Search

Transformer-based pretrained models like BERT, GPT-2 and T5 have been finetuned for a large number of natural language processing (NLP) tasks, and have been shown to be very effective. However, while finetuning, what changes across layers in these models with respect to pretrained checkpoints is under-studied. Further, how robust are these models to perturbations in input text? Does the robustness vary depending on the NLP task for which the models have been finetuned? While there exists some work on studying robustness of BERT finetuned for a few NLP tasks, there is no rigorous study which compares this robustness across encoder only, decoder only and encoder-decoder models. In this paper, we study the robustness of three language models (BERT, GPT-2 and T5) with eight different text perturbations on the General Language Understanding Evaluation (GLUE) benchmark. Also, we use two metrics (CKA and STIR) to quantify changes between pretrained and finetuned language model representations across layers. GPT-2 representations are more robust than BERT and T5 across multiple types of input perturbation. Although models exhibit good robustness broadly, dropping nouns, verbs or changing characters are the most impactful. Overall, this study provides valuable insights into perturbation-specific weaknesses of popular Transformer-based models which should be kept in mind when passing inputs., 16 pages, 8 figures

Syntactic parsing is the task of assigning a syntactic structure to a sentence. There are two popular syntactic parsing methods: constituency and dependency parsing. Recent works have used syntactic embeddings based on constituency trees, incremental top-down parsing, and other word syntactic features for brain activity prediction given the text stimuli to study how the syntax structure is represented in the brain's language network. However, the effectiveness of dependency parse trees or the relative predictive power of the various syntax parsers across brain areas, especially for the listening task, is yet unexplored. In this study, we investigate the predictive power of the brain encoding models in three settings: (i) individual performance of the constituency and dependency syntactic parsing based embedding methods, (ii) efficacy of these syntactic parsing based embedding methods when controlling for basic syntactic signals, (iii) relative effectiveness of each of the syntactic embedding methods when controlling for the other. Further, we explore the relative importance of syntactic information (from these syntactic embedding methods) versus semantic information using BERT embeddings. We find that constituency parsers help explain activations in the temporal lobe and middle-frontal gyrus, while dependency parsers better encode syntactic structure in the angular gyrus and posterior cingulate cortex. Although semantic signals from BERT are more effective compared to any of the syntactic features or embedding methods, syntactic embedding methods explain additional variance for a few brain regions., 21 pages, 22 figures

How does the brain represent different modes of information? Can we design a system that automatically understands what the user is thinking? Such questions can be answered by studying brain recordings like functional magnetic resonance imaging (fMRI). As a first step, the neuroscience community has contributed several large cognitive neuroscience datasets related to passive reading/listening/viewing of concept words, narratives, pictures and movies. Encoding and decoding models using these datasets have also been proposed in the past two decades. These models serve as additional tools for basic research in cognitive science and neuroscience. Encoding models aim at generating fMRI brain representations given a stimulus automatically. They have several practical applications in evaluating and diagnosing neurological conditions and thus also help design therapies for brain damage. Decoding models solve the inverse problem of reconstructing the stimuli given the fMRI. They are useful for designing brain-machine or brain-computer interfaces. Inspired by the effectiveness of deep learning models for natural language processing, computer vision, and speech, recently several neural encoding and decoding models have been proposed. In this survey, we will first discuss popular representations of language, vision and speech stimuli, and present a summary of neuroscience datasets. Further, we will review popular deep learning based encoding and decoding architectures and note their benefits and limitations. Finally, we will conclude with a brief summary and discussion about future trends. Given the large amount of recently published work in the `computational cognitive neuroscience' community, we believe that this survey nicely organizes the plethora of work and presents it as a coherent story., Comment: 16 pages, 10 figures

How do children acquire language through unsupervised or noisy supervision? How do their brain process language? We take this perspective to machine learning and robotics, where part of the problem is understanding how language models can perform grounded language acquisition through noisy supervision and discussing how they can account for brain learning dynamics. Most prior works have tracked the co-occurrence between single words and referents to model how infants learn wordreferent mappings. This paper studies cross-situational learning (CSL) with full sentences: we want to understand brain mechanisms that enable children to learn mappings between words and their meanings from full sentences in early language learning. We investigate the CSL task on a few training examples with two sequence-based models: (i) Echo State Networks (ESN) and (ii) Long-Short Term Memory Networks (LSTM). Most importantly, we explore several word representations including One-Hot, GloVe, pretrained BERT, and fine-tuned BERT representations (last layer token representations) to perform the CSL task. We apply our approach to three diverse datasets (two grounded language datasets and a robotic dataset) and observe that (1) One-Hot, GloVe, and pretrained BERT representations are less efficient when compared to representations obtained from fine-tuned BERT. (2) ESN online with final learning (FL) yields superior performance over ESN online continual learning (CL), offline learning, and LSTMs, indicating the more biological plausibility of ESNs and the cognitive process of sentence reading. (2) LSTM with fewer hidden units showcases higher performance for small datasets, but LSTM with more hidden units is Cross-Situational Learning needed to perform reasonably well on larger corpora. (4) ESNs demonstrate better generalization than LSTM models for increasingly large vocabularies. Overall, these models are able to learn from scratch to link complex relations between words and their corresponding meaning concepts, handling polysemous and synonymous words. Moreover, we argue that such models can extend to help current human-robot interaction studies on language grounding and better understand children's developmental language acquisition. We make the code publicly available * .

