Brian Hu

As AI systems become both more powerful and prevalent, ensuring that their actions align with human values is paramount. The challenge of AI alignment is thus an interdisciplinary one that involves not only a technical challenge for computer science but one with important ties to the psychology of moral values, decision-making, and trust. Early work identified a static set of universal values, without considering the key questions of to whom and to which values AI should be aligned. This perspective paper challenges the notion of universal alignment and instead argues for dynamic, context-specific alignability across different domains, tasks, and users. Specifically, we emphasize the need to go beyond traditional pluralism and rethink how AI alignment can be achieved through a qualitative and quantitative research process that involves identifying context-specific values, developing alignable AI algorithms using limited human feedback, and evaluating alignment through assessing both an AI’s values and actions, while considering how humans trust and delegate to the AI. We discuss several paths forward for our proposed framework, including the potential ethical and societal implications of context-specific alignability, and draw on examples ranging from chatbots to value-aligned decision-making in the medical triage domain.

Large language models (LLMs) are currently aligned using techniques such as reinforcement learning from human feedback (RLHF). However, these methods use scalar rewards that can only reflect user preferences on average. Pluralistic alignment instead seeks to capture diverse user preferences across a set of attributes, moving beyond just helpfulness and harmlessness. Toward this end, we propose a steerable pluralistic model based on few-shot comparative regression that can adapt to individual user preferences. Our approach leverages in-context learning and reasoning, grounded in a set of fine-grained attributes, to compare response options and make aligned choices. To evaluate our algorithm, we also propose two new steerable pluralistic benchmarks by adapting the Moral Integrity Corpus (MIC) and the HelpSteer2 datasets, demonstrating the applicability of our approach to value-aligned decision-making and reward modeling, respectively. Our few-shot comparative regression approach is interpretable and compatible with different attributes and LLMs, while outperforming multiple baseline and state-of-the-art methods. Our work provides new insights and research directions in pluralistic alignment, enabling a more fair and representative use of LLMs and advancing the state-of-the-art in ethical AI.

Large language models (LLMs) are currently aligned using techniques such as reinforcement learning from human feedback (RLHF). However, these methods use scalar rewards that can only reflect user preferences on average. Pluralistic alignment instead seeks to capture diverse user preferences across a set of attributes, moving beyond just helpfulness and harmlessness. Toward this end, we propose a steerable pluralistic model based on few-shot comparative regression that can adapt to individual user preferences. Our approach leverages in-context learning and reasoning, grounded in a set of fine-grained attributes, to compare response options and make aligned choices. To evaluate our algorithm, we also propose two new steerable pluralistic benchmarks by adapting the Moral Integrity Corpus (MIC) and the HelpSteer2 datasets, demonstrating the applicability of our approach to value-aligned decision-making and reward modeling, respectively. Our few-shot comparative regression approach is interpretable and compatible with different attributes and LLMs, while outperforming multiple baseline and state-of-the-art methods. Our work provides new insights and research directions in pluralistic alignment, enabling a more fair and representative use of LLMs and advancing the state-of-the-art in ethical AI.

Advances in artificial intelligence (AI) using techniques such as deep learning have fueled the recent progress in fields such as computer vision. However, these algorithms are still often viewed as "black boxes", which cannot easily explain how they arrived at their final output decisions. Saliency maps are one commonly used form of explainable AI (XAI), which indicate the input features an algorithm paid attention to during its decision process. Here, we introduce the open source xaitk-saliency package, an XAI framework and toolkit for saliency. We demonstrate its modular and flexible nature by highlighting two example use cases for saliency maps: (1) object detection model comparison and (2) doppelganger saliency for person re-identification. We also show how the xaitk-saliency package can be paired with visualization tools to support the interactive exploration of saliency maps. Our results suggest that saliency maps may play a critical role in the verification and validation of AI models, ensuring their trusted use and deployment. The code is publicly available at: https://github.com/xaitk/xaitk-saliency.

Detecting novel stimuli in the environment is critical for learning and survival, yet the neural basis of novelty processing is not understood. To characterize cell type-specific novelty processing, we surveyed the activity of ∼15,000 excitatory and inhibitory neurons in mice performing a visual task with novel and familiar stimuli. Clustering revealed a dozen functional neuron types defined by experience-dependent encoding. Vasoactive-intestinal-peptide (Vip) expressing inhibitory neurons were diverse, encoding novel stimuli, omissions of familiar stimuli, or behavioral features. Distinct Somatostatin (Sst) expressing inhibitory neurons encoded either familiar or novel stimuli. Subsets of excitatory neurons co-clustered with specific Vip or Sst subpopulations, while Sst and Vip inhibitory clusters were non-overlapping. This study establishes that novelty processing is mediated by diverse functional neuron types in the visual cortex.

As animals adapt to their environments, their brains are tasked with processing stimuli in different sensory contexts. Whether these computations are context dependent or independent, they are all implemented in the same neural tissue. A crucial question is what neural architectures can respond flexibly to a range of stimulus conditions and switch between them. This is a particular case of flexible architecture that permits multiple related computations within a single circuit. Here, we address this question in the specific case of the visual system circuitry, focusing on context integration, defined as the integration of feedforward and surround information across visual space. We show that a biologically inspired microcircuit with multiple inhibitory cell types can switch between visual processing of the static context and the moving context. In our model, the VIP population acts as the switch and modulates the visual circuit through a disinhibitory motif. Moreover, the VIP population is efficient, requiring only a relatively small number of neurons to switch contexts. This circuit eliminates noise in videos by using appropriate lateral connections for contextual spatiotemporal surround modulation, having superior denoising performance compared to circuits where only one context is learned. Our findings shed light on a minimally complex architecture that is capable of switching between two naturalistic contexts using few switching units.

