Search

Large language models (LLMs) have emerged as powerful tools in various domains. Recent studies have shown that LLMs can surpass humans in certain tasks, such as predicting the outcomes of neuroscience studies. What role does this leave for humans in the overall decision process? One possibility is that humans, despite performing worse than LLMs, can still add value when teamed with them. A human and machine team can surpass each individual teammate when team members' confidence is well-calibrated and team members diverge in which tasks they find difficult (i.e., calibration and diversity are needed). We simplified and extended a Bayesian approach to combining judgments using a logistic regression framework that integrates confidence-weighted judgments for any number of team members. Using this straightforward method, we demonstrated in a neuroscience forecasting task that, even when humans were inferior to LLMs, their combination with one or more LLMs consistently improved team performance. Our hope is that this simple and effective strategy for integrating the judgments of humans and machines will lead to productive collaborations.

Recently, large language models (LLMs) have outperformed human experts in predicting the results of neuroscience experiments (Luo et al., 2024). What is the basis for this performance? One possibility is that statistical patterns in that specific scientific literature, as opposed to emergent reasoning abilities arising from broader training, underlie LLMs' performance. To evaluate this possibility, we trained (next word prediction) a relatively small 124M-parameter GPT-2 model on 1.3 billion tokens of domain-specific knowledge. Despite being orders of magnitude smaller than larger LLMs trained on trillions of tokens, small models achieved expert-level performance in predicting neuroscience results. Small models trained on the neuroscience literature succeeded when they were trained from scratch using a tokenizer specifically trained on neuroscience text or when the neuroscience literature was used to finetune a pretrained GPT-2. Our results indicate that expert-level performance may be attained by even small LLMs through domain-specific, auto-regressive training approaches., Comment: ICML 2024 (Large Language Models and Cognition)

Scientific discoveries often hinge on synthesizing decades of research, a task that potentially outstrips human information processing capacities. Large language models (LLMs) offer a solution. LLMs trained on the vast scientific literature could potentially integrate noisy yet interrelated findings to forecast novel results better than human experts. To evaluate this possibility, we created BrainBench, a forward-looking benchmark for predicting neuroscience results. We find that LLMs surpass experts in predicting experimental outcomes. BrainGPT, an LLM we tuned on the neuroscience literature, performed better yet. Like human experts, when LLMs were confident in their predictions, they were more likely to be correct, which presages a future where humans and LLMs team together to make discoveries. Our approach is not neuroscience-specific and is transferable to other knowledge-intensive endeavors.

Machine learning (ML) provides powerful tools for predictive modeling. ML's popularity stems from the promise of sample-level prediction with applications across a variety of fields from physics and marketing to healthcare. However, if not properly implemented and evaluated, ML pipelines may contain leakage typically resulting in overoptimistic performance estimates and failure to generalize to new data. This can have severe negative financial and societal implications. Our aim is to expand understanding associated with causes leading to leakage when designing, implementing, and evaluating ML pipelines. Illustrated by concrete examples, we provide a comprehensive overview and discussion of various types of leakage that may arise in ML pipelines., Comment: second draft

Biological and artificial information processing systems form representations that they can use to categorize, reason, plan, navigate, and make decisions. How can we measure the extent to which the representations formed by these diverse systems agree? Do similarities in representations then translate into similar behavior? How can a system's representations be modified to better match those of another system? These questions pertaining to the study of representational alignment are at the heart of some of the most active research areas in cognitive science, neuroscience, and machine learning. For example, cognitive scientists measure the representational alignment of multiple individuals to identify shared cognitive priors, neuroscientists align fMRI responses from multiple individuals into a shared representational space for group-level analyses, and ML researchers distill knowledge from teacher models into student models by increasing their alignment. Unfortunately, there is limited knowledge transfer between research communities interested in representational alignment, so progress in one field often ends up being rediscovered independently in another. Thus, greater cross-field communication would be advantageous. To improve communication between these fields, we propose a unifying framework that can serve as a common language between researchers studying representational alignment. We survey the literature from all three fields and demonstrate how prior work fits into this framework. Finally, we lay out open problems in representational alignment where progress can benefit all three of these fields. We hope that our work can catalyze cross-disciplinary collaboration and accelerate progress for all communities studying and developing information processing systems. We note that this is a working paper and encourage readers to reach out with their suggestions for future revisions., Comment: Working paper, changes to be made in upcoming revisions

The Brain Imaging Data Structure (BIDS) is a community-driven standard for the organization of data and metadata from a growing range of neuroscience modalities. This paper is meant as a history of how the standard has developed and grown over time. We outline the principles behind the project, the mechanisms by which it has been extended, and some of the challenges being addressed as it evolves. We also discuss the lessons learned through the project, with the aim of enabling researchers in other domains to learn from the success of BIDS.

