As Artificial Intelligence gets more & more pervasive in a wide variety of different applications, it becomes more important to be able to train a model fast, and make it easily deployable. We have developed a set of tools to facilitate training and deployment, and provide these as a service for providers of actual Artificial Intelligence applications.
Technology Features, Specifications and Advantages
In recent years, deep neural networks have been applied to and achieved great success in many different tasks such as face recognition, cyberbully detection and speech recognition. However, there still remains a few hurdles for deep neural networks to fully realise its potential and play an effective role in the real world.
Firstly, the exact learning mechanism of deep neural networks still remains a black box. Building a good deep neural network model including hyper-parameter tuning still heavily relies on expert insight and experience from deep learning researchers. Meanwhile, good performance from deep neural network also means high complexity, which requires extremely high computational power and memory support, which is unrealistic for low-end devices such as embedded systems. Besides that, deep neural networks also require data for training, and fails to work with limited data or faulty datasets.
To tackle these three problems, we created a toolkit platform which provides the following solutions with machine learning technologies that better trains deep neural networks :
- Automatic hyper-parameter tuning with Bayesian optimization;
- Automatic model compression with singular-value decomposition (SVD) and clustering;
- Semi-supervised data cleansing for training data;
- Semi-supervised data augmentation for training data.
Any service or product that requires machine intelligence as an enhancement or value-added feature. Examples include:
- Trend predictions and classifications for commercial applications such as spam detection, emotion detection or cyberbully detection;
- Object detection and segmentation;
- Speech/speaker recognition;
- Chat-bot or FAQ system;
- Document summarization
- Predictive Maintenance of Machines;
- Big data analytics
- For the business users, our technologies are specifically for model compression, data cleansing and data augmentation that adapts deep neural networks to work in real-world applications. Also, it enables AI technology and state-of-art results from the product without requiring deep expertise in AI technology.
- For the research user, with the tool handling the tedious tuning process, he/she can focus more on model architecture and concentrating on design and improvement, thereby achieving better performance with the boost from our tools.