Abstract
Deep neural networks have gained immense popularity in the Big Data problem; however, the availability of training samples can be relatively limited in specific application domains, particularly medical imaging, and consequently leading to overfitting problems. This “Small Data” challenge may need a mindset that is entirely different from the existing Big Data paradigm. Here, under the small data scenarios, we examined whether the network structure has a substantial influence on the performance and whether the optimal structure is predominantly determined by sample size or data nature. To this end, we listed all possible combinations of layers given an upper bound of the VC-dimension to study how structural hyperparameters affected the performance. Our results showed that structural optimization improved accuracy by 27.99%, 16.44%, and 13.11% over random selection for a sample size of 100, 500, and 1,000 in the MNIST dataset, respectively, suggesting that the importance of the network structure increases as the sample size becomes smaller. Furthermore, the optimal network structure was mostly determined by the data nature (photographic, calligraphic, or medical images), and less affected by the sample size, suggesting that the optimal network structure is data-driven, not sample size driven. After network structure optimization, the convolutional neural network could achieve 91.13% accuracy with only 500 samples, 93.66% accuracy with only 1000 samples for the MNIST dataset and 94.10% accuracy with only 3300 samples for the Mitosis (microscopic) dataset. These results indicate the primary importance of the network structure and the nature of the data in facing the Small Data challenge.
Highlights
One possible solution to the small sample size problem is to use pre-trained networks[5,6], known as transfer learning
In fields using medical images, where the input pixel format (e.g., Non-RGB data such as grayscale ultrasonic/MRI images or RGB + Depth) can be entirely different from conventional photographic images, where using pre-trained networks from other image domains implicitly hypothesize that the optimal network structure is universal (“Transferrable” networks) and not data-driven
The questions are whether does the nature of the problem affect the optimal network, and the size of the set, and how does this change in sample size affect the structure of the optimal network. This can be seen in fields where the primary issue is the lack of a substantial amount of examples as well as extremely skewed classes in the training set and the ambiguity in the ad-hoc performance of the designed network
Summary
One possible solution to the small sample size problem is to use pre-trained networks[5,6], known as transfer learning. If the effect of “data size” is greater than the effect of “data nature” it means the complexity of the structure of the optimal network will increase as data size becomes larger In this case, we may blindly use pre-trained networks trained on different image modalities and always opt to increase the complexity of the used network if the pixel format is identical. We will calculate the best performance a convolutional neural network can achieve after optimizing the network structure and compare it with the average performance from randomly selected networks This heuristic serves primarily to enlist factors that affect the optimality of the network architecture as well as demonstrate the problem such networks face when provided with smaller training sets. We trained and tested each structure followed by layer dimension (layer width) optimization using small subsets (less than 5,000 samples) of these datasets, and investigated the performance difference in different network structures, as detailed
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
