Deep Learning for Applications
Deep Learning for Applications
After 2012 when deep learning based techniques won the ImageNet contest with a clear margin to competing algorithms, deep learning has been called “the revolutionary technique that quietly changed machine vision forever”. For many classification tasks deep learning has drastically surpassed previous state of the art results in classification accuracy. Currently large deep neural networks achieve the best results on speech recognition, visual object recognition, character recognition, and several language related tasks.
The power of deep learning
Deeper machine learning architectures are better capable of handling complex recognition tasks compared to previous more shallow models. Major benefits of deep networks are:
- their superior modeling capabilities of heterogeneous data in layers of increasing complexity,
- their ability to learn the best features to represent the raw data, and
- their ability to gain in performance with the availability of more training data.
Specialized deep learning
In general, deep learning based techniques require a vast amount of data for training. The system that won the ImageNet contest in 2012 had for instance been trained on 1.2 million labelled images. For most problems such amounts of data are not available. Methods that can utilize existing networks learned from large general resources, like ImageNet, to solve problems in other more specific domains can then be a solution.
Another challenge is the requirement for processing speed in the recognition process. Some problems may require real-time processing, and such constraints need to be considered when designing the system. A third problem is related to the system architecture. Although manual feature design is replaced by the deep learning network’s ability to learn the best features, there are still several challenges related to how to best design the network for various problems. To obtain good network architecture, detailed knowledge of the specific problem is needed.
Combining years of experience in pattern recognition and machine learning
The pattern recognition group at NR and the machine learning group at UiT now join forces to better understand the needs and to develop state-of-the-art specialized deep learning solutions suitable for solving specific problems for various industry-, medical and environmental applications, and within analysis of remote sensing images.
Current activities and projects involving deep learning:
- Transcription of historical Norwegian census forms
- Detection and counting of seals on ice from aerial images
- Detection of cultural heritage sites from laser data
- Segmentation and detection of small objects in high resolution remote sensing data
- Recognition of animal species from wildlife cameras
- Estimation of bone age from MR images
- Classification of behaviour from financial transaction data
- Tracking of small roads and forest paths from laser data
- Detection of avalanches from high resolution optical satellite images
- Detection of oil spills from remote sensing radar data
- Segmentation and recognition in ultrasound images
- Detection of malware-generated domain names