Deepeyes: Progressive visual analysis system for depth-neural network design (deepeyes:progressive Visual analytics for designing deep neural Networks)

Deep neural Network, the problem of pattern recognition, has achieved very good results. But it is a time-consuming process to design a well-performing neural network that requires repeated attempts. This work [1] implements a visual analysis system for deep neural network design, Deepeyes. The system can extract data in Dnns training process, analyze the operation state of neural network from the angle of network whole effect, nerve layer and neuron, and help users update Dnns.

The Deepeyes system focuses on the following 5 types of tasks:

1. Identify the stability layer. In the Dnns training process, the user can choose the stable nerve layer, carries on the detailed analysis.
2. Identify the degenerate neurons. When a neuron reacts to all samples in the same way, it is considered that the neuron cannot provide additional information, has been degraded, and can be removed.
3. Identify features that are not detected. If some features are not detected, it indicates that the current network requires more neurons and nerve layers.
4. Recognize too much of the nerve layer. When there are many degenerate neurons in a certain nerve layer, it indicates that there are too many neuron in the nerve layer, which can be reduced.
5. Identify whether the number of neural layers in the neural network is appropriate.

Based on these tasks, they designed and implemented the visual Analysis System Deepeyes (Figure 1).

Figure 1 Deepeyes System architecture

The input data for the Deepeyes system is the instance of the sensing area (instances of receptive fields) and the excitation (activation) of the neurons. An instance of the sensing region is obtained by sampling with a sampling rate of 50%. The system mainly consists of 5 views.

Training view (Training overview), which displays in real time the error degree and accuracy (loss and accuracy) of Dnns training in a line chart form. This view, combined with the confusion histogram (perplexity histogram), helps the user to identify whether a nerve layer is in a stable state.

A degree of confusion histogram used to analyze the degree of stability of a certain nerve layer. Confusion, commonly used to measure the probability distribution or probability model of the predicted results and samples of the degree of fit. The lower the confusion, the better the degree of fit. The calculation of the confusion histogram is shown in Flow 2:

Figure 2 The construction process of the confusion histogram. (a) Sampled-area instances of the sensed region, (b) the excitation of the neurons in each area of the perceptual region, the color mapping of the excitation value, (c) the excitation of a series of neurons in the layer is transformed into a probability distribution; (d) The confusion of the neural layer's response to each instance is calculated; (e) discretization of continuous Build a histogram.

In the system, the degree of confusion histogram 3 shows that the red bar shows that the confusion is in the interval of the example is decreasing; green means its number is increasing; the height is positively correlated with the specific value.

Fig. 3 A sample of the confusion degree histogram analysis. Describes the stability of the first convolutional layer in a neural network under different iterations, and describes the stability of the second convolution layer of the neural network under different iteration times.

When a layer of nerve is in a stable state, the system allows for more detailed exploration by the user. As shown in 1, detail exploration consists of three views.

The excitation thermal map (activation heatmap) (Figure 4,5) helps the user discover the degenerate neurons. In the maximum excitation value thermodynamic diagram (Figure 4), if the color of the rectangular block is darker green, the maximum excitation value of the neuron is small. Indicates that the neuron has degenerated, or even died (dead filters), in feature recognition, can not provide extra information to be removed.

Fig. 4 The thermal force diagram of the maximum excitation value. Each rectangular block represents a neuron, and the color corresponds to the maximum excitation value of the neuron in a series of input instances.

In the frequency thermal diagram (Fig. 5), if the color of the rectangular block is darker blue, it indicates that the neuron is active on almost all input instances. It is indicated that the neuron cannot provide additional information in feature recognition and can be removed.

Figure 5 Frequency thermal diagram. Each rectangular block represents a neuron, and the color is correlated with the probability that the neuron succeeds in motivating the instance.

Input map (Figure 6), using the excitation of the neuron to the instance, describe each instance, using the method of dimensionality reduction, on the two-dimensional plane, the relationship between the excitation vectors describing the example is analyzed. In Figure 6, if the two nodes are similar, the excitation vectors of the two instances are similar, it is further speculated that in the neural layer, the neurons think that the two instances are similar.

Figure 6, enter the diagram. Each node represents an instance of the perceptual region, and the color represents the label for that instance.

The Neuron view (filter map) (Figure 7) shows how similar the neuron is to the excitation instance.

Figure 7 Neuron view. Each node represents a neuron; the color represents the label of an instance of the neuron's main excitation, and the size represents the greater the degree of tightness of the neuron with the label, and the closer the representation.

In the Deepeyes system, when the user selects a neuron in a thermal or neuron view, the input graph shows an example of the neuron's excitation (Figure 8).

Fig. 8 The left image, an example of a neuron excitation, and a degenerate neuron that can hardly inspire an instance in the right image.

Next, I'll use an example Mnist network to introduce Deepeyes. In this example, the data is 70,000 handwritten Arabic numerals, 10,000 of which are tagged data. The training network consists of two convolutional layers, each with 20 and 50 neurons, and two fully connected layers, with 500 and 10 neurons in each layer.

The analysis of the Mnist network is shown in 9. The input graph of the first convolutional layer (Figure 9 (a)) is very confusing and there is no obvious clustering. It is shown that the layer can only recognize the basic characteristics and cannot classify the input images; from the maximum excitation value, we can find some degenerate neurons. Indicates that the number of neurons in this layer can be reduced a little. Observe the input graph of the second convolution layer (Figure 9 (b)). The input graph has better clustering characteristics than Figure 9 (a), but it is still not obvious enough. But in the first fully connected layer (Figure 9 (c)), we find that the input graph has very obvious clustering characteristics. Within each category, the color of the nodes is consistent, indicating that the layer has strong classification and recognition ability. But there are some fuzzy areas between the categories, and looking at the examples, we find that this is some of the more difficult numbers to identify, such as 7 and 9. But in the second fully connected graph of the input graph (Figure 9 (d)), we found that the distance between the cluster is relatively large, indicating that the level of recognition and classification is good. At the same time, in the neuron view, we find that each neuron recognizes a different label and has a greater distance from each other.

With this analysis, the Minst network can be found to reduce the number of neurons in the first convolutional layer and the first fully connected layer. Finally, they reduced the number of first convolutional neurons from 20 to 10, reducing the number of neurons in the first fully-connected layer from 500 to 100. After 2000 iterations, the network accuracy rate reached 98.2%.

Figure 9 Mnist network analysis diagram. From left to right, the first convolution layer, the second convolutional layer, the first fully connected layer, and the second fully connected layer.

In general, Deepeyes is a visual analysis system for designing deep neural networks. The system allows users to start with a stable neural layer, analyze the similarity between the excitation vectors obtained by the input instances and the similarity of the neuron excitation behavior, and help the user to analyze the neural network condition, such as whether the number of nerve layers is not enough and whether the number of neurons in the nerve layer is too many.

Deepeyes: Progressive visual analysis system for depth-neural network design (deepeyes:progressive Visual analytics for designing deep neural Networks)