Inhibitory neurons in mammalian cortex exhibit diverse physiological, morphological, molecular, and connectivity signatures. While considerable work has measured the average connectivity of several interneuron classes, there remains a fundamental lack of understanding of the connectivity distribution of distinct inhibitory cell types with synaptic resolution, how it relates to properties of target cells, and how it affects function. Here, we used large-scale electron microscopy and functional imaging to address these questions for chandelier cells in layer 2/3 of the mouse visual cortex. With dense reconstructions from electron microscopy, we mapped the complete chandelier input onto 153 pyramidal neurons. We found that synapse number is highly variable across the population and is correlated with several structural features of the target neuron. This variability in the number of axo-axonic ChC synapses is higher than the variability seen in perisomatic inhibition. Biophysical simulations show that the observed pattern of axo-axonic inhibition is particularly effective in controlling excitatory output when excitation and inhibition are co-active. Finally, we measured chandelier cell activity in awake animals using a cell-type-specific calcium imaging approach and saw highly correlated activity across chandelier cells. In the same experiments, in vivo chandelier population activity correlated with pupil dilation, a proxy for arousal. Together, these results suggest that chandelier cells provide a circuit-wide signal whose strength is adjusted relative to the properties of target neurons.

Quantifying the value of explanations in a human‐in‐the‐loop (HITL) system is difficult. Previous methods either measure explanation‐specific values that do not correspond to user tasks and needs or poll users on how useful they find the explanations to be. In this work, we quantify how much explanations help the user through a utility‐based paradigm that measures change in task performance when using explanations vs not. Our chosen task is content‐based image retrieval (CBIR), which has well‐established baselines and performance metrics independent of explainability. We extend an existing HITL image retrieval system that incorporates user feedback with similarity‐based saliency maps (SBSM) that indicate to the user which parts of the retrieved images are most similar to the query image. The system helps the user understand what it is paying attention to through saliency maps, and the user helps the system understand their goal through saliency‐guided relevance feedback. Using the MS‐COCO dataset, a standard object detection and segmentation dataset, we conducted extensive, crowd‐sourced experiments validating that SBSM improves interactive image retrieval. Although the performance increase is modest in the general case, in more difficult cases such as cluttered scenes, using explanations yields an 6.5% increase in accuracy. To the best of our knowledge, this is the first large‐scale user study showing that visual saliency map explanations improve performance on a real‐world, interactive task. Our utility‐based evaluation paradigm is general and potentially applicable to any task for which explainability can be incorporated.

Recent advances in artificial intelligence (AI), driven mainly by deep neural networks, have yielded remarkable progress in fields, such as computer vision, natural language processing, and reinforcement learning. Despite these successes, the inability to predict how AI systems will behave “in the wild” impacts almost all stages of planning and deployment, including research and development, verification and validation, and user trust and acceptance. The field of explainable artificial intelligence (XAI) seeks to develop techniques enabling AI algorithms to generate explanations of their results; generally these are human‐interpretable representations or visualizations that are meant to “explain” how the system produced its outputs. We introduce the Explainable AI Toolkit (XAITK), a DARPA‐sponsored effort that builds on results from the 4‐year DARPA XAI program. The XAITK has two goals: (a) to consolidate research results from DARPA XAI into a single publicly accessible repository; and (b) to identify operationally relevant capabilities developed on DARPA XAI and assist in their transition to interested partners. We first describe the XAITK website and associated capabilities. These place the research results from DARPA XAI in the wider context of general research in the field of XAI, and include performer contributions of code, data, publications, and reports. We then describe the XAITK analytics and autonomy software frameworks. These are Python‐based frameworks focused on particular XAI domains, and designed to provide a single integration endpoint for multiple algorithm implementations from across DARPA XAI. Each framework generalizes APIs for system‐level data and control while providing a plugin interface for existing and future algorithm implementations. The XAITK project can be followed at: https://xaitk.org.

The maintenance of short-term memories is critical for survival in a dynamically changing world. Previous studies suggest that this memory can be stored in the form of persistent neural activity or using a synaptic mechanism, such as with short-term plasticity. Here, we compare the predictions of these two mechanisms to neural and behavioral measurements in a visual change detection task. Mice were trained to respond to changes in a repeated sequence of natural images while neural activity was recorded using two-photon calcium imaging. We also trained two types of artificial neural networks on the same change detection task as the mice. Following fixed pre-processing using a pretrained convolutional neural network, either a recurrent neural network (RNN) or a feedforward neural network with short-term synaptic depression (STPNet) was trained to the same level of performance as the mice. While both networks are able to learn the task, the STPNet model contains units whose activity are more similar to the in vivo data and produces errors which are more similar to the mice. When images are omitted, an unexpected perturbation which was absent during training, mice often do not respond to the omission but are more likely to respond to the subsequent image. Unlike the RNN model, STPNet produces a similar pattern of behavior. These results suggest that simple neural adaptation mechanisms may serve as an important bottom-up memory signal in this task, which can be used by downstream areas in the decision-making process.