TR2023-014

Are Deep Neural Networks SMARTer than Second Graders?


Abstract:

Recent times have witnessed an increasing number of applications of deep neural networks towards solving tasks that require superior cognitive abilities, e.g., playing Go, generating art, question answering (e.g., ChatGPT), etc. Such a dramatic progress raises the question: how generalizable are neural networks in solving problems that demand broad skills? To answer this question, we propose SMART: a Simple Multimodal Algorithmic Reasoning Task and the associated SMART-101 dataset1, for evaluating the abstraction, deduction, and generalization abilities of neural networks in solving visuo-linguistic puzzles designed specifically for children in the 6–8 age group. Our dataset consists of 101 unique puzzles; each puzzle comprises a picture and a question, and their solution needs a mix of several elementary skills, including arithmetic, algebra, and spatial reasoning, among others. To scale our dataset towards training deep neural networks, we programmatically generate entirely new instances for each puzzle while retaining their solution algorithm. To benchmark the performance on the SMART-101 dataset, we propose a vision-and-language meta-learning model that can incorporate varied state-of- the-art neural backbones. Our experiments reveal that while powerful deep models offer reasonable performances on puzzles in a supervised setting, they are not better than random accuracy when analyzed for generalization – filling this gap may demand new multimodal learning approaches.

 

  • Software & Data Downloads

  • Related News & Events

    •  NEWS    Anoop Cherian gives a podcast interview with AI Business
      Date: September 26, 2023
      Where: Virtual
      MERL Contact: Anoop Cherian
      Research Areas: Artificial Intelligence, Computer Vision, Machine Learning
      Brief
      • Anoop Cherian, a Senior Principal Research Scientist in the Computer Vision team at MERL, gave a podcast interview with award-winning journalist, Deborah Yao. Deborah is the editor of AI Business -- a leading content platform for artificial intelligence and its applications in the real world, delivering its readers up-to-the-minute insights into how AI technologies are currently affecting the global economy and society. The podcast was based on the recent research that Anoop and his colleagues did at MERL with his collaborators at MIT; this research attempts to objectively answer the pertinent question: are current deep neural networks smarter than second graders? The podcast discusses shortcomings in the recent artificial general intelligence systems with regard to their capabilities for knowledge abstraction, learning, and generalization, which are brought out by this research.
    •  
    •  NEWS    MERL researchers presenting four papers and co-organizing a workshop at CVPR 2023
      Date: June 18, 2023 - June 22, 2023
      Where: Vancouver/Canada
      MERL Contacts: Anoop Cherian; Michael J. Jones; Suhas Lohit; Kuan-Chuan Peng
      Research Areas: Artificial Intelligence, Computer Vision, Machine Learning
      Brief
      • MERL researchers are presenting 4 papers and co-organizing a workshop at the CVPR 2023 conference, which will be held in Vancouver, Canada June 18-22. CVPR is one of the most prestigious and competitive international conferences in computer vision. Details are provided below.

        1. “Are Deep Neural Networks SMARTer than Second Graders,” by Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit, Kevin Smith, and Joshua B. Tenenbaum

        We present SMART: a Simple Multimodal Algorithmic Reasoning Task and the associated SMART-101 dataset for evaluating the abstraction, deduction, and generalization abilities of neural networks in solving visuo-linguistic puzzles designed for children in the 6-8 age group. Our experiments using SMART-101 reveal that powerful deep models are not better than random accuracy when analyzed for generalization. We also evaluate large language models (including ChatGPT) on a subset of SMART-101 and find that while these models show convincing reasoning abilities, their answers are often incorrect.

        Paper: https://arxiv.org/abs/2212.09993

        2. “EVAL: Explainable Video Anomaly Localization,” by Ashish Singh, Michael J. Jones, and Erik Learned-Miller

        This work presents a method for detecting unusual activities in videos by building a high-level model of activities found in nominal videos of a scene. The high-level features used in the model are human understandable and include attributes such as the object class and the directions and speeds of motion. Such high-level features allow our method to not only detect anomalous activity but also to provide explanations for why it is anomalous.

        Paper: https://arxiv.org/abs/2212.07900

        3. "Aligning Step-by-Step Instructional Diagrams to Video Demonstrations," by Jiahao Zhang, Anoop Cherian, Yanbin Liu, Yizhak Ben-Shabat, Cristian Rodriguez, and Stephen Gould

        The rise of do-it-yourself (DIY) videos on the web has made it possible even for an unskilled person (or a skilled robot) to imitate and follow instructions to complete complex real world tasks. In this paper, we consider the novel problem of aligning instruction steps that are depicted as assembly diagrams (commonly seen in Ikea assembly manuals) with video segments from in-the-wild videos. We present a new dataset: Ikea Assembly in the Wild (IAW) and propose a contrastive learning framework for aligning instruction diagrams with video clips.

        Paper: https://arxiv.org/pdf/2303.13800.pdf

        4. "HaLP: Hallucinating Latent Positives for Skeleton-Based Self-Supervised Learning of Actions," by Anshul Shah, Aniket Roy, Ketul Shah, Shlok Kumar Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa

        In this work, we propose a new contrastive learning approach to train models for skeleton-based action recognition without labels. Our key contribution is a simple module, HaLP: Hallucinating Latent Positives for contrastive learning. HaLP explores the latent space of poses in suitable directions to generate new positives. Our experiments using HaLP demonstrates strong empirical improvements.

        Paper: https://arxiv.org/abs/2304.00387

        The 4th Workshop on Fair, Data-Efficient, and Trusted Computer Vision

        MERL researcher Kuan-Chuan Peng is co-organizing the fourth Workshop on Fair, Data-Efficient, and Trusted Computer Vision (https://fadetrcv.github.io/2023/) in conjunction with CVPR 2023 on June 18, 2023. This workshop provides a focused venue for discussing and disseminating research in the areas of fairness, bias, and trust in computer vision, as well as adjacent domains such as computational social science and public policy.
    •  
  • Related Video

  • Related Publication

  •  Cherian, A., Peng, K.-C., Lohit, S., Smith, K., Tenenbaum, J.B., "Are Deep Neural Networks SMARTer than Second Graders?", arXiv, December 2022.
    BibTeX arXiv
    • @inproceedings{Cherian2022dec,
    • author = {Cherian, Anoop and Peng, Kuan-Chuan and Lohit, Suhas and Smith, Kevin and Tenenbaum, Joshua B.},
    • title = {Are Deep Neural Networks SMARTer than Second Graders?},
    • booktitle = {arXiv},
    • year = 2022,
    • month = dec,
    • url = {https://arxiv.org/abs/2212.09993}
    • }