skip to main content


Title: Transforming neural network representations to predict human judgments of similarity
Deep-learning vision models have shown intriguing similarities and differences with respect to human vision. We investigate how to bring machine visual represen- tations into better alignment with human representations. Human representations are often inferred from behavioral evidence such as the selection of an image most similar to a query image. We find that with appropriate linear transformations of deep embeddings, we can improve prediction of human binary choice on a data set of bird images from 72% at baseline to 89%. We hypothesized that deep embeddings have redundant, high (4096) dimensional representations; however, reducing the rank of these representations results in a loss of explanatory power. We hypothesized that the dilation transformation of representations explored in past research is too restrictive, and indeed we found that model explanatory power can be significantly improved with a more expressive linear transform. Most surprising and exciting, we found that, consistent with classic psychological literature, human similarity judgments are asymmetric: the similarity of X to Y is not necessarily equal to the similarity of Y to X, and allowing models to express this asymmetry improves explanatory power.  more » « less
Award ID(s):
1631428
NSF-PAR ID:
10299597
Author(s) / Creator(s):
Date Published:
Journal Name:
Workshop on Shared Visual Representations in Human and Machine Intelligence (SVRHM 2020)
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract

    As inspirational stimuli can assist designers with achieving enhanced design outcomes, supporting the retrieval of impactful sources of inspiration is important. Existing methods facilitating this retrieval have relied mostly on semantic relationships, e.g., analogical distances. Increasingly, data-driven methods can be leveraged to represent diverse stimuli in terms of multi-modal information, enabling designers to access stimuli in terms of less explored, non-text-based relationships. Toward improved retrieval of multi-modal representations of inspirational stimuli, this work compares human-evaluated and computationally derived similarities between stimuli in terms of non-text-based visual and functional features. A human subjects study (n = 36) was conducted where similarity assessments between triplets of 3D-model parts were collected and used to construct psychological embedding spaces. Distances between unique part embeddings were used to represent similarities in terms of visual and functional features. Obtained distances were compared with computed distances between embeddings of the same stimuli generated using artificial intelligence (AI)-based deep-learning approaches. When used to assess similarity in appearance and function, these representations were found to be largely consistent, with highest agreement found when assessing pairs of stimuli with low similarity. Alignment between models was otherwise lower when identifying the same pairs of stimuli with higher levels of similarity. Importantly, qualitative data also revealed insights regarding how humans made similarity assessments, including more abstract information not captured using AI-based approaches. Toward providing inspiration to designers that considers design problems, ideas, and solutions in terms of non-text-based relationships, further exploration of how these relationships are represented and evaluated is encouraged.

     
    more » « less
  2. Many sign languages are bona fide natural languages with grammatical rules and lexicons hence can benefit from machine translation methods. Similarly, since sign language is a visual-spatial language, it can also benefit from computer vision methods for encoding it. With the advent of deep learning methods in recent years, significant advances have been made in natural language processing (specifically neural machine translation) and in computer vision methods (specifically image and video captioning). Researchers have therefore begun expanding these learning methods to sign language understanding. Sign language interpretation is especially challenging, because it involves a continuous visual-spatial modality where meaning is often derived based on context. The focus of this article, therefore, is to examine various deep learning–based methods for encoding sign language as inputs, and to analyze the efficacy of several machine translation methods, over three different sign language datasets. The goal is to determine which combinations are sufficiently robust for sign language translation without any gloss-based information. To understand the role of the different input features, we perform ablation studies over the model architectures (input features + neural translation models) for improved continuous sign language translation. These input features include body and finger joints, facial points, as well as vector representations/embeddings from convolutional neural networks. The machine translation models explored include several baseline sequence-to-sequence approaches, more complex and challenging networks using attention, reinforcement learning, and the transformer model. We implement the translation methods over multiple sign languages—German (GSL), American (ASL), and Chinese sign languages (CSL). From our analysis, the transformer model combined with input embeddings from ResNet50 or pose-based landmark features outperformed all the other sequence-to-sequence models by achieving higher BLEU2-BLEU4 scores when applied to the controlled and constrained GSL benchmark dataset. These combinations also showed significant promise on the other less controlled ASL and CSL datasets. 
    more » « less
  3. Abstract Background

    In Alzheimer’s Diseases (AD) research, multimodal imaging analysis can unveil complementary information from multiple imaging modalities and further our understanding of the disease. One application is to discover disease subtypes using unsupervised clustering. However, existing clustering methods are often applied to input features directly, and could suffer from the curse of dimensionality with high-dimensional multimodal data. The purpose of our study is to identify multimodal imaging-driven subtypes in Mild Cognitive Impairment (MCI) participants using a multiview learning framework based on Deep Generalized Canonical Correlation Analysis (DGCCA), to learn shared latent representation with low dimensions from 3 neuroimaging modalities.