Aligning model representations to humans has been found to improve robustness and generalization. However, such methods often focus on standard observational data. Synthetic data is proliferating and powering many advances in machine learning; yet, it is not always clear whether synthetic labels are perceptually aligned to humans -- rendering it likely model representations are not human aligned. We focus on the synthetic data used in mixup: a powerful regularizer shown to improve model robustness, generalization, and calibration. We design a comprehensive series of elicitation interfaces, which we release as HILL MixE Suite, and recruit 159 participants to provide perceptual judgments along with their uncertainties, over mixup examples. We find that human perceptions do not consistently align with the labels traditionally used for synthetic points, and begin to demonstrate the applicability of these findings to potentially increase the reliability of downstream models, particularly when incorporating human uncertainty. We release all elicited judgments in a new data hub we call H-Mix.

Machine learning (ML) approaches to data analysis are now widely adopted in many fields including epidemiology and medicine. To apply these approaches, confounds must first be removed as is commonly done by featurewise removal of their variance by linear regression before applying ML. Here, we show this common approach to confound removal biases ML models, leading to misleading results. Specifically, this common deconfounding approach can leak information such that what are null or moderate effects become amplified to near-perfect prediction when nonlinear ML approaches are subsequently applied. We identify and evaluate possible mechanisms for such confound-leakage and provide practical guidance to mitigate its negative impact. We demonstrate the real-world importance of confound-leakage by analyzing a clinical dataset where accuracy is overestimated for predicting attention deficit hyperactivity disorder (ADHD) with depression as a confound. Our results have wide-reaching implications for implementation and deployment of ML workflows and beg caution against na\"ive use of standard confound removal approaches., Comment: Revised Introduction, added CoI, results unchanged

An evaluation criterion for safe and trustworthy deep learning is how well the invariances captured by representations of deep neural networks (DNNs) are shared with humans. We identify challenges in measuring these invariances. Prior works used gradient-based methods to generate identically represented inputs (IRIs), ie, inputs which have identical representations (on a given layer) of a neural network, and thus capture invariances of a given network. One necessary criterion for a network's invariances to align with human perception is for its IRIs look 'similar' to humans. Prior works, however, have mixed takeaways; some argue that later layers of DNNs do not learn human-like invariances (\cite{jenelle2019metamers}) yet others seem to indicate otherwise (\cite{mahendran2014understanding}). We argue that the loss function used to generate IRIs can heavily affect takeaways about invariances of the network and is the primary reason for these conflicting findings. We propose an adversarial regularizer on the IRI generation loss that finds IRIs that make any model appear to have very little shared invariance with humans. Based on this evidence, we argue that there is scope for improving models to have human-like invariances, and further, to have meaningful comparisons between models one should use IRIs generated using the regularizer-free loss. We then conduct an in-depth investigation of how different components (eg architectures, training losses, data augmentations) of the deep learning pipeline contribute to learning models that have good alignment with humans. We find that architectures with residual connections trained using a (self-supervised) contrastive loss with $\ell_p$ ball adversarial data augmentation tend to learn invariances that are most aligned with humans. Code: \url{github.com/nvedant07/Human-NN-Alignment}., Comment: AAAI 2023

Top-down attention allows neural networks, both artificial and biological, to focus on the information most relevant for a given task. This is known to enhance performance in visual perception. But it remains unclear how attention brings about its perceptual boost, especially when it comes to naturalistic settings like recognising an object in an everyday scene. What aspects of a visual task does attention help to deal with? We aim to answer this with a computational experiment based on a general framework called task-oriented ablation design. First we define a broad range of visual tasks and identify six factors that underlie task variability. Then on each task we compare the performance of two neural networks, one with top-down attention and one without. These comparisons reveal the task-dependence of attention's perceptual boost, giving a clearer idea of the role attention plays. Whereas many existing cognitive accounts link attention to stimulus-level variables, such as visual clutter and object scale, we find greater explanatory power in system-level variables that capture the interaction between the model, the distribution of training data and the task format. This finding suggests a shift in how attention is studied could be fruitful. We make publicly available our code and results, along with statistics relevant to ImageNet-based experiments beyond this one. Our contribution serves to support the development of more human-like vision models and the design of more informative machine-learning experiments.

