quality and model footprint. An in-depth discussion of data augmentation and distillation follows right after. Following the lead from the previous chapters, the theory is complemented with programming Let’s take brightness transformation as an example. Figure 3-6 shows an image 2x bright (bottom-right corner) as compared to the original image (center). This transformation causes the whale fin to visually doubles the pixel intensity values leading to a brighter image. A 2X brightness transformation (bottom-right corner in figure 3-6) of our whale image flattens the water texture causing the pectoral fin to stand

to research and experiment with these and other similar ideas. With that being said, let’s jump right in. Self-Supervised Learning The vanilla supervised learning paradigm that we are familiar has two part. You can also play around with the arrangement of the input, and make the model predict the right order of the elements of . The next question is where do we get the data for creating these tasks but in other cases you might have to experiment and figure out the learning techniques and their right combinations that work well for your model training setup. With that being said, let’s jump to how

set of tools and techniques that can help us make models pareto-optimal and let the user pick the right tradeoff. To that end, we can think of work on efficient deep learning to be categorized in roughly Figure 1-9: Illustration of the pruning process. On the left is the unpruned graph, and on the right is a pruned graph with the unimportant connections and neurons removed. Learning Techniques Learning data. The intuition is that the soft labels from the teacher can help the student capture how wrong/right the prediction is. For example, given an image of a truck if the student model predicts that it is

Python模块-Pandas ⚫ 数据合并 pd.merge(left, right) 类数 据库的数据融合操作. 参数：how，融合方式，包括左连接、右连接、内连 接（默认）和外连接；on，连接键；left_on，左 键；right_on，右键；left_index，是否将left 行索引作 为左键；right_index，是否将right行 索引作为右键. 66 Python模块-Pandas ⚫数据融合

‘cute’ occupies the x-axis, and the feature ‘dangerous’ occupies the y-axis. The animals on the bottom-right are cute and safe to play with. The dangerous animals occupy the top-left area of the plot. Note validation_data=(x_test_vectorized, y_test)) The model training without pre-trained embeddings follows right after: # Training without pre-trained embeddings. cnn_model_no_w2v = get_cnn_model(get_untrained_embedding_layer()) left shows a recurrent cell processing the input sequence element at time step t. The image on the right explains the processing of the entire input sequence across n time steps. RNNs are also used for

Python模块-Pandas ⚫ 数据合并 pd.merge(left, right) 类数 据库的数据融合操作. 参数：how，融合方式，包括左连接、右连接、内连 接（默认）和外连接；on，连接键；left_on，左 键；right_on，右键；left_index，是否将left 行索引作 为左键；right_index，是否将right行 索引作为右键. 67 Python模块-Pandas ⚫数据融合

on the left is smaller which was used to classify the cifar10 dataset. The larger network on the right is designed to learn from the ImageNet dataset. Both of these networks are largely composed of alternate predicted parameters is . In the original NASNet paper, the value for is chosen to be 5. Figure 7-8 (right) shows a predicted block. Figure 7-8: The structure of a block used to compose normal and reduction the timesteps predicting the hidden states, primitive operations and the combinations operations. Right image shows the structure of a block after applying the predictions from NASNet. NASNet selects the

to the idea of compression. Details of quantization and its applications in deep learning follow right after. The quantization section delves into the implementation details using code samples. We finish about the loss in quality. Figure 2-2: On the left is a high quality image of a cat. The cat on the right is a lower quality compressed image. Source Both the cat images in figure 2-2 might serve their uses 2-bits to represent each value. The quality degradation is apparent. The image in the bottom right corner (b=8) is the highest quality image. It’s hard to notice much difference for the b values 6

2023 39 / 122 Multivariate Gaussian Distribution: A 2D Example From left to right: Σ = I, Σ = 0.6I, Σ = 2I From left to right: Σ = I, Σ = � 1 0.5 0.5 1 � , Σ = � 1 0.8 0.8 1 � Feng Li (SDU) GDA, NB and EM September 27, 2023 40 / 122 Multivariate Gaussian Distribution: A 2D Example From left to right: µ = � 1 0 � , µ = � −0.5 0 � , µ = � −1 −1.5 � Feng Li (SDU) GDA, NB and EM September 27, 2023 Q[t] i (z(i)) where Q[t] i (z(i)) = p(z(i) | x(i); θ[t]) θ[t+1] is then obtained by maximizing the right hand side of the above equation Feng Li (SDU) GDA, NB and EM September 27, 2023 100 / 122 Convergence

(symmetric_height_crop, symmetric_width_crop)。 • 如果为 2 个整数的 2 个元组：解释为 ((top_crop, bottom_crop), (left_crop, right_crop))。 • data_format: 字符串，channels_last (默认) 或 channels_first 之一，表示输入中 维度的顺序。channels_last 对应输入尺寸为 symmetric_dim3_crop)。 • 如 果 为 2 个 整 数 的 3 个 元 组： 解 释 为 ((left_dim1_crop, right_dim1_crop), (left_dim2_crop, right_dim2_crop), (left_dim3_crop, right_dim3_crop))。 • data_format: 字符串，channels_last (默认) 或 channels_first • 如果为整数：在填充维度（第一个轴）的开始和结束处添加多少个零。 关于 KERAS 网络层 77 • 长度为 2 的整数元组：在填充维度的开始和结尾处添加多少个零 ((left_pad, right_pad))。 输入尺寸 3D 张量，尺寸为 (batch, axis_to_pad, features)。 输出尺寸 3D 张量，尺寸为 (batch, padded_axis, features)。