Abstract.
Large Convolutional Network models have recently demonstrated impressive classification performance on the ImageNet benchmark Krizhevsky et al. [18]. However there is no clear understanding of why they perform so well, or how they might be improved. In this paper we explore both issues. We introduce a novel visualization technique that gives insight into the function of intermediate feature layers and the operation of the classifier. Used in a diagnostic role, these visualizations allow us to find model architectures that outperform Krizhevsky et al. on the ImageNet classification benchmark. We also perform an ablation study to discover the performance contribution from different model layers. We show our ImageNet model generalizes well to other datasets: when the softmax classifier is retrained, it convincingly beats the current state-ofthe-art results on Caltech-101 and Caltech-256 datasets.
Since their introduction by LeCun et al. [20] in the early 1990’s, Convolutional Networks (convnets) have demonstrated excellent performance at tasks such as hand-written digit classification and face detection. In the last 18 months, several papers have shown that they can also deliver outstanding performance on more challenging visual classification tasks. Ciresan et al. [4] demonstrate state-ofthe-art performance on NORB and CIFAR-10 datasets. Most notably, Krizhevsky et al. [18] show record beating performance on the ImageNet 2012 classification benchmark, with their convnet model achieving an error rate of 16.4%, compared to the 2nd place result of 26.1%. Following on from this work, Girshick et al. [10] have shown leading detection performance on the PASCAL VOC dataset. Several factors are responsible for this dramatic improvement in performance: (i) the availability of much larger training sets, with millions of labeled examples; (ii) powerful GPU implementations, making the training of very large models practical and (iii) better model regularization strategies, such as Dropout [14].
Despite this encouraging progress, there is still little insight into the internal operation and behavior of these complex models, or how they achieve such good performance. From a scientific standpoint, this is deeply unsatisfactory. Without clear understanding of how and why they work, the development of better models is reduced to trial-and-error. In this paper we introduce a visualization technique that reveals the input stimuli that excite individual feature maps at any layer in the model. It also allows us to observe the evolution of features during training and to diagnose potential problems with the model. The visualization technique we propose uses a multi-layered Deconvolutional Network (deconvnet), as proposed by Zeiler et al. [29], to project the feature activations back to the input pixel space. We also perform a sensitivity analysis of the classifier output by occluding portions of the input image, revealing which parts of the scene are important for classification.
Using these tools, we start with the architecture of Krizhevsky et al. [18] and explore different architectures, discovering ones that outperform their results on ImageNet. We then explore the generalization ability of the model to other datasets, just retraining the softmax classifier on top. As such, this is a form of supervised pre-training, which contrasts with the unsupervised pre-training methods popularized by Hinton et al. [13] and others [1,26].
1.1 Related Work
Visualization: Visualizing features to gain intuition about the network is common practice, but mostly limited to the 1st layer where projections to pixel space are possible. In higher layers alternate methods must be used. [8] find the optimal stimulus for each unit by performing gradient descent in image space to maximize the unit’s activation. This requires a careful initialization and does not give any information about the unit’s invariances. Motivated by the latter’s short-coming, [19] (extending an idea by [2]) show how the Hessian of a given unit may be computed numerically around the optimal response, giving some insight into invariances. The problem is that for higher layers, the invariances are extremely complex so are poorly captured by a simple quadratic approximation. Our approach, by contrast, provides a non-parametric view of invariance, showing which patterns from the training set activate the feature map. Our approach is similar to contemporary work by Simonyan et al. [23] who demonstrate how saliency maps can be obtained from a convnet by projecting back from the fully connected layers of the network, instead of the convolutional features that we use. Girshick et al. [10] show visualizations that identify patches within a dataset that are responsible for strong activations at higher layers in the model. Our visualizations differ in that they are not just crops of input images, but rather top-down projections that reveal structures within each patch that stimulate a particular feature map.
Feature Generalization: Our demonstration of the generalization ability of convnet features is also explored in concurrent work by Donahue et al. [7] and Girshick et al. [10]. They use the convnet features to obtain state-of-the-art performance on Caltech-101 and the Sun scenes dataset in the former case, and for object detection on the PASCAL VOC dataset, in the latter.
