict the fMRI brain activation patterns evoked by propositions and words which are entirely new to the model with reliably above-chance rank accuracy. The two core components implemented in the model that reflect the theory were the choice of intermediate semantic features and the brain regions associated with the neurosemantic dimensions. This approach also predicts the neural representation of object nouns across participants, studies, and sentence contexts. Moreover, we find that the neural representation of an agent-verb-object proto-sentence is more accurately characterized by the neural signatures of its components as they occur in a similar context than by the neural signatures of these components as they occur in isolation.

Alterations of functional connectivity and intrinsic activity within the cingulate cortex of suicidal ideators.

The 'default mode network' (DMN), a collection of brain regions including the posterior cingulate cortex (PCC), shows reliable inter-regional functional connectivity at rest. It has been implicated in rumination and other negative affective states, but its role in suicidal ideation is not well understood. We employed seed based functional connectivity methods to analyze resting state fMRI data in 34 suicidal ideators and 40 healthy control participants. Whole-brain connectivity with dorsal PCC or ventral PCC was broadly intact between the two groups, but while the control participants showed greater coupling between the dorsal anterior cingulate cortex (dACC) and dorsal PCC, compared to the dACC and ventral PCC, this difference was reversed in the ideators. Furthermore, ongoing low frequency BOLD signal in these three regions (dorsal, ventral PCC, dACC) was reduced in the ideators. The structural integrity of the cingulum bundle, as measured using diffusion tensor imaging (DTI), also explained variation in the functional connectivity measures but did not abolish the group differences. Together, these findings provide evidence of abnormalities in the DMN underlying the tendency towards suicidal ideation.

SVM-Based System for Prediction of Epileptic Seizures From iEEG Signal.

OBJECTIVE: This paper describes a data-analytic modeling approach for the prediction of epileptic seizures from intracranial electroencephalogram (iEEG) recording of brain activity. Even though it is widely accepted that statistical characteristics of iEEG signal change prior to seizures, robust seizure prediction remains a challenging problem due to subject-specific nature of data-analytic modeling. METHODS: Our work emphasizes the understanding of clinical considerations important for iEEG-based seizure prediction, and proper translation of these clinical considerations into data-analytic modeling assumptions. Several design choices during preprocessing and postprocessing are considered and investigated for their effect on seizure prediction accuracy. RESULTS: Our empirical results show that the proposed support vector machine-based seizure prediction system can achieve robust prediction of preictal and interictal iEEG segments from dogs with epilepsy. The sensitivity is about 90-100%, and the false-positive rate is about 0-0.3 times per day. The results also suggest that good prediction is subject specific (dog or human), in agreement with earlier studies. CONCLUSION: Good prediction performance is possible only if the training data contain sufficiently many seizure episodes, i.e., at least 5-7 seizures. SIGNIFICANCE: The proposed system uses subject-specific modeling and unbalanced training data. This system also utilizes three different time scales during training and testing stages.

Machine learning of neural representations of suicide and emotion concepts identifies suicidal youth.

The clinical assessment of suicidal risk would be significantly complemented by a biologically-based measure that assesses alterations in the neural representations of concepts related to death and life in people who engage in suicidal ideation. This study used machine-learning algorithms (Gaussian Naïve Bayes) to identify such individuals (17 suicidal ideators vs 17 controls) with high (91%) accuracy, based on their altered fMRI neural signatures of death and life-related concepts. The most discriminating concepts were death, cruelty, trouble, carefree, good, and praise. A similar classification accurately (94%) discriminated 9 suicidal ideators who had made a suicide attempt from 8 who had not. Moreover, a major facet of the concept alterations was the evoked emotion, whose neural signature served as an alternative basis for accurate (85%) group classification. The study establishes a biological, neurocognitive basis for altered concept representations in participants with suicidal ideation, which enables highly accurate group membership classification.

Commonality of neural representations of sentences across languages: Predicting brain activation during Portuguese sentence comprehension using an English-based model of brain function.

The aim of the study was to test the cross-language generative capability of a model that predicts neural activation patterns evoked by sentence reading, based on a semantic characterization of the sentence. In a previous study on English monolingual speakers (Wang et al., submitted), a computational model performed a mapping from a set of 42 concept-level semantic features (Neurally Plausible Semantic Features, NPSFs) as well as 6 thematic role markers to neural activation patterns (assessed with fMRI), to predict activation levels in a network of brain locations. The model used two types of information gained from the English-based fMRI data to predict the activation for individual sentences in Portuguese. First, it used the mapping weights from NPSFs to voxel activation levels derived from the model for English reading. Second, the brain locations for which the activation levels were predicted were derived from a factor analysis of the brain activation patterns during English reading. These meta-language locations were defined by the clusters of voxels with high loadings on each of the four main dimensions (factors), namely people, places, actions and feelings, underlying the neural representations of the stimulus sentences. This cross-language model succeeded in predicting the brain activation patterns associated with the reading of 60 individual Portuguese sentences that were entirely new to the model, attaining accuracies reliably above chance level. The prediction accuracy was not affected by whether the Portuguese speaker was monolingual or Portuguese-English bilingual. The model's confusion errors indicated an accurate capture of the events or states described in the sentence at a conceptual level. Overall, the cross-language predictive capability of the model demonstrates the neural commonality between speakers of different languages in the representations of everyday events and states, and provides an initial characterization of the common meta-language neural basis.

