There has recently been a surge of interest in Wearable Assistant Agents: agents embodied in a wearable form factor such as smart glasses, who can take actions toward a user’s stated goal — a high-level language-expressed command such as “where did I leave my keys?”, “Text Alice I will be late”, or “What’s the weather in Cancun?”. In this work, we consider the complementary problem of eliminating the effort required to interact with such an agent by proactively inferring the user’s goal from multimodal contextual observations. As vision-language models (VLMs) hold strong potential to ultimately solve this problem, our work focuses on creating a strong benchmark to measure progress toward this end. Given the limited prior work in this area, establishing the benchmark required collecting a novel multimodal goal-inference dataset; our dataset comprises ~30 hours of data from 363 participants across 3,482 recordings, featuring ground-truth reference goals alongside accompanying visual, audio, digital, and longitudinal contextual observations. We ran a human predictability study, where we found that humans set a strong baseline that comprises a de facto upper bound on model performance: they show multiple choice question (MCQ) accuracy of 93%, with the best VLM achieving about 84% accuracy. However, MCQ assesses discrimination, not the model’s ultimate task of generating the goal through open-ended text generation. Through a meta-evaluation, we find that a VLM judging the generated goals is as good as a human judge if it has access to a human-authored script of the video or a correct reference goal. Finally, we evaluate several families of modern vision-language models on the benchmark, showing that larger models have a significant performance advantage, but are still far from being practically useful, as they produce relevant goals only ~57% of the time. The best-performing smaller models—whose size makes them better suited to wearable applications—perform significantly worse than their counterparts, generating ~49% accuracy on the benchmark. Through a modality ablation, we show that models benefit from extra information in relevant modalities with minimal performance degradation from irrelevant modalities, but don’t gain as much when noisy modalities are included (e.g., in the case of digital context when most of the app state is irrelevant).

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 ǣshortcutsǥ 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.

Advances in supervised learning approaches to 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 a large set of human similarity judgments that supplements the existing ILSVRC validation set. The new dataset supports a range of task and performance metrics, including evaluation of unsupervised 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. We were able to scale to the full 50,000 image ILSVRC validation set through a selective sampling process that used variational Bayesian inference and model ensembles to sample aspects of the embedding space that were most uncertain. 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 the 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. ImageNet-HSJ supports the appraisal of internal representations and the development of more humanlike models.

Abstract Our goal is to understand and optimize human concept learning by predicting the ease of learning of a particular exemplar or category. We propose a method for estimating ease values, quantitative measures of ease of learning, as an alternative to conducting costly empirical training studies. Our method combines a psychological embedding of domain exemplars with a pragmatic categorization model. The two components are integrated using a radial basis function network (RBFN) that predicts ease values. The free parameters of the RBFN are fit using human similarity judgments, circumventing the need to collect human training data to fit more complex models of human categorization. We conduct two category-training experiments to validate predictions of the RBFN. We demonstrate that an instance-based RBFN outperforms both a prototype-based RBFN and an empirical approach using the raw data. Although the human data were collected across diverse experimental conditions, the predicted ease values strongly correlate with human learning performance. Training can be sequenced by (predicted) ease, achieving what is known as fading in the psychology literature and curriculum learning in the machine-learning literature, both of which have been shown to facilitate learning.

Deep networks are increasingly being applied to problems involving image synthesis, e.g., generating images from textual descriptions and reconstructing an input image from a compact representation. Supervised training of image-synthesis networks typically uses a pixel-wise loss (PL) to indicate the mismatch between a generated image and its corresponding target image. We propose instead to use a loss function that is better calibrated to human perceptual judgments of image quality: the multiscale structural-similarity score (MS-SSIM) [1]. Because MS-SSIM is differentiable, it is easily incorporated into gradient-descent learning. We compare the consequences of using MS-SSIM versus PL loss on training autoencoders. Human observers reliably prefer images synthesized by MS-SSIM-optimized models over those synthesized by PL-optimized models, for two distinct PL measures (L1and L2distances). We also explore the effect of training objective on image encoding and analyze conditions under which perceptually-optimized representations yield better performance on image classification. Finally, we demonstrate the superiority of perceptually-optimized networks for super-resolution imaging. We argue that significant advances can be made in modeling images through the use of training objectives that are well aligned to characteristics of human perception.

We use Bayesian optimization methods to design games that maximize user engagement. Participants are paid to try a game for several minutes, at which point they can quit or continue to play voluntarily with no further compensation. Engagement is measured by player persistence, projections of how long others will play, and a post-game survey. Using Gaussian process surrogate-based optimization, we conduct efficient experiments to identify game design characteristics---specifically those influencing difficulty---that lead to maximal engagement. We study two games requiring trajectory planning, the difficulty of each is determined by a three-dimensional continuous design space. Two of the design dimensions manipulate the game in user-transparent manner (e.g., the spacing of obstacles), the third in a subtle and possibly covert manner (incremental trajectory corrections). Converging results indicate that overt difficulty manipulations are effective in modulating engagement only when combined with the covert manipulation, suggesting the critical role of a user's self-perception of competence.