Knowledge distillation deploys complex machine learning models in resource-constrained environments by training a smaller student model to emulate internal representations of a complex teacher model. However, the teacher’s representations can also encode nuisance or additional information not relevant to the downstream task. Distilling such irrelevant information can actually impede the performance of a capacity-limited student model. This observation motivates our primary question: What are the information-theoretic limits of knowledge distillation? To this end, we leverage Partial Information Decomposition to quantify and explain the transferred knowledge and knowledge left to distill for a downstream task. We theoretically demonstrate that the task-relevant transferred knowledge is succinctly captured by the measure of redundant information about the task between the teacher and student. We propose a novel multi-level optimization to incorporate redundant information as a regularizer, leading to our framework of Redundant Information Distillation (RID). RID leads to more resilient and effective distillation under nuisance teachers as it succinctly quantifies task-relevant knowledge rather than simply aligning student and teacher representations.

In order for AI systems to communicate effectively with people, they must understand how we make decisions. However, people's decisions are not always rational, so the implicit internal models of human decision-making in Large Language Models (LLMs) must account for this. Previous empirical evidence seems to suggest that these implicit models are accurate --- LLMs offer believable proxies of human behavior, acting how we expect humans would in everyday interactions. However, by comparing LLM behavior and predictions to a large dataset of human decisions, we find that this is actually not the case: when both simulating and predicting people's choices, a suite of cutting-edge LLMs (GPT-4o \& 4-Turbo, Llama-3-8B \& 70B, Claude 3 Opus) assume that people are more rational than we really are. Specifically, these models deviate from human behavior and align more closely with a classic model of rational choice --- expected value theory. Interestingly, people also tend to assume that other people are rational when interpreting their behavior. As a consequence, when we compare the inferences that LLMs and people draw from the decisions of others using another psychological dataset, we find that these inferences are highly correlated. Thus, the implicit decision-making models of LLMs appear to be aligned with the human expectation that other people will act rationally, rather than with how people actually act.

Chain-of-thought (CoT) prompting has become a widely used strategy for improving large language and multimodal model performance. However, it is still an open question under which settings CoT systematically reduces performance. In this paper, we seek to identify the characteristics of tasks where CoT reduces performance by drawing inspiration from cognitive psychology, focusing on six representative tasks from the psychological literature where deliberation hurts performance in humans. In three of these tasks, state-of-the-art models exhibit significant performance drop-offs with CoT (up to 36.3% absolute accuracy for OpenAI o1-preview compared to GPT-4o), while in others, CoT effects are mixed, with positive, neutral, and negative changes. While models and humans do not exhibit perfectly parallel cognitive processes, considering cases where thinking has negative consequences for humans helps identify settings where it negatively impacts models. By connecting the literature on human verbal thinking and deliberation with evaluations of CoT, we offer a perspective for understanding the impact of inference-time reasoning.

Conversational tones -the manners and attitudes in which speakers communicate -are essential to effective communication.Amidst the increasing popularization of Large Language Models (LLMs) over recent years, it becomes necessary to characterize the divergences in their conversational tones relative to humans.However, existing investigations of conversational modalities rely on pre-existing taxonomies or text corpora, which suffer from experimenter bias and may not be representative of real-world distributions for the studies' psycholinguistic domains.Inspired by methods from cognitive science, we propose an iterative method for simultaneously eliciting conversational tones and sentences, where participants alternate between two tasks: (1) one participant identifies the tone of a given sentence and (2) a different participant generates a sentence based on that tone.We run 100 iterations of this process with human participants and GPT-4, then obtain a dataset of sentences and frequent conversational tones.In an additional experiment, humans and GPT-4 annotated all sentences with all tones.With data from 1,339 human participants, 33,370 human judgments, and 29,900 GPT-4 queries, we show how our approach can be used to create an interpretable geometric representation of relations between conversational tones in humans and GPT-4.This work demonstrates how combining ideas from machine learning and cognitive science can address challenges in human-computer interactions.B Sampling paradigm A Problem statement C Quality-of-fit rating D Shared space E Benchmark What are similarities and divergences in conversational tones in humans and LLMs?I am pretty certain, Tom ate all the cookies from the jar.Could it be possible that Tom ate all the cookies from the jar? polite?Humans

Learning from demonstrations is a common way for users to teach robots, but it is prone to spurious feature correlations. Recent work constructs state abstractions, i.e. visual representations containing task-relevant features, from language as a way to perform more generalizable learning. However, these abstractions also depend on a user's preference for what matters in a task, which may be hard to describe or infeasible to exhaustively specify using language alone. How do we construct abstractions to capture these latent preferences? We observe that how humans behave reveals how they see the world. Our key insight is that changes in human behavior inform us that there are differences in preferences for how humans see the world, i.e. their state abstractions. In this work, we propose using language models (LMs) to query for those preferences directly given knowledge that a change in behavior has occurred. In our framework, we use the LM in two ways: first, given a text description of the task and knowledge of behavioral change between states, we query the LM for possible hidden preferences; second, given the most likely preference, we query the LM to construct the state abstraction. In this framework, the LM is also able to ask the human directly when uncertain about its own estimate. We demonstrate our framework's ability to construct effective preference-conditioned abstractions in simulated experiments, a user study, as well as on a real Spot robot performing mobile manipulation tasks.