2 Approach
We use standard fully supervised convnet models throughout the paper, as defined by LeCun et al. [20] and Krizhevsky et al. [18]. These models map a color 2D input image x , via a series of layers, to a probability vector yˆ over the C different classes. Each layer consists of (i) convolution of the previous layer output (or, in the case of the 1st layer, the input image) with a set of learned filters; (ii) passing the responses through a rectified linear function (relu(x) = max(x, 0)); (iii) [optionally] max pooling over local neighborhoods and (iv) [optionally] a local contrast operation that normalizes the responses across feature maps. For more details of these operations, see [18] and [16]. The top few layers of the network are conventional fully-connected networks and the final layer is a softmax classifier. Fig. 3 shows the model used in many of our experiments.
We train these models using a large set of N labeled images {x, y}, where label yi is a discrete variable indicating the true class. A cross-entropy loss function,suitable for image classification, is used to compare yˆ and y . The parameters ii of the network (filters in the convolutional layers, weight matrices in the fullyconnected layers and biases) are trained by back-propagating the derivative of the loss with respect to the parameters throughout the network, and updating the parameters via stochastic gradient descent. Details of training are given in Section 3.
2.1 Visualization with a Deconvnet
Understanding the operation of a convnet requires interpreting the feature activity in intermediate layers. We present a novel way to map these activities back to the input pixel space, showing what input pattern originally caused a given activation in the feature maps. We perform this mapping with a Deconvolutional Network (deconvnet) Zeiler et al. [29]. A deconvnet can be thought of as a convnet model that uses the same components (filtering, pooling) but in reverse, so instead of mapping pixels to features does the opposite. In Zeiler et al. [29], deconvnets were proposed as a way of performing unsupervised learning. Here, they are not used in any learning capacity, just as a probe of an already trained convnet.
Fig. 1. Top: A deconvnet layer (left) attached to a convnet layer (right). The deconvnet will reconstruct an approximate version of the convnet features from the layer beneath. Bottom: An illustration of the unpooling operation in the deconvnet, using switches which record the location of the local max in each pooling region (colored zones) during pooling in the convnet. The black/white bars are negative/positive activations within the feature map.
To examine a convnet, a deconvnet is attached to each of its layers, as illustrated in Fig. 1(top), providing a continuous path back to image pixels. To start, an input image is presented to the convnet and features computed throughout the layers. To examine a given convnet activation, we set all other activations in the layer to zero and pass the feature maps as input to the attached deconvnet layer. Then we successively (i) unpool, (ii) rectify and (iii) filter to reconstruct the activity in the layer beneath that gave rise to the chosen activation. This is then repeated until input pixel space is reached.
Unpooling: In the convnet, the max pooling operation is non-invertible, however we can obtain an approximate inverse by recording the locations of the maxima within each pooling region in a set of switch variables. In the deconvnet, the unpooling operation uses these switches to place the reconstructions from the layer above into appropriate locations, preserving the structure of the stimulus. See Fig. 1(bottom) for an illustration of the procedure.
Rectification: The convnet uses relu non-linearities, which rectify the feature maps thus ensuring the feature maps are always positive. To obtain valid feature reconstructions at each layer (which also should be positive), we pass the reconstructed signal through a relu non-linearity1.
Filtering: The convnet uses learned filters to convolve the feature maps from the previous layer. To approximately invert this, the deconvnet uses transposed versions of the same filters (as other autoencoder models, such as RBMs), but applied to the rectified maps, not the output of the layer beneath. In practice this means flipping each filter vertically and horizontally.
