We present causal head gating (CHG), a scalable method for interpreting the functional roles of attention heads in transformer models. CHG learns soft gates over heads and assigns them a causal taxonomy - facilitating, interfering, or irrelevant - based on their impact on task performance. Unlike prior approaches in mechanistic interpretability, which are hypothesis-driven and require prompt templates or target labels, CHG applies directly to any dataset using standard next-token prediction. We evaluate CHG across multiple large language models (LLMs) in the Llama 3 model family and diverse tasks, including syntax, commonsense, and mathematical reasoning, and show that CHG scores yield causal, not merely correlational, insight validated via ablation and causal mediation analyses. We also introduce contrastive CHG, a variant that isolates sub-circuits for specific task components. Our findings reveal that LLMs contain multiple sparse task-sufficient sub-circuits, that individual head roles depend on interactions with others (low modularity), and that instruction following and in-context learning rely on separable mechanisms.

BACKGROUND: Exposure and sensitivity to early-life stress (ELS) are related to increased risk for psychopathology in adolescence. While cross-sectional studies have reported blunted nucleus accumbens (NAcc) activation in the context of these associations, researchers have not yet assessed the effects of ELS on developmental trajectories of activation. We examined whether trajectories are affected by stress and the moderating role of biological sex in predicting vulnerability to symptoms of psychopathology. METHODS: Adolescents (n = 173) completed 3 assessments at 2-year intervals across puberty (ages 9-18 years). At baseline, we assessed objective ELS and stress sensitivity using the Traumatic Events Screening Inventory for Children. At all time points, we assessed NAcc activation using the Monetary Incentive Delay task and externalizing, internalizing, and total problems using the Youth Self-Report. We examined correlations between NAcc trajectories (extracted using linear mixed-effects models) with ELS and stress sensitivity and conducted multivariate regression analysis to examine the interaction of NAcc trajectories and biological sex in predicting symptoms of psychopathology. RESULTS: Symptoms increased over adolescence. Stress sensitivity, but not objective ELS, was associated with decreasing trajectories of NAcc activation. Biological sex interacted with NAcc trajectories to predict psychopathology; boys, but not girls, with decreasing NAcc activation had more severe externalizing problems in adolescence. These findings were replicated in the putamen and caudate but not in the medial prefrontal cortex or control brain regions. CONCLUSIONS: NAcc activation may be a sex-specific marker of externalizing problems in adolescence. Efforts to reduce stress sensitivity may help to decrease symptoms of psychopathology in adolescent boys.

Neural networks have long been used to model human intelligence, capturing elements of behavior and cognition, and their neural basis. Recent advancements in deep learning have enabled neural network models to reach and even surpass human levels of intelligence in many respects, yet unlike humans, their ability to learn new tasks quickly remains a challenge. People can reason not only in familiar domains, but can also rapidly learn to reason through novel problems and situations, raising the question of how well modern neural network models capture human intelligence and in which ways they diverge. In this work, we explore this gap by investigating human adults' ability to learn an abstract reasoning task based on Sudoku from a brief instructional tutorial with explanatory feedback for incorrect responses using a narrow range of training examples. We find that participants who master the task do so within a small number of trials and generalize well to puzzles outside of the training range. We also find that most of those who master the task can describe a valid solution strategy, and such participants perform better on transfer puzzles than those whose strategy descriptions are vague or incomplete. Interestingly, fewer than half of our human participants were successful in acquiring a valid solution strategy, and this ability is associated with high school mathematics education. We consider the challenges these findings pose for building computational models that capture all aspects of our findings and point toward a possible role for learning to engage in explanation-based reasoning to support rapid learning and generalization.

Compositionality is an important feature of discrete symbolic systems, such as language and programs, as it enables them to have infinite capacity despite a finite symbol set. It serves as a useful abstraction for reasoning in both cognitive science and in AI, yet the interface between continuous and symbolic processing is often imposed by fiat at the algorithmic level, such as by means of quantization or a softmax sampling step. In this work, we explore how discretization could be implemented in a more neurally plausible manner through the modeling of attractor dynamics that partition the continuous representation space into basins that correspond to sequences of symbols. Building on established work in attractor networks and introducing novel training methods, we show that imposing structure in the symbolic space can produce compositionality in the attractor-supported representation space of rich sensory inputs. Lastly, we argue that our model exhibits the process of an information bottleneck that is thought to play a role in conscious experience, decomposing the rich information of a sensory input into stable components encoding symbolic information.

