Article Source

• Title: Deep Learning Part 2: Transfer Learning and Fine-tuning Deep Convolutional Neural Networks
• Authors: Anusua Trivedi, Microsoft Data Scientist

Transfer Learning and Fine-tuning Deep Convolutional Neural Networks

This is a blog series in several parts — where I describe my experiences and go deep into the reasons behind my choices. In Part 1, I discussed the pros and cons of different symbolic frameworks, and my reasons for choosing Theano (with Lasagne) as my platform of choice.

Part 2 of this blog series is based on my upcoming talk at The Data Science Conference, 2016. Here in Part 2, I describe Deep Convolutional Neural Networks (DCNNs) and how Transfer learning and Fine-tuning helps better the training process for domain specific images.

Please feel free to email me at trivedianusua23@gmail.com if you have questions.

Introduction

The eye disease Diabetic Retinopathy (DR) is a common cause of vision loss. Screening diabetic patients using fluorescein angiography images can potentially reduce the risk of blindness. Current trends in the research have demonstrated that DCNNs are very effective in automatically analyzing large collections of images and identifying features that can categorize images with minimum error. DCNNs are rarely trained from scratch, as it is relatively uncommon to have a domain-specific dataset of sufficient size. Since modern DCNNs take 2-3 weeks to train across GPUs, Berkley Vision and Learning Center (BVLC) have released some final DCNN checkpoints. In this blog, we use such a pre-trained network: GoogLeNet. This GoogLeNet network is pre-trained on a large collection of natural ImageNet images. We transfer the learned ImageNet weights as initial weights for the network, and fine-tune these pre-trained generic network to recognize fluorescein angiography images of eyes and improve DR prediction.

Using explicit feature extraction to predict Diabetic Retinopathy

Much work has been done in developing algorithms and morphological image processing techniques that explicitly extract features prevalent in patients with DR. The generic workflow used in a standard image classification technique is as follows:

• Image preprocessing techniques for noise removal and contrast enhancement
• Feature extraction technique
• Classification
• Prediction

Faust et al. provide a very comprehensive analysis of models that use explicit feature extraction for DR screening. Vujosevic et al. build a binary classifier on a dataset of 55 patients by explicitly forming single lesion features. These authors use morphological image processing techniques to extract blood vessel, and hemorrhage features and then train an SVM on a data set of 331 images. These authors report accuracy of 90% and sensitivity of 90% on binary classification task with a dataset of 140 images.

However, all these processes are very time and effort consuming. Further improvements in prediction accuracy require large quantities of labeled data. Image processing and feature extraction of image datasets is very complex and time-consuming. Thus, we choose to automate the image processing and feature extraction step by using DCNNs.

Deep convolutional neural network (DCNN)

Image data requires subject-matter expertise to extract key features. DCNNs extract features automatically from domain-specific images, without any feature engineering techniques. This process makes DCNNs suitable for image analysis:

• DCNNs train networks with many layers
• Multiple layers work to build an improved feature space
• Initial layers learn 1st order features (e.g. color, edges etc.)
• Later layers learn higher order features (specific to input dataset)
• Lastly, final layer features are fed into classification layer(s)


 C layers are convolutions, S layers are pool/sample

Convolution: Convolution layers consist of a rectangular grid of neurons. The weights for this are the same for each neuron in the convolution layer. The convolution layer weights specify the convolution filter.

Convolution

Pooling: The pooling layer takes small rectangular blocks from the convolutional layer and subsamples it to produce a single output from that block.

Pooling  

In this post, we are using GoogLeNet DCNN, which was developed at Google. GoogLeNet won the ImageNet challenge in 2014, setting the record for the best contemporaneous results. Motivations for this model were a simultaneously deeper as well as computationally inexpensive architecture.

GoogLeNet

Transfer Learning and Fine-tuning DCNNs

In practice, we don’t usually train an entire DCNN from scratch with random initialization. This is because it is relatively rare to have a dataset of sufficient size that is required for the depth of network required. Instead, it is common to pre-train a DCNN on a very large dataset and then use the trained DCNN weights either as an initialization or a fixed feature extractor for the task of interest.

