Lousy dataset? Tips for training your neural network

Kornel Dylski
May 20, 2019

Machine learning is often described as a way to make computers perform some actions by showing them initial data and action results instead of giving them explicit instructions of a particular action. For image recognition tasks we provide the ML model images together with described images to make the model finally recognize the images on its own and provide recognized images. Machine learning models attempt to approximate information from the received data to use them further. When a dataset is vast enough, a model can learn the majority of distribution, and new samples are easily approximated. On the other hand, poor dataset results in lousy approximation. Unluckily gathering and preparing data is usually the most expensive task and takes the most time and sometimes you just cannot afford to get more data.

Recently we were creating a neural network for image recognition. The size of the training dataset was so limited that we knew a standard approach was doomed for failure. We didn’t have tons of data to train the NN and therefore to succeed we had to try something different. In this article, I describe the reasoning process behind our efforts. I would like not to focus on specific libraries but more on general procedures one may use. I strongly encourage you to dig more into each approach if you consider it useful.

When we can search the internet for data #

The first technique, called transfer learning comes from the idea that it is perfectly ok to recycle and reuse a model that has already been trained on a large general set instead of training model from scratch. For general cases, we usually can find a similar data set and a model that is already trained on it. Because random parameters of the model have much less chance to approximate solution, almost always a better idea is to train using a model that was pretrained before. I encourage you to search through the model zoo and state-of-the-art to draw from the experiences of others.

Regardless of whether you found a pretrained network that suits your needs, you have to collect training data for your specific task. You should spend time with the search engine, even if you are able and ready to collect data manually. Maybe someone has already done this for you.

When collecting pictures, it’s easy to forget that the model can not find a solution if the information it needs is not included in stored pictures. Assume you want to classify mammals based on their pictures. The dataset needs to be carefully planned if you want the network to be well-trained. There is a need to collect additional photos to make the results more independent of the style and the source of data — for example, a sketch of an elephant or a picture of a cat at night. Additional categories may be useful if you expect that model to see objects other than mammals. Otherwise, the network has to assign an unrecognized object to one of the categories of mammals. Next, do not forget about exceptions like dolphins. The model could easily confuse the fact that their pictures almost always contains water environment. Perhaps the model should see a few dolphins in unique environments to learn that the actual shape of the dolphin causes it to be a dolphin, not water around. Also, the picture of the dolphin’s rear fin can be very different from the picture of his head. In the subject of pictures of objects, people or nature, almost everything has been already said, so you will quickly find many examples in Google Images to fill your dataset. Dolphins from each side and many more. And the same concept of trying to cover the entire distribution also applies to other, less obvious themes.

dolphin and a dog, both classified as a dolphin
both pictures classified as a dolphin

When we have exhausted the internet data sources #

Data augmentation, which is another useful process, is necessary for the model to prevent learning by heart when it sees the same picture repeatedly. To make the model more aware of what’s in the picture, each picture is used with random transformations. These transformations may include rotation, scaling, inverting the image but also adding noise and distortion. Actually, we can use any function if the resulting image still unambiguously represents the same object. Data augmentation is a simple technique to ensure that the model is not paying too much attention to the specific location of the pixels.

picture of squirrel with applied random transformations
squirrel image with applied six random transformations

Worth to mention step before data augmentation is scaling. Usually, using a pretrained network, you are condemned to the shape and size. Every picture has to be rescaled to fit into the network, and if the picture does not match the same proportion also has to be cropped or get some padding. Cropping can lead to information loss from the picture, so the preferred option is to add padding. What can fill for extra pixels? It may be a solid color, blurred edges or side reflection of the picture. For instance for space photos, black background as padding would work well. For landscape photos, reflection operation will do the job.

elephants picture with different paddings
reflection seems to be the most natural here

When pictures are prepared, you want to perform data augmentation. So here comes one of not the very common but yet an effective technique is the cutout regularisation. The idea is to cut out a randomly placed rectangle from the picture. The network should figure out what is missing in a similar way as humans guess which puzzle is missing. Thanks to the example with the GAN used, we can see that the network can restore the missing part. This may remind you of a commonly used dropout regularisation technique, in which some neurons can be randomly switched off during training to prevent overfitting and to improve results.

lion picture with and without cutout
cutout regularisation

Another effective technique is a mixup. In simple words, when two images are overlapping on each other in some proportion, a neural network can be trained to distinguish that. For the picture on which 50% of pixels are from the image of the dog and rest is from the image of the cat, the network has to be told that the solution is a half-dog half-cat picture.

Among the more sophisticated techniques used for augmentation are Generative Adversarial Networks. GAN is a kind of network that is composed of two subnetworks. First one (the generator) is consistently trying to generate images from random noise. The second subnetwork (discriminator) is processing alternately images generated from the generator and real images. The discriminator is judging which image is real and which one is generated. Both subnetworks are trained until the generator fools the discriminator, and the generated images no longer stand out from the real ones. Generated images can be successfully used to enlarge the size of the data set. You can read about it further in Low-Shot Learning from Imaginary Data paper where this approach was used.

However, there are many cases when we are the first in the field, and there are no pretrained models yet. If there are also no datasets available, data must be collected manually, and it is not easy to speed up the process. For such a sparse dataset data augmentation may be not enough and creativity must come in place.

When our dataset is still poor #

If the current data set is still not large enough to improve training, you may want to start labeling images. Images can be labeled manually or by others with some polling. However, both are money- and time-consuming. There is a branch of machine learning called semi-supervised learning. Basically, the network is figuring out the solution based on parameters on subsequent layers. The last layer outputs solution. The previous layer outputs something that represents the features of the image. Using various methods like t-SNE, you can place each picture on the plot according to their features. Pictures with similar features will be plotted close to each other and you can expect that these pictures represent the same category. Thanks to this, you have to categorize only pictures with features far from others. Then you can retrain the network using more labeled pictures. Labels generated this way are somewhat unsure. Assuming that the network can already recognize trivial examples with a low error, you should accept the labels generated this way, only for pictures with a similar low error.

DIY - make your own dataset #

When there is no way to gather more pictures and typical transformations are not enough, you may consider creating a totally custom generator. If you think about it more, creating sufficient generator should be possible.

Implementation of a generator is quite obvious. Take as an example network for recognizing objects packed into a box. Arrangement of objects or shape of the box does not play such a significant role for recognition. Every object and every box can be separated from each other. When the set is built with separated objects and boxes, you need to write and tune up the generator which always takes a few of the objects for a newly generated picture one box underneath and place them on the picture. You should think about how to place them naturally and how objects should overlap others. Thanks to this, an infinite number of pictures can be created by combining the boxes with objects. For a dataset with tens, not thousands of pictures it is a blessing.

In some products, a more advanced generator is desired. It is very interesting to employ a rendering engine, which can reproduce a real object. Creating a realistic picture requires a lot of talent and precision. Usually, you will be condemned to basic models that can be created and rendered in a relatively short time. Fortunately, with domain adaptation, poor reproductions can be used as well. GAN network can be trained to erase differences between a source (rendered picture) and a target (real picture) domains. This time generator subnetwork has to deceive a discriminator until the differences of pictures from different domains, representing the same object, will not cease to affect the classification.

Conclusion #

A truth nobody can deny is that a big enough dataset is crucial to perform machine learning successfully. However, due to the lack of necessary data, many parallel topics are being developed to reduce the effort required to provide them. I can’t wait to see further progress in that area. And of course, if you reading this article, have something to add, please share!

Now, let's talk about your project!

We don't have one standard offer.
Each project is unique, rest assured that we will approach the next one full of energy and engagement.