Identifying thematic roles from neural representations measured by functional magnetic resonance imaging.

The generativity and complexity of human thought stem in large part from the ability to represent relations among concepts and form propositions. The current study reveals how a given object such as rabbit is neurally encoded differently and identifiably depending on whether it is an agent ("the rabbit punches the monkey") or a patient ("the monkey punches the rabbit"). Machine-learning classifiers were trained on functional magnetic resonance imaging (fMRI) data evoked by a set of short videos that conveyed agent-verb-patient propositions. When tested on a held-out video, the classifiers were able to reliably identify the thematic role of an object from its associated fMRI activation pattern. Moreover, when trained on one subset of the study participants, classifiers reliably identified the thematic roles in the data of a left-out participant (mean accuracy = .66), indicating that the neural representations of thematic roles were common across individuals.

Correction: Forecasting Seizures Using Bivariate Intracranial EEG Measures and SVM in Naturally Occurring Canine Epilepsy.

[This corrects the article DOI: 10.1371/journal.pone.0133900.].

Modality-independent representations of small quantities based on brain activation patterns.

Machine learning or MVPA (Multi Voxel Pattern Analysis) studies have shown that the neural representation of quantities of objects can be decoded from fMRI patterns, in cases where the quantities were visually displayed. Here we apply these techniques to investigate whether neural representations of quantities depicted in one modality (say, visual) can be decoded from brain activation patterns evoked by quantities depicted in the other modality (say, auditory). The main finding demonstrated, for the first time, that quantities of dots were decodable by a classifier that was trained on the neural patterns evoked by quantities of auditory tones, and vice-versa. The representations that were common across modalities were mainly right-lateralized in frontal and parietal regions. A second finding was that the neural patterns in parietal cortex that represent quantities were common across participants. These findings demonstrate a common neuronal foundation for the representation of quantities across sensory modalities and participants and provide insight into the role of parietal cortex in the representation of quantity information.

Aberrant functioning of the theory-of-mind network in children and adolescents with autism.

BACKGROUND: Theory-of-mind (ToM), the ability to infer people's thoughts and feelings, is a pivotal skill in effective social interactions. Individuals with autism spectrum disorders (ASD) have been found to have altered ToM skills, which significantly impacts the quality of their social interactions. Neuroimaging studies have reported altered activation of the ToM cortical network, especially in adults with autism, yet little is known about the brain responses underlying ToM in younger individuals with ASD. This functional magnetic resonance imaging (fMRI) study investigated the neural mechanisms underlying ToM in high-functioning children and adolescents with ASD and matched typically developing (TD) peers. METHODS: fMRI data were acquired from 13 participants with ASD and 13 TD control participants while they watched animations involving two "interacting" geometrical shapes. RESULTS: Participants with ASD showed significantly reduced activation, relative to TD controls, in regions considered part of the ToM network, the mirror network, and the cerebellum. Functional connectivity analyses revealed underconnectivity between frontal and posterior regions during task performance in the ASD participants. CONCLUSIONS: Overall, the findings of this study reveal disruptions in the brain circuitry underlying ToM in ASD at multiple levels, including decreased activation and decreased functional connectivity.

Forecasting Seizures Using Intracranial EEG Measures and SVM in Naturally Occurring Canine Epilepsy.

Management of drug resistant focal epilepsy would be greatly assisted by a reliable warning system capable of alerting patients prior to seizures to allow the patient to adjust activities or medication. Such a system requires successful identification of a preictal, or seizure-prone state. Identification of preictal states in continuous long- duration intracranial electroencephalographic (iEEG) recordings of dogs with naturally occurring epilepsy was investigated using a support vector machine (SVM) algorithm. The dogs studied were implanted with a 16-channel ambulatory iEEG recording device with average channel reference for a mean (st. dev.) of 380.4 (+87.5) days producing 220.2 (+104.1) days of intracranial EEG recorded at 400 Hz for analysis. The iEEG records had 51.6 (+52.8) seizures identified, of which 35.8 (+30.4) seizures were preceded by more than 4 hours of seizure-free data. Recorded iEEG data were stratified into 11 contiguous, non-overlapping frequency bands and binned into one-minute synchrony features for analysis. Performance of the SVM classifier was assessed using a 5-fold cross validation approach, where preictal training data were taken from 90 minute windows with a 5 minute pre-seizure offset. Analysis of the optimal preictal training time was performed by repeating the cross validation over a range of preictal windows and comparing results. We show that the optimization of feature selection varies for each subject, i.e. algorithms are subject specific, but achieve prediction performance significantly better than a time-matched Poisson random predictor (p<0.05) in 5/5 dogs analyzed.