We describe a framework for using natural language to design state abstractions for imitation learning. Generalizable policy learning in high-dimensional observation spaces is facilitated by well-designed state representations, which can surface important features of an environment and hide irrelevant ones. These state representations are typically manually specified, or derived from other labor-intensive labeling procedures. Our method, LGA (\textit{language-guided abstraction}), uses a combination of natural language supervision and background knowledge from language models (LMs) to automatically build state representations tailored to unseen tasks. In LGA, a user first provides a (possibly incomplete) description of a target task in natural language; next, a pre-trained LM translates this task description into a state abstraction function that masks out irrelevant features; finally, an imitation policy is trained using a small number of demonstrations and LGA-generated abstract states. Experiments on simulated robotic tasks show that LGA yields state abstractions similar to those designed by humans, but in a fraction of the time, and that these abstractions improve generalization and robustness in the presence of spurious correlations and ambiguous specifications. We illustrate the utility of the learned abstractions on mobile manipulation tasks with a Spot robot.

How can we build AI systems that can learn any set of individual human values both quickly and safely, avoiding causing harm or violating societal standards for acceptable behavior during the learning process? We explore the effects of representational alignment between humans and AI agents on learning human values. Making AI systems learn human-like representations of the world has many known benefits, including improving generalization, robustness to domain shifts, and few-shot learning performance. We demonstrate that this kind of representational alignment can also support safely learning and exploring human values in the context of personalization. We begin with a theoretical prediction, show that it applies to learning human morality judgments, then show that our results generalize to ten different aspects of human values -- including ethics, honesty, and fairness -- training AI agents on each set of values in a multi-armed bandit setting, where rewards reflect human value judgments over the chosen action. Using a set of textual action descriptions, we collect value judgments from humans, as well as similarity judgments from both humans and multiple language models, and demonstrate that representational alignment enables both safe exploration and improved generalization when learning human values.

Placing a human in the loop may help abate the risks of deploying AI systems in safety-critical settings (e.g., a clinician working with a medical AI system). However, mitigating risks arising from human error and uncertainty within such human-AI interactions is an important and understudied issue. In this work, we study human uncertainty in the context of concept-based models, a family of AI systems that enable human feedback via concept interventions where an expert intervenes on human-interpretable concepts relevant to the task. Prior work in this space often assumes that humans are oracles who are always certain and correct. Yet, real-world decision-making by humans is prone to occasional mistakes and uncertainty. We study how existing concept-based models deal with uncertain interventions from humans using two novel datasets: UMNIST, a visual dataset with controlled simulated uncertainty based on the MNIST dataset, and CUB-S, a relabeling of the popular CUB concept dataset with rich, densely-annotated soft labels from humans. We show that training with uncertain concept labels may help mitigate weaknesses of concept-based systems when handling uncertain interventions. These results allow us to identify several open challenges, which we argue can be tackled through future multidisciplinary research on building interactive uncertainty-aware systems. To facilitate further research, we release a new elicitation platform, UElic, to collect uncertain feedback from humans in collaborative prediction tasks.

Diffusion models are a class of generative models that learn to synthesize samples by inverting a diffusion process that gradually maps data into noise. While these models have enjoyed great success recently, a full theoretical understanding of their observed properties is still lacking, in particular, their weak sensitivity to the choice of noise family and the role of adequate scheduling of noise levels for good synthesis. By identifying a correspondence between diffusion models and a well-known paradigm in cognitive science known as serial reproduction, whereby human agents iteratively observe and reproduce stimuli from memory, we show how the aforementioned properties of diffusion models can be explained as a natural consequence of this correspondence. We then complement our theoretical analysis with simulations that exhibit these key features. Our work highlights how classic paradigms in cognitive science can shed light on state-of-the-art machine learning problems.

Should we care whether AI systems have representations of the world that are similar to those of humans? We provide an information-theoretic analysis that suggests that there should be a U-shaped relationship between the degree of representational alignment with humans and performance on few-shot learning tasks. We confirm this prediction empirically, finding such a relationship in an analysis of the performance of 491 computer vision models. We also show that highly-aligned models are more robust to both natural adversarial attacks and domain shifts. Our results suggest that human-alignment is often a sufficient, but not necessary, condition for models to make effective use of limited data, be robust, and generalize well.

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.