    Results

    DGCCA applies non-linear transformation to input views using neural networks and is able to learn correlated embeddings with low dimensions that capture more variance than its linear counterpart, generalized CCA (GCCA). We designed experiments to compare DGCCA embeddings with single modality features and GCCA embeddings by generating 2 subtypes from each feature set using unsupervised clustering. In our validation studies, we found that amyloid PET imaging has the most discriminative features compared with structural MRI and FDG PET which DGCCA learns from but not GCCA. DGCCA subtypes show differential measures in 5 cognitive assessments, 6 brain volume measures, and conversion to AD patterns. In addition, DGCCA MCI subtypes confirmed AD genetic markers with strong signals that existing late MCI group did not identify.

    Conclusion

    Overall, DGCCA is able to learn effective low dimensional embeddings from multimodal data by learning non-linear projections. MCI subtypes generated from DGCCA embeddings are different from existing early and late MCI groups and show most similarity with those identified by amyloid PET features. In our validation studies, DGCCA subtypes show distinct patterns in cognitive measures, brain volumes, and are able to identify AD genetic markers. These findings indicate the promise of the imaging-driven subtypes and their power in revealing disease structures beyond early and late stage MCI.

     
    more » « less
  4. In natural language processing, most models try to learn semantic representations merely from texts. The learned representations encode the “distributional semantics” but fail to connect to any knowledge about the physical world. In contrast, humans learn language by grounding concepts in perception and action and the brain encodes “grounded semantics” for cognition. Inspired by this notion and recent work in vision-language learning, we design a two-stream model for grounding language learning in vision. The model includes a VGG-based visual stream and a Bert-based language stream. The two streams merge into a joint representational space. Through cross-modal contrastive learning, the model first learns to align visual and language representations with the MS COCO dataset. The model further learns to retrieve visual objects with language queries through a cross-modal attention module and to infer the visual relations between the retrieved objects through a bilinear operator with the Visual Genome dataset. After training, the model’s language stream is a stand-alone language model capable of embedding concepts in a visually grounded semantic space. This semantic space manifests principal dimensions explainable with human intuition and neurobiological knowledge. Word embeddings in this semantic space are predictive of human-defined norms of semantic features and are segregated into perceptually distinctive clusters. Furthermore, the visually grounded language model also enables compositional language understanding based on visual knowledge and multimodal image search with queries based on images, texts, or their combinations. 
    more » « less
  5. null (Ed.)
    Deep Learning (DL) has attracted a lot of attention for its ability to reach state-of-the-art performance in many machine learning tasks. The core principle of DL methods consists of training composite architectures in an end-to-end fashion, where inputs are associated with outputs trained to optimize an objective function. Because of their compositional nature, DL architectures naturally exhibit several intermediate representations of the inputs, which belong to so-called latent spaces. When treated individually, these intermediate representations are most of the time unconstrained during the learning process, as it is unclear which properties should be favored. However, when processing a batch of inputs concurrently, the corresponding set of intermediate representations exhibit relations (what we call a geometry) on which desired properties can be sought. In this work, we show that it is possible to introduce constraints on these latent geometries to address various problems. In more detail, we propose to represent geometries by constructing similarity graphs from the intermediate representations obtained when processing a batch of inputs. By constraining these Latent Geometry Graphs (LGGs), we address the three following problems: (i) reproducing the behavior of a teacher architecture is achieved by mimicking its geometry, (ii) designing efficient embeddings for classification is achieved by targeting specific geometries, and (iii) robustness to deviations on inputs is achieved via enforcing smooth variation of geometry between consecutive latent spaces. Using standard vision benchmarks, we demonstrate the ability of the proposed geometry-based methods in solving the considered problems. 
    more » « less