Despite their impressive performance in object recognition and other tasks under standard testing conditions, deep networks often fail to generalize to out-of-distribution (o.o.d.) samples. One cause for this shortcoming is that modern architectures tend to rely on "shortcuts" - superficial features that correlate with categories without capturing deeper invariants that hold across contexts. Real-world concepts often possess a complex structure that can vary superficially across contexts, which can make the most intuitive and promising solutions in one context not generalize to others. One potential way to improve o.o.d. generalization is to assume simple solutions are unlikely to be valid across contexts and avoid them, which we refer to as the too-good-to-be-true prior. A low-capacity network (LCN) with a shallow architecture should only be able to learn surface relationships, including shortcuts. We find that LCNs can serve as shortcut detectors. Furthermore, an LCN's predictions can be used in a two-stage approach to encourage a high-capacity network (HCN) to rely on deeper invariant features that should generalize broadly. In particular, items that the LCN can master are downweighted when training the HCN. Using a modified version of the CIFAR-10 dataset in which we introduced shortcuts, we found that the two-stage LCN-HCN approach reduced reliance on shortcuts and facilitated o.o.d. generalization., Comment: 10 pages, 8 figures

Advances in object recognition flourished in part because of the availability of high-quality datasets and associated benchmarks. However, these benchmarks---such as ILSVRC---are relatively task-specific, focusing predominately on predicting class labels. We introduce a publicly-available dataset that embodies the task-general capabilities of human perception and reasoning. The Human Similarity Judgments extension to ImageNet (ImageNet-HSJ) is composed of human similarity judgments that supplement the ILSVRC validation set. The new dataset supports a range of task and performance metrics, including the evaluation of unsupervised learning algorithms. We demonstrate two methods of assessment: using the similarity judgments directly and using a psychological embedding trained on the similarity judgments. This embedding space contains an order of magnitude more points (i.e., images) than previous efforts based on human judgments. Scaling to the full 50,000 image set was made possible through a selective sampling process that used variational Bayesian inference and model ensembles to sample aspects of the embedding space that were most uncertain. This methodological innovation not only enables scaling, but should also improve the quality of solutions by focusing sampling where it is needed. To demonstrate the utility of ImageNet-HSJ, we used the similarity ratings and the embedding space to evaluate how well several popular models conform to human similarity judgments. One finding is that more complex models that perform better on task-specific benchmarks do not better conform to human semantic judgments. In addition to the human similarity judgments, pre-trained psychological embeddings and code for inferring variational embeddings are made publicly available. Collectively, ImageNet-HSJ assets support the appraisal of internal representations and the development of more human-like models.

Induction benefits from useful priors. Penalized regression approaches, like ridge regression, shrink weights toward zero but zero association is usually not a sensible prior. Inspired by simple and robust decision heuristics humans use, we constructed non-zero priors for penalized regression models that provide robust and interpretable solutions across several tasks. Our approach enables estimates from a constrained model to serve as a prior for a more general model, yielding a principled way to interpolate between models of differing complexity. We successfully applied this approach to a number of decision and classification problems, as well as analyzing simulated brain imaging data. Models with robust priors had excellent worst-case performance. Solutions followed from the form of the heuristic that was used to derive the prior. These new algorithms can serve applications in data analysis and machine learning, as well as help in understanding how people transition from novice to expert performance., Comment: 5 figures

Top-down attention allows people to focus on task-relevant visual information. Is the resulting perceptual boost task-dependent in naturalistic settings? We aim to answer this with a large-scale computational experiment. First, we design a collection of visual tasks, each consisting of classifying images from a chosen task set (subset of ImageNet categories). The nature of a task is determined by which categories are included in the task set. Second, on each task we train an attention-augmented neural network and then compare its accuracy to that of a baseline network. We show that the perceptual boost of attention is stronger with increasing task-set difficulty, weaker with increasing task-set size and weaker with increasing perceptual similarity within a task set., Comment: Published as a workshop paper at "Bridging AI and Cognitive Science" (ICLR 2020)

