Machine learning advancements in neural network training and industrial monitoring

Researchers have conducted studies on optimizing machine learning applications for both deep learning training and industrial process monitoring.

In Colombia, an experimental study focused on the compute-memory trade-off during neural network training. By using rematerialization techniques—which selectively discard and recompute intermediate activations—researchers aimed to reduce memory consumption on GPUs from NVIDIA and AMD. Using the GUANE supercomputer, the study found that while the technique provides no benefit for MLP and LeNet-5 architectures, it offers moderate improvements for AlexNet and significant memory reduction for VGG-16 with limited computational overhead.

In Chile, research at the Universidad de Concepción addressed industrial efficiency by developing soft sensors and predictive tools for green liquor filters. Due to the latency and poor quality of physical sensors in recovery boilers, the study utilized machine learning to estimate reduction levels from historical data. The findings indicate that predictive solutions are feasible, though performance depends heavily on variable selection, temporal structure, and training modes, with mixed training and explicit temporal structures yielding the best results.