Note that we do not use any contrast normalization operations when in this reconstruction path. Projecting down from higher layers uses the switch settings generated by the max pooling in the convnet on the way up. As these switch settings are peculiar to a given input image, the reconstruction obtained from a single activation thus resembles a small piece of the original input image, with structures weighted according to their contribution toward to the feature activation. Since the model is trained discriminatively, they implicitly show which parts of the input image are discriminative. Note that these projections are not samples from the model, since there is no generative process involved. The whole procedure is similar to backpropping a single strong activation (rather than the usual gradients), i.e. computing ∂h , where h is the element of the feature map ∂Xn with the strong activation and Xn is the input image. However, it differs in that (i) the the relu is imposed independently and (ii) contrast normalization operations are not used. A general shortcoming of our approach is that it only visualizes a single activation, not the joint activity present in a layer. Nevertheless, as we show in Fig. 6, these visualizations are accurate representations of the input pattern that stimulates the given feature map in the model: when the parts of the original input image corresponding to the pattern are occluded, we see a distinct drop in activity within the feature map.
3 Training Details
Fig. 3. Architecture of our 8 layer convnet model. A 224 by 224 crop of an image (with 3 color planes) is presented as the input. This is convolved with 96 different 1st layer filters (red), each of size 7 by 7, using a stride of 2 in both x and y. The resulting feature maps are then: (i) passed through a rectified linear function (not shown), (ii) pooled (max within 3x3 regions, using stride 2) and (iii) contrast normalized across feature maps to give 96 different 55 by 55 element feature maps. Similar operations are repeated in layers 2,3,4,5. The last two layers are fully connected, taking features from the top convolutional layer as input in vector form (6 · 6 · 256 = 9216 dimensions). The final layer is a C-way softmax function, C being the number of classes. All filters and feature maps are square in shape.
We now describe the large convnet model that will be visualized in Section 4. The architecture, shown in Fig. 3, is similar to that used by Krizhevsky et al. [18] for ImageNet classification. One difference is that the sparse connections used in Krizhevsky’s layers 3,4,5 (due to the model being split across 2 GPUs) are replaced with dense connections in our model. Other important differences relating to layers 1 and 2 were made following inspection of the visualizations in Fig. 5, as described in Section 4.1.
The model was trained on the ImageNet 2012 training set (1.3 million images, spread over 1000 different classes) [6]. Each RGB image was preprocessed by resizing the smallest dimension to 256, cropping the center 256x256 region, subtracting the per-pixel mean (across all images) and then using 10 different sub-crops of size 224x224 (corners + center with(out) horizontal flips). Stochastic gradient descent with a mini-batch size of 128 was used to update the parameters, starting with a learning rate of 10−2, in conjunction with a momentum term of 0.9. We anneal the learning rate throughout training manually when the validation error plateaus. Dropout [14] is used in the fully connected layers (6 and 7) with a rate of 0.5. All weights are initialized to 10−2 and biases are set to 0.
Visualization of the first layer filters during training reveals that a few of them dominate. To combat this, we renormalize each filter in the convolutional layers whose RMS value exceeds a fixed radius of 10−1 to this fixed radius. This is crucial, especially in the first layer of the model, where the input images are roughly in the [-128,128] range. As in Krizhevsky et al. [18], we produce multiple different crops and flips of each training example to boost training set size. We stopped training after 70 epochs, which took around 12 days on a single GTX580 GPU, using an implementation based on [18].
4 Convnet Visualization
Using the model described in Section 3, we now use the deconvnet to visualize the feature activations on the ImageNet validation set.
Fig. 2. Visualization of features in a fully trained model. For layers 2-5 we show the top 9 activations in a random subset of feature maps across the validation data, projected down to pixel space using our deconvolutional network approach. Our reconstructions are not samples from the model: they are reconstructed patterns from the validation set that cause high activations in a given feature map. For each feature map we also show the corresponding image patches. Note: (i) the the strong grouping within each feature map, (ii) greater invariance at higher layers and (iii) exaggeration of discriminative parts of the image, e.g. eyes and noses of dogs (layer 4, row 1, cols 1). Best viewed in electronic form. The compression artifacts are a consequence of the 30Mb submission limit, not the reconstruction algorithm itself.