Supervised learning typically focuses on learning transferable representations from training examples annotated by humans. While rich annotations (like soft labels) carry more information than sparse annotations (like hard labels), they are also more expensive to collect. For example, while hard labels only provide information about the closest class an object belongs to (e.g., “this is a dog”), soft labels provide information about the object’s relationship with multiple classes (e.g., “this is most likely a dog, but it could also be a wolf or a coyote”). We use information theory to compare how a number of commonly-used supervision signals contribute to representation-learning performance, as well as how their capacity is affected by factors such as the number of labels, classes, dimensions, and noise. Our framework provides theoretical justification for using hard labels in the big-data regime, but richer supervision signals for few-shot learning and out-of-distribution generalization. We validate these results empirically in a series of experiments with over 1 million crowdsourced image annotations and conduct a cost-benefit analysis to establish a tradeoff curve that enables users to optimize the cost of supervising representation learning on their own datasets.

Deep neural networks require large training sets but suffer from high computational cost and long training times. Training on much smaller training sets while maintaining nearly the same accuracy would be very beneficial. In the few-shot learning setting, a model must learn a new class given only a small number of samples from that class. One-shot learning is an extreme form of few-shot learning where the model must learn a new class from a single example. We propose the 'less than one'-shot learning task where models must learn N new classes given only M

We leverage what are typically considered the worst qualities of deep learning algorithms - high computational cost, requirement for large data, no explainability, high dependence on hyper-parameter choice, overfitting, and vulnerability to adversarial perturbations - in order to create a method for the secure and efficient training of remotely deployed neural networks over unsecure channels.

Increasingly large datasets are rapidly driving up the computational costs of machine learning. Prototype generation methods aim to create a small set of synthetic observations that accurately represent a training dataset but greatly reduce the computational cost of learning from it. Assigning soft labels to prototypes can allow increasingly small sets of prototypes to accurately represent the original training dataset. Although foundational work on ‘less than one’ -shot learning has proven the theoretical plausibility of learning with fewer than one observation per class, developing practical algorithms for generating such prototypes remains an unexplored territory. We propose a novel, modular method for generating soft-label prototypical lines that still maintains representational accuracy even when there are fewer prototypes than the number of classes in the data. In addition, we propose the Hierarchical Soft-Label Prototype k-Nearest Neighbor classification algorithm based on these prototypical lines. We show that our method maintains high classification accuracy while greatly reducing the number of prototypes required to represent a dataset, even when working with severely imbalanced and difficult data. Our code is available at https://github.com/ilia10000/SLkNN.

Dataset distillation is a method for reducing dataset sizes by learning a small number of representative synthetic samples. This has several benefits such as speeding up model training, reducing energy consumption, and reducing required storage space. These benefits are especially crucial in settings like federated learning where initial overhead costs are justified by the speedup they enable. Currently, 1) each synthetic sample is assigned a single ‘hard’ label, and 2) dataset distillation can only be used with image data. We propose to simultaneously distill both images and their labels, thus assigning each synthetic sample a ‘soft’ label (a distribution of labels). Our algorithm increases accuracy by 2-4% for several image classification tasks. Using ‘soft’ labels also enables distilled datasets to consist of fewer samples than there are classes as each sample encodes information for multiple classes. For example, training a LeNet model with 10 distilled images (one per class) results in over 96% accuracy on MNIST, and almost 92% accuracy when trained on just 5 distilled images. We also extend the dataset distillation algorithm to distill text data. We demonstrate that text distillation outperforms other methods across multiple datasets. For example, models attain almost their original accuracy on the IMDB sentiment analysis task using just 20 distilled sentences. Our code can be found at https://github.com/ilia10000/dataset-distillation.

Most real-world datasets, and particularly those collected from physical systems, are full of noise, packet loss, and other imperfections. However, most specification mining, anomaly detection and other such algorithms assume, or even require, perfect data quality to function properly. Such algorithms may work in lab conditions when given clean, controlled data, but will fail in the field when given imperfect data. We propose a method for accurately reconstructing discrete temporal or sequential system traces affected by data loss, using Long Short-Term Memory Networks (LSTMs). The model works by learning to predict the next event in a sequence of events, and uses its own output as an input to continue predicting future events. As a result, this method can be used for data restoration even with streamed data. Such a method can reconstruct even long sequence of missing events, and can also help validate and improve data quality for noisy data. The output of the model will be a close reconstruction of the true data, and can be fed to algorithms that rely on clean data. We demonstrate our method by reconstructing automotive CAN traces consisting of long sequences of discrete events. We show that given even small parts of a CAN trace, our LSTM model can predict future events with an accuracy of almost 90%, and can successfully reconstruct large portions of the original trace, greatly outperforming a Markov Model benchmark. We separately feed the original, lossy, and reconstructed traces into a specification mining framework to perform downstream analysis of the effect of our method on state-of-the-art models that use these traces for understanding the behavior of complex systems.