Analysing Semi-Supervised ConvNet Model Performance with Computation Processes

The emergence of new Semi-Supervised Machine Learning (SSML) algorithms is becoming increasingly versatile leveraging different components in their structure. This versatility of design in SSML algorithms offered an improved overall performance, in terms of accuracy, training loss, and training time. However, identifying the key contributing factors for those SSML algorithms is still unknown. In the previous studies, researchers have utilized Deep Neural Networks (DNNs) to develop the primary learning models employing SSML algorithms with better classification accuracy. However, these research works lack a systematic study to grasp the impacts of primary contributing factors. In this article, we will describe some of those performance factors of the SSML models, their expected impact, and potential biases using Post-hoc Explainable artificial intelligence (XAI) methods. The audiences for this work are experienced AI researchers that have insight into SSML models. We start by generalizing the common components of SSML algorithms found previously (Π-Model, Temporal Ensembling, and Mean Teacher) compared to the new emerging contemporary SSML algorithms (MixMatch and RemixMatch). We then reform the Semi-Supervised Computation Processes (SSCPs) of both preprocessing focused and classifier focused designs. Subsequently, we introduce various types of DNNs in the network classifier module during the computational process. This integration demonstrates two main aspects to explore the performance effects: (1) parameter updates throughout the training process across both labelled and unlabelled data; (2) the labelled/unlabelled data ratio. Furthermore, we demonstrate our in-depth analysis of the 45 experiments with five different SSML algorithms, three DNNs, and three different labelled/unlabelled data ratios. Our experimental results show that one of the dominating factors to the learning precision and training loss is the parameter updates. Moreover, the experimental results also exhibit that training loss has been reduced by 8% and learning precision is improved by 6.75% using the network classifier, Shake-Shake26, in the RemixMatch SSML algorithm, compared to various other combinations. Similarly, the R score value of 0.701 and the p -value of 0.003597 clearly emphasize a positive correlation between the training time and the labelled/unlabelled data ratio. Finally, we conclude that the introduction of more labelled data precedes to prolonged training time that prompts additional parameter updates during the forward- and backward propagations.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.