Document summarization aims to create a precise and coherent summary of a text document. Many deep learning summarization models are developed mainly for English, often requiring a large training corpus and efficient pre-trained language models and tools. However, English summarization models for low-resource Indian languages are often limited by rich morphological variation, syntax, and semantic differences. In this paper, we propose GAE-ISumm, an unsupervised Indic summarization model that extracts summaries from text documents. In particular, our proposed model, GAE-ISumm uses Graph Autoencoder (GAE) to learn text representations and a document summary jointly. We also provide a manually-annotated Telugu summarization dataset TELSUM, to experiment with our model GAE-ISumm. Further, we experiment with the most publicly available Indian language summarization datasets to investigate the effectiveness of GAE-ISumm on other Indian languages. Our experiments of GAE-ISumm in seven languages make the following observations: (i) it is competitive or better than state-of-the-art results on all datasets, (ii) it reports benchmark results on TELSUM, and (iii) the inclusion of positional and cluster information in the proposed model improved the performance of summaries., 9 pages, 7 figures

Language models have been shown to be very effective in predicting brain recordings of subjects experiencing complex language stimuli. For a deeper understanding of this alignment, it is important to understand the alignment between the detailed processing of linguistic information by the human brain versus language models. In NLP, linguistic probing tasks have revealed a hierarchy of information processing in neural language models that progresses from simple to complex with an increase in depth. On the other hand, in neuroscience, the strongest alignment with high-level language brain regions has consistently been observed in the middle layers. These findings leave an open question as to what linguistic information actually underlies the observed alignment between brains and language models. We investigate this question via a direct approach, in which we eliminate information related to specific linguistic properties in the language model representations and observe how this intervention affects the alignment with fMRI brain recordings obtained while participants listened to a story. We investigate a range of linguistic properties (surface, syntactic and semantic) and find that the elimination of each one results in a significant decrease in brain alignment across all layers of a language model. These findings provide direct evidence for the role of specific linguistic information in the alignment between brain and language models, and opens new avenues for mapping the joint information processing in both systems., 15 pages, 7 figures

Several popular sequence-based and pretrained language models have been found to be successful for text-driven prediction of brain activations. However, these models still lack long-term cognitive plausibility as well as insights on the underlying neural substrate mechanisms. This paper studies the influence of context representations of different language models such as sequence-based models: Long short-term memory networks (LSTMs) and ELMo, and popular pretrained Transformer language model (Longformer). In particular, we study how the internal hidden representations align with the brain activity observed via fMRI when the subjects listen to several narrative stories. We use brain imaging recordings of subjects listening to narrative stories to interpret word and sequence embeddings. We further investigate how the representations of language models layers reveal better semantic context during listening. Experiments across all language model representations provide the following cognitive insights: (i) the representations of LSTM cell states are better aligned with brain recordings than LSTM (hidden state), the cell state activity can represent more long-term information, (ii) the representations of ELMo and Longformer display a good predictive performance across brain regions for listening stimuli; (iii) Posterior Medial Cortex (PMC), Temporo-Parieto-Occipital junction (TPOJ), and Dorsal Frontal Lobe (DFL) have higher correlation versus Early Auditory (EAC) and Auditory Association Cortex (AAC).

How the brain captures the meaning of linguistic stimuli across multiple views is still a critical open question in neuroscience. Consider three different views of the concept apartment: (1) picture (WP) presented with the target word label, (2) sentence (S) using the target word, and (3) word cloud (WC) containing the target word along with other semantically related words. Unlike previous efforts, which focus only on single view analysis, in this paper, we study the effectiveness of brain decoding in a zero-shot cross-view learning setup. Further, we propose brain decoding in the novel context of cross-view-translation tasks like image captioning (IC), image tagging (IT), keyword extraction (KE), and sentence formation (SF). Using extensive experiments, we demonstrate that cross-view zero-shot brain decoding is practical leading to ~0.68 average pairwise accuracy across view pairs. Also, the decoded representations are sufficiently detailed to enable high accuracy for cross-view-translation tasks with following pairwise accuracy: IC (78.0), IT (83.0), KE (83.7) and SF (74.5). Analysis of the contribution of different brain networks reveals exciting cognitive insights: (1) A high percentage of visual voxels are involved in image captioning and image tagging tasks, and a high percentage of language voxels are involved in the sentence formation and keyword extraction tasks. (2) Zero-shot accuracy of the model trained on S view and tested on WC view is better than same-view accuracy of the model trained and tested on WC view., Comment: 11 pages, 10 figures

