Long-term Recurrent Convolutional Network (LRCN) vs Large Language Model (LLM)
Can an LLM do everything that an LRCN can do? Yes. But…
The LLM is extremely powerful, general, and useful. It’s possible to load a “lightweight” or “pruned down” LLM on an expensive edge computer, but mostly inferences are run on the cloud where the trained LLM is deployed. Every cloud inference costs money. An internet connection must exist. There is latency.
The LRCN is a customized, highly-specific, streamlined model that can be loaded on a small, inexpensive edge computer such as an Nvidia Jetson. The model is built and trained for specific applications and therefore is the most computationally efficient possible.
Consider an application for a patient monitor in the ER. Using the incoming metabolic data, the LRCN is trained to predict if the patient is going to experience cardiac arrest during the next 30 minutes. The model will constantly be performing inferences at the edge and if necessary, post alerts/warnings to the screen.
Do you want all those inferences depending on a cloud model through the internet? No, you want complete independence from the internet to make sure that the system is always up with minimal latency.
Or, do you want a large, power-intensive edge computer onsite? No, you want a tiny, low-power edge device that goes unnoticed.
Consider an application that monitors multiple bearings on a vehicle. Using accelerometers, the LRCN will predict if any of the bearings it is monitoring will fail during the next 5 hours.
For the same reasons, the application-specific LRCN is a better choice than the general LLM for this or any similar critical application.
Previous work by Tor Hydra’s Principal involved converting data to pictures for classification of patient conditions (US 12061670, Machine-learning image recognition for classifying conditions based on ventilatory data, August 2024). Tor Hydra’s method takes it to a much higher level, converting data into videos for prediction of future events. Feel free to contact us with any questions.