There are at least three ways of learning how the world works: learning from observations, from interventions, and from explanations. Prior work on causal inference focused on how people learn causal structures through observation and intervention. Our study is the first to look at how explanations support causal structure learning. We develop a normative inference model that learns from observations and explanations, and compare the model's predictions to participants' judgments. The task is to infer the causal connections in 3-node graphs, based on information about their co-activation, and explanations of the kind "B activated because A activated". We find that participants learn better from explanations than from observations. However, while the normative model benefits from having observations in addition to explanations, participants did not.

Large language models have recently shown promising progress in mathematical reasoning when fine-tuned with human-generated sequences walking through a sequence of solution steps. However, the solution sequences are not formally structured and the resulting model-generated sequences may not reflect the kind of systematic reasoning we might expect an expert human to produce. In this paper, we study how to build stronger reasoning capability in language models using the idea of relational abstractions. We introduce new types of sequences that more explicitly provide an abstract characterization of the transitions through intermediate solution steps to the goal state. We find that models that are supplied with such sequences as prompts can solve tasks with a significantly higher accuracy, and models that are trained to produce such sequences solve problems better than those that are trained with previously used human-generated sequences and other baselines. Our work thus takes several steps toward elucidating and improving how language models perform on tasks requiring multi-step mathematical reasoning.

What can be learned about causality and experimentation from passive data? This question is salient given recent successes of passively-trained language models in interactive domains such as tool use. Passive learning is inherently limited. However, we show that purely passive learning can in fact allow an agent to learn generalizable strategies for determining and using causal structures, as long as the agent can intervene at test time. We formally illustrate that learning a strategy of first experimenting, then seeking goals, can allow generalization from passive learning in principle. We then show empirically that agents trained via imitation on expert data can indeed generalize at test time to infer and use causal links which are never present in the training data; these agents can also generalize experimentation strategies to novel variable sets never observed in training. We then show that strategies for causal intervention and exploitation can be generalized from passive data even in a more complex environment with high-dimensional observations, with the support of natural language explanations. Explanations can even allow passive learners to generalize out-of-distribution from perfectly-confounded training data. Finally, we show that language models, trained only on passive next-word prediction, can generalize causal intervention strategies from a few-shot prompt containing examples of experimentation, together with explanations and reasoning. These results highlight the surprising power of passive learning of active causal strategies, and may help to understand the behaviors and capabilities of language models.

Out-of-distribution generalization (OODG) is a longstanding challenge for neural networks. This challenge is quite apparent in tasks with well-defined variables and rules, where explicit use of the rules could solve problems independently of the particular values of the variables, but networks tend to be tied to the range of values sampled in their training data. Large transformer-based language models have pushed the boundaries on how well neural networks can solve previously unseen problems, but their complexity and lack of clarity about the relevant content in their training data obfuscates how they achieve such robustness. As a step toward understanding how transformer-based systems generalize, we explore the question of OODG in small scale transformers trained with examples from a known distribution. Using a reasoning task based on the puzzle Sudoku, we show that OODG can occur on a complex problem if the training set includes examples sampled from the whole distribution of simpler component tasks. Successful generalization depends on carefully managing positional alignment when absolute position encoding is used, but we find that suppressing sensitivity to absolute positions overcomes this limitation. Taken together our results represent a small step toward understanding and promoting systematic generalization in transformers.

Knowledge representation has gained in relevance as data from the ubiquitous digitization of behaviors amass and academia and industry seek methods to understand and reason about the information they encode. Success in this pursuit has emerged with data from natural language, where skip-grams and other linear connectionist models of distributed representation have surfaced scrutable relational structures which have also served as artifacts of anthropological interest. Natural language is, however, only a fraction of the big data deluge. Here we show that latent semantic structure can be informed by behavioral data and that domain knowledge can be extracted from this structure through visualization and a novel mapping of the text descriptions of elements onto this behaviorally informed representation. In this study, we use the course enrollment histories of 124,000 students at a public university to learn vector representations of its courses. From these course selection informed representations, a notable 88% of course attribute information was recovered, as well as 40% of course relationships constructed from prior domain knowledge and evaluated by analogy (e.g., Math 1B is to Honors Math 1B as Physics 7B is to Honors Physics 7B). To aid in interpretation of the learned structure, we create a semantic interpolation, translating course vectors to a bag-of-words of their respective catalog descriptions via regression. We find that representations learned from enrollment histories resolved courses to a level of semantic fidelity exceeding that of their catalog descriptions, revealing nuanced content differences between similar courses, as well as accurately describing departments the dataset had no course descriptions for. We end with a discussion of the possible mechanisms by which this semantic structure may be informed and implications for the nascent research and practice of data science.

