High performance machine learning platform development and applications

Abstract

Machine learning (ML) and Deep learning (DL) are booming. However, the algorithms beneath the MLs and the DLs involve much high-density computing work. Traditional Central Processing Unit (CPU) only has few computing threads. And each thread must execute instructions in sequence. So even the CPUs own a higher frequency, they have to process tens of thousands of computing jobs one by one. On the contrary, Graphics Processing Unit (GPU) and Field Programmable Gate Array (FPGA) often own tens of thousands of computing units, which can efficiently conduct parallel computing to enhance the model’s training and inferencing performance dramatically. We propose in this thesis a novel hybrid GPU-FPGA-based design methodology to address the above training and inferencing high-density computing challenge. According to the design methodology, we developed a new heterogeneous GPU-FPGA-based ML or DL platform. Since the training algorithms are often changed, and the GPU programming is much easier and more flexible than FPGAs, the training phase is implemented on the GPU. Otherwise, we perform the inferencing phase on the FPGA based on the following two reasons: One is that we scarcely change the inferencing algorithm design. The other is FPGAs have higher energy efficiency and lower delay than GPUs’. Moreover, since the two platforms have different model file formats and can not be substituted directly, we designed a model converter between the two phases to convert the model from the training platform to the inferencing platform. To evaluate the above methodology and platform’s performance, we have implemented a convolutional neural network (CNN) for recognizing handwritten digits and a deep neural network (DNN) for predicting the data center’s Power Usage Effectiveness (PUE) with the hybrid design methodology on the heterogeneous platform. Moreover, the experimental results presented that our approach has gained significant performance improvement on the ML training and inferencing. In addition, in order to fully evaluate our hybrid ML design, we extended the experiments to include two spectral reconstruction approaches. Both of these methods are designed to solve the severely under-constrained problem of reconstructing multispectral images from RGB images. The core of these two methods revolves around how to generate multispectral information that is lost due to compression. The first approach is based on Variational Autoencoder (VAE) and Generative Adversarial Network (GAN). The VAE extracts the key feature information from the input RGB images through the encoder, reparameterizes it with a value randomly sampled from the normal distribution, and then restores the MSI-like outputs through the decoder. GAN is responsible for training the generator to generate MSI-like pictures from re-parameterized latent vectors. GAN is accountable for teaching the generator to create MSI-like images from re-parameterized latent vectors. The problem of reconstructing MSIs from RGB can be solved with low computational cost. The second method is called Taiji Generative Neural Network (TaijiGNN), which combines cycled GAN and ancient Chinese "Taiji" philosophy. TaijiGNN consists of a pair of generators performing in opposite directions. The output of one generator is connected to the input of the other one, forming a loop structure. This loop structure can pass the input through the output domain and then back to the input domain. Therefore, TaijiGNN can convert the problem of comparing images in different domains into the problem of comparing images in the same domain. In the same domain, the severely under-constrained problem mentioned above can be solved naturally. Moreover, TaijiGNN has absorbed the essence of Taiji to train the pair of generators. During the training process, the pair of generators work like a couple, using their own advantages to complement each other, helping each other to achieve convergence, so that the entire system enters a state of dynamic equilibrium, which is very similar to Taiji’s "Yin" and "Yang" bipolar working mode. We use two classic spectral datasets, CAVE and ICVL, to evaluate VAE-GAN and TaijiGNN. And both of these two approaches use much less training data than state-of-the-art and reach or exceed their results. Moreover, the two approaches are implemented and verified on a heterogeneous computing platform designed using our hybrid machine learning methodology. Their training and inferencing speeds have been greatly improved.

Date	14 Oct 2021
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Abdelouahed Gherbi (Supervisor) & Mohamed Cheriet (Co-supervisor)