I'm a 4th-year undergraduate student advised by Prof. Jong Hwan Ko in the IRIS Lab at Sungkyunkwan University. My research interests are:
On-device AI · Dynamic Neural Networks · Model Quantization.
- Reduce the cost of inference, which impacts key user requirements (latency & performance)
- Streamline the development process.
What are dynamic NNs? More here.
RECENT FOCUSES
More efficient flexible bit precision models.
During runtime, dynamic neural networks generally use quantization as a means to change the bit precision of a model. However, the existing quantization method does not allow the sharing of most-significant-bits (MSBs), which leads to the model having to reference the full-precision value every time it wants to convert to a lower precision version of itself. This incurs memory and computational overhead, which can be reduced by a simple solution- bit-shifting. My current research explores how neural networks react to straightfowardly bit-shifting their weights, and the impact this has to the overall performance.
Shorter training time for dynamic neural networks.
Dynamic neural networks achieve great robustness during runtime by training for every scenario beforehand. Naturally, this leads to an exponential growth in training time, which 1) not only hinders the deployment of these models but also 2) results in huge carbon emissions. Interestingly, dynamic neural networks need not to train for every setting it will encounter in runtime, but just a judiciously selected few which can maintain performance for most settings of deployment. My current research focuses on how to find these selected few, and ultimately optimize the training process for dynamic neural networks.