Feature Visualization: Fig. 2 shows feature visualizations from our model once training is complete. For a given feature map, we show the top 9 activations, each projected separately down to pixel space, revealing the different structures that excite that map and showing its invariance to input deformations. Alongside these visualizations we show the corresponding image patches. These have greater variation than visualizations which solely focus on the discriminant structure within each patch. For example, in layer 5, row 1, col 2, the patches appear to have little in common, but the visualizations reveal that this particular feature map focuses on the grass in the background, not the foreground objects.
The projections from each layer show the hierarchical nature of the features in the network. Layer 2 responds to corners and other edge/color conjunctions. Layer 3 has more complex invariances, capturing similar textures (e.g. mesh patterns (Row 1, Col 1); text (R2,C4)). Layer 4 shows significant variation, and is more class-specific: dog faces (R1,C1); bird’s legs (R4,C2). Layer 5 shows entire objects with significant pose variation, e.g. keyboards (R1,C11) and dogs (R4).
Feature Evolution during Training: Fig. 4 visualizes the progression during training of the strongest activation (across all training examples) within a given feature map projected back to pixel space. Sudden jumps in appearance result from a change in the image from which the strongest activation originates. The lower layers of the model can be seen to converge within a few epochs. However, the upper layers only develop develop after a considerable number of epochs (40-50), demonstrating the need to let the models train until fully converged.
4.1 Architecture Selection
While visualization of a trained model gives insight into its operation, it can also assist with selecting good architectures in the first place. By visualizing the first and second layers of Krizhevsky et al. ’s architecture (Fig. 5(a) & (c)), various problems are apparent. The first layer filters are a mix of extremely high and low frequency information, with little coverage of the mid frequencies. Additionally, the 2nd layer visualization shows aliasing artifacts caused by the large stride 4 used in the 1st layer convolutions. To remedy these problems, we (i) reduced the 1st layer filter size from 11x11 to 7x7 and (ii) made the stride of the convolution 2, rather than 4. This new architecture retains much more information in the 1st and 2nd layer features, as shown in Fig. 5(b) & (d). More importantly, it also improves the classification performance as shown in Section 5.1.
4.2 Occlusion Sensitivity
With image classification approaches, a natural question is if the model is truly identifying the location of the object in the image, or just using the surrounding context. Fig. 6 attempts to answer this question by systematically occluding different portions of the input image with a grey square, and monitoring the output of the classifier. The examples clearly show the model is localizing the objects within the scene, as the probability of the correct class drops significantly when the object is occluded. Fig. 6 also shows visualizations from the strongest feature map of the top convolution layer, in addition to activity in this map (summed over spatial locations) as a function of occluder position.
When the occluder covers the image region that appears in the visualization, we see a strong drop in activity in the feature map. This shows that the visualization genuinely corresponds to the image structure that stimulates that feature map, hence validating the other visualizations shown in Fig. 4 and Fig. 2.
This supports the premise that as the feature hierarchies become deeper, they learn increasingly powerful features.
We explored large convolutional neural network models, trained for image classification, in a number ways. First, we presented a novel way to visualize the activity within the model. This reveals the features to be far from random, uninterpretable patterns. Rather, they show many intuitively desirable properties such as compositionality, increasing invariance and class discrimination as we ascend the layers. We also show how these visualization can be used to identify problems with the model and so obtain better results, for example improving on Krizhevsky et al. ’s [18] impressive ImageNet 2012 result. We then demonstrated through a series of occlusion experiments that the model, while trained for classification, is highly sensitive to local structure in the image and is not just using broad scene context. An ablation study on the model revealed that having a minimum depth to the network, rather than any individual section, is vital to the model’s performance.
Finally, we showed how the ImageNet trained model can generalize well to other datasets. For Caltech-101 and Caltech-256, the datasets are similar enough that we can beat the best reported results, in the latter case by a significant margin. Our convnet model generalized less well to the PASCAL data, perhaps suffering from dataset bias [25], although it was still within 3.2% of the best reported result, despite no tuning for the task. For example, our performance might improve if a different loss function was used that permitted multiple objects per image. This would naturally enable the networks to tackle the object detection as well.