fMRI semantic category understanding using linguistic encoding models attempts to learn a forward mapping that relates stimuli to the corresponding brain activation. State-of-the-art encoding models use a single global model (linear or non-linear) to predict brain activation given the stimulus. However, the critical assumption in these methods is that a priori different brain regions respond the same way to all the stimuli, that is, there is no modularity or specialization assumed for any region. This goes against the modularity theory, supported by many cognitive neuroscience investigations suggesting that there are functionally specialized regions in the brain. In this paper, we achieve this by clustering similar regions together and for every cluster we learn a different linear regression model using a mixture of linear experts model. The key idea here is that each linear expert captures the behaviour of similar brain regions. Given a new stimulus, the utility of the proposed model is twofold (i) predicts the brain activation as a weighted linear combination of the activations of multiple linear experts and (ii) to learn multiple experts corresponding to different brain regions. We argue that each expert captures activity patterns related to a particular region of interest (ROI) in the human brain. This study helps in understanding the brain regions that are activated together given different kinds of stimuli. Importantly, we suggest that the mixture of regression experts (MoRE) framework successfully combines the two principles of organization of function in the brain, namely that of specialization and integration. Experiments on fMRI data from paradigm 1 [1]where participants view linguistic stimuli show that the proposed MoRE model has better prediction accuracy compared to that of conventional models., Comment: 15 pages

fMRI semantic category understanding using linguistic encoding models attempt to learn a forward mapping that relates stimuli to the corresponding brain activation. Classical encoding models use linear multi-variate methods to predict the brain activation (all voxels) given the stimulus. However, these methods essentially assume multiple regions as one large uniform region or several independent regions, ignoring connections among them. In this paper, we present a mixture of experts-based model where a group of experts captures brain activity patterns related to particular regions of interest (ROI) and also show the discrimination across different experts. The model is trained word stimuli encoded as 25-dimensional feature vectors as input and the corresponding brain responses as output. Given a new word (25-dimensional feature vector), it predicts the entire brain activation as the linear combination of multiple experts brain activations. We argue that each expert learns a certain region of brain activations corresponding to its category of words, which solves the problem of identifying the regions with a simple encoding model. We showcase that proposed mixture of experts-based model indeed learns region-based experts to predict the brain activations with high spatial accuracy., Comment: 8 pages, 3 figures, Workshop on Visually Grounded Interaction and Language @ 32nd Conference on Neural Information Processing Systems (NeurIPS 2018), Montr\'eal, Canada

The dispute of how the human brain represents conceptual knowledge has been argued in many scientific fields. Brain imaging studies have shown that the spatial patterns of neural activation in the brain are correlated with thinking about different semantic categories of words (for example, tools, animals, and buildings) or when viewing the related pictures. In this paper, we present a computational model that learns to predict the neural activation captured in functional magnetic resonance imaging (fMRI) data of test words. Unlike the models with hand-crafted features that have been used in the literature, in this paper we propose a novel approach wherein decoding models are built with features extracted from popular linguistic encodings of Word2Vec, GloVe, Meta-Embeddings in conjunction with the empirical fMRI data associated with viewing several dozen concrete nouns. We compared these models with several other models that use word features extracted from FastText, Randomly-generated features, Mitchell's 25 features [1]. The experimental results show that the predicted fMRI images using Meta-Embeddings meet the state-of-the-art performance. Although models with features from GloVe and Word2Vec predict fMRI images similar to the state-of-the-art model, model with features from Meta-Embeddings predicts significantly better. The proposed scheme that uses popular linguistic encoding offers a simple and easy approach for semantic decoding from fMRI experiments., Comment: 12 pages, 7 Figures

International audience; How does the brain represent different modes of information? Can we design a system that can automatically understand what the user is thinking? We can make progress towards answering such questions by studying brain recordings from devices such as functional magnetic resonance imaging (fMRI). The brain encoding problem aims to automatically generate fMRI brain representations given a stimulus. The brain decoding problem is the inverse problem of reconstructing the stimuli given the fMRI brain representation. Both the brain encoding and decoding problems have been studied in detail in the past two decades and the foremost attraction of studying these solutions is that they serve as additional tools for basic research in cognitive science and cognitive neuroscience. Recently, inspired by the effectiveness of deep learning models for natural language processing and computer vision, such models have been applied for neuroscience as well. In this tutorial, we plan to discuss different kinds of stimulus representations, and popular encoding and decoding architectures in detail. The tutorial will provide a working knowledge of the state of the art methods for encoding and decoding, a thorough understanding of the literature, and a better understanding of the benefits and limitations of encoding/decoding with deep learning.