In previous blogs, I have referred to the huge datasets used to train LLM generative AI models, and the need for huge power consumption and computing resources. According to OpenAI, the state-of-the-art deep learning model can emit up to 2.85 lac kg. of CO2, which equals the lifetime emissions of five cars. AI models have become complex.

In order to make AI more useful to the society at large, we do need lightweight AI models, smaller and efficient. They consume less power, and can be used on smartphones, tablets and IoT devices. They are easily accessible and affordable. They are environment-friendly. And of course, they are speedier. A heavyweight model takes several minutes or hours. Lightweight model makes a prediction in seconds. In autonomous cars or medical devices, this quality is highly desirable.

Lightweight models are trained on smaller datasets. Heavyweight models require millions of labelled data points to achieve accuracy. It makes the whole thing expensive. Lightweight models give comparable accuracy with smaller datasets. These make the models affordable.

However, this requires careful selection of features and alogrithms used. There should be optimisation of hyperparameters. Lightweight models are vulnerable to adversarial attacks. The attacker manipulates the input data so as to deceive the model to make false predictions.

It is necessary to use transfer learning where a pretrained model is fine-tuned on a smaller dataset. One can use model compression techniques, such as pruning or quantisation to reduce the size of the model. Instead of deep learning, some researchers suggest the use of alternative algorithms such as decision-trees or rule-based systems.

Lightweight models are useful for real-time decision making and in remote areas.