References
1. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. In: NIPS, pp. 153–160 (2007)
2. Berkes, P., Wiskott, L.: On the analysis and interpretation of inhomogeneous quadratic forms as receptive fields. Neural Computation (2006)
3. Bo, L., Ren, X., Fox, D.: Multipath sparse coding using hierarchical matching pursuit. In: CVPR (2013)
4. Ciresan, D.C., Meier, J., Schmidhuber, J.: Multi-column deep neural networks for image classification. In: CVPR (2012)
5. Dalal, N., Triggs, B.: Histograms of oriented gradients for pedestrian detection. In: CVPR (2005)
6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-Scale Hierarchical Image Database. In: CVPR 2009 (2009)
7. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrell, T.: DeCAF: A deep convolutional activation feature for generic visual recognition. arXiv:1310.1531 (2013)
8. Erhan, D., Bengio, Y., Courville, A., Vincent, P.: Visualizing higher-layer features of a deep network. Technical report, University of Montreal (2009)
9. Fei-fei, L., Fergus, R., Perona, P.: One-shot learning of object categories. IEEE Trans. PAMI (2006)
10. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. arXiv:1311.2524 (2014)
11. Griffin, G., Holub, A., Perona, P.: The caltech 256. Caltech Technical Report (2006)
12. Gunji, N., Higuchi, T., Yasumoto, K., Muraoka, H., Ushiku, Y., Harada, T., Ku-niyoshi, Y.: Classification entry. Imagenet Competition (2012)
13. Hinton, G.E., Osindero, S., Teh, Y.: A fast learning algorithm for deep belief nets.Neural Computation 18, 1527–1554 (2006)
14. Hinton, G.E., Srivastave, N., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.R.:Improving neural networks by preventing co-adaptation of feature detectors. In:arXiv:1207.0580 (2012)
15. Howard, A.G.: Some improvements on deep convolutional neural network based image classification. arXiv 1312.5402 (2013)
16. Jarrett, K., Kavukcuoglu, K., Ranzato, M., LeCun, Y.: What is the best multi-stage architecture for object recognition? In: ICCV (2009)
17. Jianchao, Y., Kai, Y., Yihong, G., Thomas, H.: Linear spatial pyramid matching using sparse coding for image classification. In: CVPR (2009)
18. Krizhevsky, A., Sutskever, I., Hinton, G.: Imagenet classification with deep convolutional neural networks. In: NIPS (2012)
19. Le, Q.V., Ngiam, J., Chen, Z., Chia, D., Koh, P., Ng, A.Y.: Tiled convolutional neural networks. In: NIPS (2010)
20. LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W., Jackel, L.D.: Backpropagation applied to handwritten zip code recognition. Neural Comput. 1(4), 541–551 (1989)
21. Oquab, M., Bottou, L., Laptev, I., Sivic, J.: Learning and transferring mid-level image representations using convolutional neural networks. In: CVPR (2014)
22. Sande, K., Uijlings, J., Snoek, C., Smeulders, A.: Hybrid coding for selective search.In: PASCAL VOC Classification Challenge 2012 (2012)
23. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Vi-sualising image classification models and saliency maps. arXiv 1312.6034v1 (2013)
24. Sohn, K., Jung, D., Lee, H., Hero III, A.: Efficient learning of sparse, distributed,convolutional feature representations for object recognition. In: ICCV (2011)
25. Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: CVPR (2011)
26. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: ICML, pp. 1096–1103 (2008)
27. Yan, S., Dong, J., Chen, Q., Song, Z., Pan, Y., Xia, W., Huang, Z., Hua, Y., Shen, S.: Generalized hierarchical matching for sub-category aware object classification.In: PASCAL VOC Classification Challenge 2012 (2012)
28.Zeiler,M.:Clarifai(2013),http://www.imagenet.org/challenges/LSVRC/2013/results.php
29. Zeiler, M., Taylor, G., Fergus, R.: Adaptive deconvolutional networks for mid and high level feature learning. In: ICCV (2011)
177
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
