Brains and algorithms partially converge in natural language processing

Abstract

Deep learning algorithms trained to predict masked words from large amount of text have recently been shown to generate activations similar to those of the human brain. However, what drives this similarity remains currently unknown. Here, we systematically compare a variety of deep language models to identify the computational principles that lead them to generate brain-like representations of sentences. Specifically, we analyze the brain responses to 400 isolated sentences in a large cohort of 102 subjects, each recorded for two hours with functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG). We then test where and when each of these algorithms maps onto the brain responses. Finally, we estimate how the architecture, training, and performance of these models independently account for the generation of brain-like representations. Our analyses reveal two main findings. First, the similarity between the algorithms and the brain primarily depends on their ability to predict words from context. Second, this similarity reveals the rise and maintenance of perceptual, lexical, and compositional representations within each cortical region. Overall, this study shows that modern language algorithms partially converge towards brain-like solutions, and thus delineates a promising path to unravel the foundations of natural language processing.

Fig. 1. Approach.

a Subjects read isolated sentences while their brain activity was recorded with fMRI and MEG. b To compute the similarity between a deep language model and the brain, we (1) fit a linear regression W from the model’s activations X to predict brain responses Y and (2) evaluate this mapping with a correlation between the predicted and true brain responses to held-out sentences Y_test. c We consider different types of embedding depending on whether they vary with neighboring words during training and/or during inference. Visual embeddings refer, here, to the activations of a deep convolutional neural network trained on character recognition. Lexical embeddings refer, here, to the non-contextualized activations associated with a word independently of its context. Here, we use the word-embedding layer of language transformers (bottom green), as opposed to algorithms like Word2Vec (middle, green). Compositional embeddings refer, here, to the context-dependent activations of a deep language model (see SI.4 for a discussion of our terminology). d The three panels represent three hypotheses on the link between deep language models and the brain. Each dot represents one embedding. Algorithm are said to converge to brain-like computations if their performance (x-axis: i.e., accuracy at predicting a word from its previous context) indexes their ability to map onto brain responses to the same stimuli (i.e., y-axis: brain score). High-dimensional neural networks can, in principle, capture relevant information^, and thus lead to a fortunate similarity with brain responses, and event a systematic divergence.

Fig. 2. Average and shared response modeling (or noise ceiling).

a Grand average MEG source estimates to word onset (t = 0 ms) for seven regions typically associated with reading (V1: purple, M1: green, fusiform gyrus: dark blue, supramarginal gyrus: light blue, superior temporal gyrus: orange, infero-frontal gyrus: yellow and fronto-polar gyrus: red), normalized to their peak response. Vertical bars indicate the peak time of each region. The full (not normalized) spatio-temporal time course of the whole-brain activity is displayed in Supplementary Movie 1. b MEG shared response model (or noise ceilings), approximated by predicting brain responses of a given subject from those of all other subjects. Colored lines depict the mean noise ceiling in each region of interest. The gray line depicts the best noise ceiling across sources. c Same as b in sensor space. d Shared response model of fMRI recordings.

Fig. 3. Brain-score comparison across embeddings.

Lexical and compositional representations (see Supplementary Note 4 for the definition of compositionality) can be isolated from (i) the word embedding layer (green) and (ii) one middle layer (red) of a typical language transformer (here, the ninth layer of a 12-layer causal transformer), respectively. We also report the brain scores of a convolutional neural network trained on visual character recognition (blue) to account for low-level visual representations. a Mean (across subjects) fMRI scores obtained with the visual, word, and compositional embeddings. All colored regions display significant fMRI scores across subjects (n = 100) after false discovery rate (FDR) correction. b Mean MEG scores averaged across all time samples and subjects (n = 95 subjects). c Left: mean MEG scores averaged across all sensors. Right: mean MEG gains averaged across all sensors: i.e., the gain in MEG score of one level relative to the level below (blue: R[visual]; green: R[word] − R[visual]; red: R[compositional] − R[word]). d Mean MEG gains in four regions of interest. For a whole-brain depiction of the MEG gains, see Supplementary Movie 2. For the raw scores (without subtraction), see Supplementary Fig. 6. For the distribution of scores across channels and voxels, see Supplementary Fig. 4.

Fig. 4. Language transformers tend to converge towards brain-like representations.

a Bar plots display the average MEG score (across time and channels) of six representative transformers varying in tasks (causal vs. masked language modeling) and depth (4–12 layers). The green and red bars correspond to the word-embedding and middle layers, respectively. The star indicates the layer with the highest MEG score. b Average MEG scores (across subjects, time, and channels) of each of the embeddings (dots) extracted from 18 causal architectures, separately for the input layer (word embedding, green) and the middle layers (red). c Zoom of b, focusing on the best neural networks (i.e., word-prediction accuracy >35%). The results reveal a plateau and/or a divergence of the middle and input layers. d Permutation importance quantifies the extent to which each property of the language transformers specifically contribute to making its embeddings more-or-less similar to brain activity (ΔR). All properties (training task. dimensionality etc.) significantly contribute to the brain scores (ΔR > 0, all p < 0.0001 across subjects). Ordered pairwise comparisons of the permutation scores are marked with a star (*p < 0.05, **p < 0.01, ***p < 0.001). e–h Same as a–d, but evaluated on fMRI recordings. All error bars are the 95% confidence intervals across subjects (n = 95 for MEG, n = 100 for fMRI).