In this study, we introduce a novel platform Resource-Aware AutoML (RA-AutoML) which enables flexible and generalized algorithms to build machine learning models subjected to multiple objectives, as well as resource and hard-ware constraints. RA-AutoML intelligently conducts Hyper-Parameter Search(HPS) as well as Neural Architecture Search (NAS) to build models optimizing predefined objectives. RA-AutoML is a versatile framework that allows user to prescribe many resource/hardware constraints along with objectives demanded by the problem at hand or business requirements. At its core, RA-AutoML relies on our in-house search-engine algorithm,MOBOGA, which combines a modified constraint-aware Bayesian Optimization and Genetic Algorithm to construct Pareto optimal candidates. Our experiments on CIFAR-10 dataset shows very good accuracy compared to results obtained by state-of-art neural network models, while subjected to resource constraints in the form of model size.

Automated speech recognition (ASR) systems, which use sophisticated machine-learning algorithms to convert spoken language to text, have become increasingly widespread, powering popular virtual assistants, facilitating automated closed captioning, and enabling digital dictation platforms for health care. Over the last several years, the quality of these systems has dramatically improved, due both to advances in deep learning and to the collection of large-scale datasets used to train the systems. There is concern, however, that these tools do not work equally well for all subgroups of the population. Here, we examine the ability of five state-of-the-art ASR systems—developed by Amazon, Apple, Google, IBM, and Microsoft—to transcribe structured interviews conducted with 42 white speakers and 73 black speakers. In total, this corpus spans five US cities and consists of 19.8 h of audio matched on the age and gender of the speaker. We found that all five ASR systems exhibited substantial racial disparities, with an average word error rate (WER) of 0.35 for black speakers compared with 0.19 for white speakers. We trace these disparities to the underlying acoustic models used by the ASR systems as the race gap was equally large on a subset of identical phrases spoken by black and white individuals in our corpus. We conclude by proposing strategies—such as using more diverse training datasets that include African American Vernacular English—to reduce these performance differences and ensure speech recognition technology is inclusive.

The humans brain possesses intuitive understandings of physics [1] learned during infanthood [2], allowing them to reason about the world using mental simulations [3]. For example, when given tasks to predict the final position of an object given its initial state, humans simulate possible trajectories that might occur given the environment and starting configurations [4, 5]. Understanding the neural architectures and computations that may lead to accurate physical simulations that reflect the processing of the human brain.

Escherichia coli detects and follows chemical gradients in its environment in a process known as chemotaxis. The performance of chemotaxis approaches fundamental biosensor speed and sensitivity limits, but there have been relatively few attempts to incorporate the response into a functional biosensor. Toward that end, we have developed software to process digital microscope images of a large number of tethered E. coli responding to different chemical perturbations. Upwards of fifty cells can be recorded in one experiment, allowing for rapid labeling of the chemotactic responses of multiple cells. After we collected hundreds of wild-type chemotactic E. coli motor responses to dilutions of aspartate and leucine, we trained a support vector classifier (SVC) to estimate the order of magnitude of aspartate concentration between 0M, 100nM, and 1μM with a single cell classification subset accuracy of 69%. We trained another SVC to differentiate between aspartate and leucine with a single cell classification subset accuracy of 83%. Using a majority-vote method on a bacterial population of size N, estimates have 95% confidence for N = 27 bacteria for concentration detection and N = 9 bacteria for chemical differentiation. These methods are a step towards adaptable chemotaxis-based biosensing.

Knowledge representation has gained in relevance as data from the ubiquitous digitization of behaviors amass and academia and industry seek methods to understand and reason about the information they encode. Success in this pursuit has emerged with data from natural language, where skip-grams and other linear connectionist models of distributed representation have surfaced scrutable relational structures which have also served as artifacts of anthropological interest. Natural language is, however, only a fraction of the big data deluge. Here we show that latent semantic structure, comprised of elements from digital records of our interactions, can be informed by behavioral data and that domain knowledge can be extracted from this structure through visualization and a novel mapping of the literal descriptions of elements onto this behaviorally informed representation. We use the course enrollment behaviors of 124,000 students at a public university to learn vector representations of its courses. From these behaviorally informed representations, a notable 88% of course attribute information were recovered (e.g., department and division), as well as 40% of course relationships constructed from prior domain knowledge and evaluated by analogy (e.g., Math 1B is to Math H1B as Physics 7B is to Physics H7B). To aid in interpretation of the learned structure, we create a semantic interpolation, translating course vectors to a bag-of-words of their respective catalog descriptions. We find that the representations learned from enrollments resolved course vectors to a level of semantic fidelity exceeding that of their catalog descriptions, depicting a vector space of high conceptual rationality. We end with a discussion of the possible mechanisms by which this knowledge structure may be informed and its implications for data science.