People deploy top-down, goal-directed attention to accomplish tasks, such as finding lost keys. By tuning the visual system to relevant information sources, object recognition can become more efficient (a benefit) and more biased toward the target (a potential cost). Motivated by selective attention in categorisation models, we developed a goal-directed attention mechanism that can process naturalistic (photographic) stimuli. Our attention mechanism can be incorporated into any existing deep convolutional neural network (DCNNs). The processing stages in DCNNs have been related to ventral visual stream. In that light, our attentional mechanism incorporates top-down influences from prefrontal cortex (PFC) to support goal-directed behaviour. Akin to how attention weights in categorisation models warp representational spaces, we introduce a layer of attention weights to the mid-level of a DCNN that amplify or attenuate activity to further a goal. We evaluated the attentional mechanism using photographic stimuli, varying the attentional target. We found that increasing goal-directed attention has benefits (increasing hit rates) and costs (increasing false alarm rates). At a moderate level, attention improves sensitivity (i.e., increases $d^\prime$) at only a moderate increase in bias for tasks involving standard images, blended images, and natural adversarial images chosen to fool DCNNs. These results suggest that goal-directed attention can reconfigure general-purpose DCNNs to better suit the current task goal, much like PFC modulates activity along the ventral stream. In addition to being more parsimonious and brain consistent, the mid-level attention approach performed better than a standard machine learning approach for transfer learning, namely retraining the final network layer to accommodate the new task.

Concept induction requires the extraction and naming of concepts from noisy perceptual experience. For supervised approaches, as the number of concepts grows, so does the number of required training examples. Philosophers, psychologists, and computer scientists, have long recognized that children can learn to label objects without being explicitly taught. In a series of computational experiments, we highlight how information in the environment can be used to build and align conceptual systems. Unlike supervised learning, the learning problem becomes easier the more concepts and systems there are to master. The key insight is that each concept has a unique signature within one conceptual system (e.g., images) that is recapitulated in other systems (e.g., text or audio). As predicted, children's early concepts form readily aligned systems., Comment: This is a post-peer-review, pre-copyedit version of an article published in Nature Machine Intelligence. The final authenticated version is available online at: https://doi.org/10.1038/s42256-019-0132-2

Meaning may arise from an element's role or interactions within a larger system. For example, hitting nails is more central to people's concept of a hammer than its particular material composition or other intrinsic features. Likewise, the importance of a web page may result from its links with other pages rather than solely from its content. One example of meaning arising from extrinsic relationships are approaches that extract the meaning of word concepts from co-occurrence patterns in large, text corpora. The success of these methods suggest that human activity patterns may reveal conceptual organization. However, texts do not directly reflect human activity, but instead serve a communicative function and are usually highly curated or edited to suit an audience. Here, we apply methods devised for text to a data source that directly reflects thousands of individuals' activity patterns, namely supermarket purchases. Using product co-occurrence data from nearly 1.3m shopping baskets, we trained a topic model to learn 25 high-level concepts (or "topics"). These topics were found to be comprehensible and coherent by both retail experts and consumers. Topics ranged from specific (e.g., ingredients for a stir-fry) to general (e.g., cooking from scratch). Topics tended to be goal-directed and situational, consistent with the notion that human conceptual knowledge is tailored to support action. Individual differences in the topics sampled predicted basic demographic characteristics. These results suggest that human activity patterns reveal conceptual organization and may give rise to it., Comment: 7 pages, 4 figures

Partisan gerrymandering poses a threat to democracy. Moreover, the complexity of the districting task may exceed human capacities. One potential solution is using computational models to automate the districting process by optimizing objective and open criteria, such as how spatially compact districts are. We formulated one such model that minimized pairwise distance between voters within a district. Using US Census Bureau data, we confirmed our prediction that the difference in compactness between the computed and actual districts would be greatest for states that are large and therefore difficult for humans to properly district given their limited capacities. The computed solutions highlighted differences in how humans and machines solve this task with machine solutions more fully optimized and displaying emergent properties not evident in human solutions. These results suggest a division of labour in which humans debate and formulate districting criteria whereas machines optimize the criteria to draw the district boundaries. We discuss how criteria can be expanded beyond notions of compactness to include other factors, such as respecting municipal boundaries, historic communities, relevant legislation, etc.