Low-rank Random Tensor for Bilinear Pooling

Yan Zhang, Krikamol Muandet, Qianli Ma, Heiko Neumann, Siyu Tang

Abstract

Fine-grained action segmentation in long untrimmed videos is an important task for many applications such as surveillance, robotics, and human-computer interaction. To understand subtle and precise actions within a long time period, second-order information (e.g. feature covariance) or higher is reported to be effective in the literature. However, extracting such high-order information is considerably non-trivial. In particular, the dimensionality increases exponentially with the information order, and hence gaining more representation power also increases the computational cost and the risk of overfitting. In this paper, we propose an approach to representing high-order information for temporal action segmentation via a simple yet effective bilinear form. Specifically, our contributions are: (1) From the multilinear perspective, we derive a bilinear form of low complexity, assuming that the three-way tensor has low-rank frontal slices. (2) Rather than learning the tensor entries from data, we sample the entries from different underlying distributions, and prove that the underlying distribution influences the information order. (3) We employed our bilinear form as an intermediate layer in state-of-the-art deep neural networks, enabling to represent high-order information in complex deep models effectively and efficiently. Our experimental results demonstrate that the proposed bilinear form outperforms the previous state-of-the-art methods on the challenging temporal action segmentation task.

Introduction

Fine-grained action segmentation involves identifying subtle, precise movements in long, untrimmed videos (e.g., "slicing a tomato" vs. "peeling a tomato"). To achieve high performance, models must look beyond simple features and examine high-order information, such as how different features co-occur or correlate over time. However, standard bilinear pooling, which calculates these interactions, suffers from the "curse of dimensionality." The number of parameters often grows cubically with the feature dimension, making it memory-intensive and slow to use as an intermediate layer in deep neural networks.

Frontal Low-rank Random Tensors

To address the computational tax of high-order feature fusion, we propose a novel Frontal Low-rank Random Tensor approach. This method relies on two key innovations:

  1. Low-Rank Assumption: We assume that the large three-way tensor used to calculate feature interactions can be decomposed into "frontal slices" that are low-rank. This allows for projecting input features into a lower-dimensional space before performing a vector outer product, drastically reducing complexity from $O(D^3)$ to roughly $O(D\sqrt{d})$.
  2. Random Projections (RP): Instead of learning the weights of these projection matrices, we sample them from specific distributions. RPBinary uses Rademacher distributions to approximate linear kernels, while RPGaussian uses Gaussian distributions to approximate RBF kernels, capturing "infinite-order" information. By using random weights, the model avoids overfitting and gains a solid theoretical foundation based on Reproducing Kernel Hilbert Space (RKHS) theory.
Architecture comparison: (a) standard TCN block and (b) the proposed block with the bilinear pooling module integrated as an intermediate layer.

Architecture Integration

Because our bilinear operation is lightweight, it can be used as an intermediate layer in deep models, whereas previous bilinear methods were often restricted to the final layer. We integrated this module into state-of-the-art architectures like the Temporal Convolutional Network (TCN) and the Multi-Stage TCN (MS-TCN). As shown in the figure above, the bilinear pooling module (comprising pooling, convolution, and dropout) follows the standard dilated convolution blocks, enabling the network to capture complex feature interactions throughout its depth.

Experimental Results

We evaluated the model on three standard benchmarks: 50Salads, GTEA, and Breakfast.

  • Performance: The proposed RPGaussian method consistently outperformed previous state-of-the-art (SOTA) models. On the GTEA dataset, it achieved an [email protected] score of 74.6, compared to the baseline MS-TCN's 70.4.
  • Efficiency: Our methods were significantly faster than other bilinear techniques like Temporal Bilinear Pooling (TBP), often by a factor of 8x or more, while maintaining superior performance.
  • Robustness: Incorporating a dropout layer within the bilinear module proved critical for preventing overfitting when dealing with high-dimensional feature interactions.

Conclusion

This work demonstrates that capturing high-order information for action segmentation does not require a massive computational overhead. By leveraging low-rank assumptions and random projections, we can effectively fuse features in a way that is both theoretically sound and practically efficient. Future work may explore using deep generative models to learn optimal distributions for these random tensors, further tailoring the kernels to specific video data types.

Summary updated

Abstract

Fine-grained action segmentation in long untrimmed videos is an important task for many applications such as surveillance, robotics, and human-computer interaction. To understand subtle and precise actions within a long time period, second-order information (e.g. feature covariance) or higher is reported to be effective in the literature. However, extracting such high-order information is considerably non-trivial. In particular, the dimensionality increases exponentially with the information order, and hence gaining more representation power also increases the computational cost and the risk of overfitting. In this paper, we propose an approach to representing high-order information for temporal action segmentation via a simple yet effective bilinear form. Specifically, our contributions are: (1) From the multilinear perspective, we derive a bilinear form of low complexity, assuming that the three-way tensor has low-rank frontal slices. (2) Rather than learning the tensor entries from data, we sample the entries from different underlying distributions, and prove that the underlying distribution influences the information order. (3) We employed our bilinear form as an intermediate layer in state-of-the-art deep neural networks, enabling to represent high-order information in complex deep models effectively and efficiently. Our experimental results demonstrate that the proposed bilinear form outperforms the previous state-of-the-art methods on the challenging temporal action segmentation task. One can see our project page for data, model and code: \url{https://vlg.inf.ethz.ch/projects/BilinearTCN/}.
BibTeX
@misc{zhang2019low,
  author = {Yan Zhang and Krikamol Muandet and Qianli Ma and Heiko Neumann and Siyu Tang},
  title = {{Low-rank Random Tensor for Bilinear Pooling}},
  year = {2019},
  eprint = {1906.01004},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1906.01004},
}
Figure