Artificial intelligence students need GPUs for their projects
In the past, university projects were not as costly as nowadays. AI students working on interesting projects need to be prepared to invest significant time and financial resources in computing power which is needed to optimize modern AI models such as deep neural networks.
These GPU hungry AI models can be trained locally (e.g. if you have an NVIDIA RTX 30 Series which is getting more affordable these days) or in the cloud. Many students pay for Colab PRO, where older NVIDIA Tesla P100 are available.
The best option is when their university can provide GPU computing infrastructure that is well accessible and available for students any time they need it. This is however not often the case. When universities run their private GPU clouds, they are often available mainly for researchers and student projects do not have access or priority.
This year at Faculty of Information Technology, Czech Technical University in Prague, thanks to generous support of NVIDIA and M Computers, a GPU cloud dedicated to student AI projects was made available. It was based on the NVIDIA DGX V100 machine managed by Petr Kasalický, our doctoral student.
We have prepared a short overview of selected student projects that were using the GPU cloud. As you see, artificial intelligence can be applied in many different fields.
Flow modeling around airfoil with graph neural networks
David Horský in his thesis used Graph neural networks for simulation of flow around the airfoil. He used the provided server for training of the neural networks. He trained several networks with differing structures. The best network is capable of simulating airflow with reasonable precision while increasing the speed of simulation by up to two orders of magnitude.
Improving Interactive Voice Response
Martin Nykodém used the DGX machine to train the automatic speech recognition model wav2vec2 for the practical part of the thesis, which focused on dictated number recognition for an IVR company. The thesis results were so compelling that he is currently working on founding a startup that will create a human interaction level AI solution for companies to communicate with customers over the phone, which will train directly on the dialogues between coordinator and customer and system actions taken during that interaction.
Merlin: Recommender Systems pipeline implemented on GPUs
Čeněk Sůva used the DGX machine as a computational resource for exploring and using the NVIDIA Merlin framework. The goal was to create a recommendation system model. Tools that were being used daily are Python 3, Pandas, JupyterLab, Tensorflow, and NVIDIA Merlin components. A typical workload consisted of data preprocessing, model training, and inference. The process is explained in his thesis.
Deep learning techniques for identifying selected heart conditions from electrocardiography signals
Valeria Pak used DGX to build and train deep learning models for signal processing in medical applications, as part of her diploma thesis. The thesis focused on identifying selected heart conditions from electrocardiography signals. The thesis explored various deep learning techniques for signal processing and compared their performance on publicly available datasets. GPU cloud helped her to explore more architectures in a shorter period of time.
Age prediction based on 3D facial scans
Filip Žďánský I used the machine to estimate age from three-dimensional scans of the human face, which were in the form of a point cloud and two-dimensional views of the scan. The results are documented in his bachelor thesis. It explored various machine learning methods targeting this problem.
Improving cold start recommendations using convolutional neural networks enriched by interaction data
Kamil Agha Kader used the DGX Station V100 for his diploma thesis research. The study aimed to build a recommender system prototype based on image data and user interactions with the help of convolutional neural networks. The thesis shows how modern vision CNN can be fine-tuned for the recommendation task.
Evaluating defect indications by magnetic powder method and fluorescent penetration method by neural networks
Matěj Latka used the DGX Station V100 as part of his diploma thesis for training a detection model to automatically evaluate defect indications by magnetic powder method and fluorescent penetration method. He experimented with several architectures of advanced convolutional neural networks.
Conclusion
As you can see, student projects can be really interesting and diverse. Most of them are not toy projects, but real AI applications with a potential to end up in industry and improve the efficiency of existing solutions.
Therefore, the Faculty of Information Technology decided to support student projects with the more powerful NVIDIA DGX A100. The existing DGX V100 was purchased by Recombee and it will support student research in the area of recommender systems. We believe more universities and companies will follow our example and dedicate GPU resources to student AI projects.