Fine-Tuning: Transfer learning strategies depend on various factors, but the two most important ones are the size of the new dataset, and its similarity to the original dataset. Keeping in mind that DCNN features are more generic in early layers and more dataset-specific in later layers, there are four major scenarios:

1. New dataset is smaller in size and similar in content compared to original dataset: If the data is small, it is not a good idea to fine-tune the DCNN due to overfitting concerns. Since the data is similar to the original data, we expect higher-level features in the DCNN to be relevant to this dataset as well. Hence, the best idea might be to train a linear classifier on the CNN-features.

2. New dataset is relatively large in size and similar in content compared to the original dataset: Since we have more data, we can have more confidence that we would not over fit if we were to try to fine-tune through the full network.

3. New dataset is smaller in size but very different in content compared to the original dataset: Since the data is small, it is likely best to only train a linear classifier. Since the dataset is very different, it might not be best to train the classifier from the top of the network, which contains more dataset-specific features. Instead, it might work better to train a classifier from activations somewhere earlier in the network.

4. New dataset is relatively large in size and very different in content compared to the original dataset: Since the dataset is very large, we may expect that we can afford to train a DCNN from scratch. However, in practice it is very often still beneficial to initialize with weights from a pre-trained model. In this case, we would have enough data and confidence to fine-tune through the entire network.

Fine-tuning DCNNs: For this DR prediction problem, we fall under scenario iv. We fine-tune the weights of the pre-trained DCNN by continuing the backpropagation. It is possible to fine-tune all the layers of the DCNN, or it’s possible to keep some of the earlier layers fixed (due to overfitting concerns) and only fine-tune some higher-level portion of the network. This is motivated by the observation that the earlier features of a DCNN contain more generic features (e.g. edge detectors or color blob detectors) that should be useful to many tasks, but later layers of the DCNN becomes progressively more specific to the details of the classes contained in the DR dataset.

Transfer learning constraints: As we use a pre-trained network, we are slightly constrained in terms of the model architecture. For example, we can’t arbitrarily take out convolutional layers from the pre-trained network. However, due to parameter sharing, we can easily run a pre-trained network on images of different spatial size. This is clearly evident in the case of Convolutional and Pool layers because their forward function is independent of the input volume spatial size. In case of Fully Connected (FC) layers, this still holds true because FC layers can be converted to a Convolutional Layer.

Learning rates: We use a smaller learning rate for DCNN weights that are being fine-tuned under the assumption that the pre-trained DCNN weights are relatively good. We don’t wish to distort them too quickly or too much, so we keep both our learning rate and learning rate decay really small.

Data Augmentation: One of the drawbacks of non-regularized neural networks is that they are extremely flexible: they learn both features and noise equally well, increasing the potential for overfitting. In our model, we apply L2 regularization to avoid overfitting. But even after that, we observed a large gap in model performance on the training and validation DR images, indicating that the fine tuning process is overfitting to the training set. To combat this overfitting, we leverage data augmentation for the DR image dataset.

There are many ways to do data augmentation, such as the popular horizontally flipping, random crops and color jittering. As the color information in these images is very important, we only rotate the images at different angles – at 0, 90, 180, and 270 degrees.

Replacing the input layer of the pre-trained GoogLeNet network with DR images. We fine-tune all layers except for the top 2 pre-trained layers which contains more generic data-independent weights.

Fine-tuning GoogLeNet: The GoogLeNet network we use here for DR screening was initially trained on ImageNet. The ImageNet dataset contains about 1 million natural images and 1000 labels/categories. In contrast, our labeled DR dataset has only about 30,000 domain-specific images and 4 labels/ categories. Thus, the DR dataset is insufficient to train a network as complex as GoogLeNet and so we use weights from the ImageNet-trained GoogLeNet network. We fine-tune all layers, except for the top 2 pre-trained layers which contains more generic data-independent weights. The original classification layer “loss3/classifier” outputs predictions for 1000 classes. We replace it with a new binary layer.

Fine-tuning GoogLeNet

Conclusion

Fine-tuning allows us to bring the power of state-of-the-art DCNN models to new domains where insufficient data and time/cost constraints might otherwise prevent their use. This approach achieves a significant improvement of average accuracy and improves the state-of-the-art of image-based medical classification.

In my Part 3 of this blog series (coming soon), I will explain re-usability of these trained DCNN models.