machine learning). Deep Learning models have beaten previous baselines significantly in many tasks in computer vision, natural language understanding, speech, and so on. Their rise can be attributed to a combination number-crunching at the heart of deep learning. AlexNet1 was one of the earliest models to rely on Graphics Processing Units (GPUs) for training, which could 1 Krizhevsky, Alex, Ilya Sutskever, and Geoffrey to create deeper networks with an ever larger number of parameters and increased complexity. In Computer Vision, several model architectures such as VGGNet, Inception, ResNet etc. (refer to Figure 1-2)

特征应该由多个共同学习的神经网络层组成，每个层都有可学习的参数。在机器视觉中，最底层可能检测边 88 https://en.wikipedia.org/wiki/Bag‐of‐words_model_in_computer_vision 248 7. 现代卷积神经网络 缘、颜色和纹理。事实上，Alex Krizhevsky、Ilya Sutskever和Geoff Hinton提出了一种新的卷积神经网络变 迭代都需要通过代价高昂的许多线性代 数层传递数据。这也是为什么在20世纪90年代至21世纪初，优化凸目标的简单算法是研究人员的首选。然而， 用GPU训练神经网络改变了这一格局。图形处理器（Graphics Processing Unit，GPU）早年用来加速图形处 理，使电脑游戏玩家受益。GPU可优化高吞吐量的4 × 4矩阵和向量乘法，从而服务于基本的图形任务。幸运 的是，这些数学运算与卷 Tuytelaars, T., & Van Gool, L. (2006). Surf: speeded up robust features. European conference on computer vision (pp. 404–417). [Bengio et al., 2003] Bengio, Y., Ducharme, R., Vincent, P., & Jauvin, C

这种算法固然简单直接，但是面对大规模、高维度数据的优化问题时计算效率极低， 基本不可行。梯度下降算法(Gradient Descent)是神经网络训练中最常用的优化算法，配合 强大的图形处理芯片 GPU(Graphics Processing Unit)的并行加速计算能力，非常适合优化海 量数据的神经网络模型，自然也适合优化这里的神经元线性模型。这里先简单地应用梯度 下降算法，来解决神经元模型预测的问题。由于梯度下降算法是深度学习的核心算法之 表示，图片中越白的像素点，对应矩阵位置中数值也就越大。 28行28列 图 3.3 图片的表示示意图① ① 素材来自 https://towardsdatascience.com/how-to-teach-a-computer-to-see-with-convolutional-neural-networks- 96c120827cd1 预览版202112 3.1 手写数字图片数据集 3 目前常用的深度学习框架，如 [5] M. D. Zeiler 和 R. Fergus, “Visualizing and Understanding Convolutional Networks,” 出处 Computer Vision -- ECCV 2014, Cham, 2014. [6] S. Ioffe 和 C. Szegedy, “Batch Normalization: Accelerating

������a��� �������a����� ����“��”���� ���������������������� ��GPU������������ ������ CVPR (Computer Vision and Pattern Recognition� 2015 ��������� ����FaceNet �FaceNet �������������������������� unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 815-823). ������ – �� ������ – �� ������ – ������ ������ Fang Wen, Jian Sun. Bayesian face revisited: a joint formulation. 2012, european conference on computer vision. MSRA “Feature Master” �� ��3���� � �21� �0�1 ����e�FD3F����A�L��]�gP[�����o��� �3�h���

reserved. Table Detection – How Do We Do It © 2018 Bloomberg Finance L.P. All rights reserved. Computer Vision Tasks Modified from https://commons.wikimedia.org/wiki/File:Cats_Petunia_and_Mimosa_2004 https://creativecommons.org/licenses/by-sa/4.0/deed.en © 2018 Bloomberg Finance L.P. All rights reserved. Computer Vision Tasks Modified from https://commons.wikimedia.org/wiki/File:Cats_Petunia_and_Mimosa_2004 https://creativecommons.org/licenses/by-sa/4.0/deed.en © 2018 Bloomberg Finance L.P. All rights reserved. Computer Vision Tasks Modified from https://commons.wikimedia.org/wiki/File:Cats_Petunia_and_Mimosa_2004

Mapillary • Computer vision for crowd sourced maps Hudl • Predictive analytics on sports plays Upserve • Restaurant table mgmt & POS for forecasting customer traffic TuSimple • Computer Vision for Autonomous Autonomous Driving Clarifai • Computer Vision APIs AWS 上的 AI 应用 • Pinterest Lens • Netflix 推荐引擎 数千名员工致力于人工智能领域 发现& 搜索 执行 &物流 现有产品的增强 定义新的产品分类 将机器学习拓 展更广领域 Amazon 的人工智能应用 在Amazon 最初的人 工智能应用 (1995)

parameters. Figure 7-1: The plethora of choices that we face when training a deep learning model in the computer vision domain. A Search Space for n parameters is a n-dimensional region such that a point in such transferable architectures for scalable image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018. blocks for the child networks. NASNet searches for the cells Platform-aware neural architecture search for mobile." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019. directly on the target devices and weighing the model accuracy

simply use e-books, Wikipedia and other sources for NLU related models, and web images & videos for computer vision models. We can then construct the final dataset for the pretext task by simply masking the to think about the significance of these results. It is kind of like receiving upgrades to your computer, except that these new upgrades don’t make it slower because your hardware is older, but actually 5555/3524938.3525536. 16 Szegedy, Christian, et al. "Rethinking the Inception Architecture for Computer Vision." arXiv, 2 Dec. 2015, doi:10.48550/arXiv.1512.00567. In curriculum learning, we start to

introduce Quantization, a popular compression technique which is also used in various fields of computer science in addition to deep learning. Quantization Before we jump to working with a deep learning Quantization Quantization is a common compression technique that has been used across different parts of Computer Science especially in signal processing. It is a process of converting high precision continuous "Xnor-net: Imagenet classification using binary convolutional neural networks." European conference on computer vision. Springer, Cham, 2016. Figure 2-10 compares the accuracies and the latencies between unoptimized

of the art NLP model architectures such as the Transformer, which is now showing great promise in computer vision applications as well! Learn Long-Term Dependencies Using Attention Imagine yourself in "Xception: Deep learning with depthwise separable convolutions." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017. on mobile and edge devices. Let’s say you want to design resize(image, [IMG_SIZE, IMG_SIZE]) 30 Parkhi, Omkar M., et al. "Cats and dogs." 2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012. mask = tf.image.resize(mask, [IMG_SIZE, IMG_SIZE])