Revealing Vision-Language Integration in the Brain with Multimodal Networks.

We use (multi)modal deep neural networks (DNNs) to probe for sites of multimodal integration in the human brain by predicting stereoen-cephalography (SEEG) recordings taken while human subjects watched movies. We operationalize sites of multimodal integration as regions where a multimodal vision-language model predicts recordings better than unimodal language, unimodal vision, or linearly-integrated language-vision models. Our target DNN models span different architectures (e.g., convolutional networks and transformers) and multimodal training techniques (e.g., cross-attention and contrastive learning). As a key enabling step, we first demonstrate that trained vision and language models systematically outperform their randomly initialized counterparts in their ability to predict SEEG signals. We then compare unimodal and multimodal models against one another. Because our target DNN models often have different architectures, number of parameters, and training sets (possibly obscuring those differences attributable to integration), we carry out a controlled comparison of two models (SLIP and SimCLR), which keep all of these attributes the same aside from input modality. Using this approach, we identify a sizable number of neural sites (on average 141 out of 1090 total sites or 12.94%) and brain regions where multimodal integration seems to occur. Additionally, we find that among the variants of multimodal training techniques we assess, CLIP-style training is the best suited for downstream prediction of the neural activity in these sites.

Population Transformer: Learning Population-level Representations of Intracranial Activity.

We present a self-supervised framework that learns population-level codes for intracranial neural recordings at scale, unlocking the benefits of representation learning for a key neuroscience recording modality. The Population Transformer (PopT) lowers the amount of data required for decoding experiments, while increasing accuracy, even on never-before-seen subjects and tasks. We address two key challenges in developing PopT: sparse electrode distribution and varying electrode location across patients. PopT stacks on top of pretrained representations and enhances downstream tasks by enabling learned aggregation of multiple spatially-sparse data channels. Beyond decoding, we interpret the pretrained PopT and fine-tuned models to show how it can be used to provide neuroscience insights learned from massive amounts of data. We release a pretrained PopT to enable off-the-shelf improvements in multi-channel intracranial data decoding and interpretability, and code is available at https://github.com/czlwang/PopulationTransformer.

CELER: A 365-Participant Corpus of Eye Movements in L1 and L2 English Reading.

We present CELER (Corpus of Eye Movements in L1 and L2 English Reading), a broad coverage eye-tracking corpus for English. CELER comprises over 320,000 words, and eye-tracking data from 365 participants. Sixty-nine participants are L1 (first language) speakers, and 296 are L2 (second language) speakers from a wide range of English proficiency levels and five different native language backgrounds. As such, CELER has an order of magnitude more L2 participants than any currently available eye movements dataset with L2 readers. Each participant in CELER reads 156 newswire sentences from the Wall Street Journal (WSJ), in a new experimental design where half of the sentences are shared across participants and half are unique to each participant. We provide analyses that compare L1 and L2 participants with respect to standard reading time measures, as well as the effects of frequency, surprisal, and word length on reading times. These analyses validate the corpus and demonstrate some of its strengths. We envision CELER to enable new types of research on language processing and acquisition, and to facilitate interactions between psycholinguistics and natural language processing (NLP).

Compositional RL Agents That Follow Language Commands in Temporal Logic.

We demonstrate how a reinforcement learning agent can use compositional recurrent neural networks to learn to carry out commands specified in linear temporal logic (LTL). Our approach takes as input an LTL formula, structures a deep network according to the parse of the formula, and determines satisfying actions. This compositional structure of the network enables zero-shot generalization to significantly more complex unseen formulas. We demonstrate this ability in multiple problem domains with both discrete and continuous state-action spaces. In a symbolic domain, the agent finds a sequence of letters that satisfy a specification. In a Minecraft-like environment, the agent finds a sequence of actions that conform to a formula. In the Fetch environment, the robot finds a sequence of arm configurations that move blocks on a table to fulfill the commands. While most prior work can learn to execute one formula reliably, we develop a novel form of multi-task learning for RL agents that allows them to learn from a diverse set of tasks and generalize to a new set of diverse tasks without any additional training. The compositional structures presented here are not specific to LTL, thus opening the path to RL agents that perform zero-shot generalization in other compositional domains.

How the Brain Represents Language and Answers Questions? Using an AI System to Understand the Underlying Neurobiological Mechanisms.

To understand the computations that underlie high-level cognitive processes we propose a framework of mechanisms that could in principle implement START, an AI program that answers questions using natural language. START organizes a sentence into a series of triplets, each containing three elements (subject, verb, object). We propose that the brain similarly defines triplets and then chunks the three elements into a spatial pattern. A complete sentence can be represented using up to 7 triplets in a working memory buffer organized by theta and gamma oscillations. This buffer can transfer information into long-term memory networks where a second chunking operation converts the serial triplets into a single spatial pattern in a network, with each triplet (with corresponding elements) represented in specialized subregions. The triplets that define a sentence become synaptically linked, thereby encoding the sentence in synaptic weights. When a question is posed, there is a search for the closest stored memory (having the greatest number of shared triplets). We have devised a search process that does not require that the question and the stored memory have the same number of triplets or have triplets in the same order. Once the most similar memory is recalled and undergoes 2-level dechunking, the sought for information can be obtained by element-by-element comparison of the key triplet in the question to the corresponding triplet in the retrieved memory. This search may require a reordering to align corresponding triplets, the use of pointers that link different triplets, or the use of semantic memory. Our framework uses 12 network processes; existing models can implement many of these, but in other cases we can only suggest neural implementations. Overall, our scheme provides the first view of how language-based question answering could be implemented by the brain.