Cookie Consent

By clicking “Accept”, you agree to the storing of cookies on your device to analyze site usage, and assist in our marketing efforts. View our Cookies Policy for more information.


Retraining Test: GreenLightningAI vs ResNet 152

January 11, 2024


Gínes Sánchez

Unleashing GreenLightningAI: A Revolution in AI Efficiency and Sustainability


In the rapidly evolving landscape of artificial intelligence (AI), the demand for more efficient and environmentally sustainable models is paramount. We introduce GreenLightningAI, a groundbreaking AI system designed with a focus on efficiency, sustainability, and superior performance. In this blog post, we delve into the results of our extensive testing, comparing GreenLightningAI with well-known deep neural networks (DNNs) like ResNet 152, and explore some of its applications in computer vision, edge computing, and large language models (LLMs).


GreenLightningAI takes a divergent path from conventional neural network training. Unlike the resource-intensive retraining of DNNs, where old and new samples need to be processed to prevent catastrophic forgetting, GreenLightningAI just requires processing the new samples when retraining. This has been made possible by adopting a system design that decouples structural and quantitative knowledge. More precisely, GreenLightningAI utilizes a linear model capable of emulating the behavior of DNNs by subsetting the model for each specific sample. This linear model stores the quantitative knowledge while the structural knowledge is used to perform model subsetting.


Our proof of concept demonstrates that the structural knowledge stabilizes significantly earlier than the quantitative counterpart, and does not need to be updated when retraining. Thus, only the linear component of our AI system needs to be updated every time the system is retrained. This is the key to enable incremental retraining, and is responsible for the efficiency gains of this innovative design.

The result is an amazing reduction in retraining times and energy consumption by two or three orders of magnitude (x100-x1000), compared to DNNs.


Comparative Analysis with ResNet 152


ResNet 152, a residual DNN architecture, has been a stalwart in image classification tasks. Renowned for its depth, ResNet 152 has found applications in various domains, including computer vision, medical image analysis, and more. However, the increasing retraining time associated with ResNet 152 has become a significant challenge in the face of the growing demand for continual model updates.


In our head-to-head comparison with ResNet 152, using the ImageNet dataset, GreenLightningAI, powered by Qsimov, showcased unprecedented efficiency. While ResNet 152 required retraining with the entire ImageNet dataset (90% + 10%), GreenLightningAI only needed the additional 10%, highlighting its streamlined approach.

The impact on speed was staggering. Qsimov retrained the model over 30 times faster than ResNet 152 while maintaining inference accuracy. To put it succinctly, if we were to retrain with just 1% of new images, GreenLightningAI's expected acceleration would surpass a remarkable 300 times compared to the traditional ResNet 152 approach. These results, graphically represented, clearly illustrate the transformative efficiency gains achieved by GreenLightningAI in handling large datasets and redefining the benchmarks for retraining speed.


Use cases and applications

GreenLightningAI's superiority extends across diverse use cases such as incremental retraining, federated learning, incremental federated learning, and applications of these methods, making it a standout solution in the ever-expanding realm of artificial intelligence.


In computer vision, our revolutionary AI system not only outshines DNNs in computational efficiency while maintaining accuracy, but also offers a sustainable approach for continual learning and adaptation. GreenLightningAI ensures consistent stability in retraining times, eliminating the increasing burden faced by traditional DNNs like ResNet 152. The result is a more efficient and environmentally friendly solution for computer vision applications.


In the domain of edge computing, GreenLightningAI not only stands out for its capacity to work with low-precision arithmetic without reducing accuracy, but also eliminates the need to send data over communication channels to the cloud for model retraining. GreenLightningAI incremental retraining capability enables retraining directly on the edge device, a feature that is crucial for scenarios where real-time adaptability is paramount.


GreenLightningAI emerges as a cost-saving powerhouse. Traditional retraining methods for LLMs, exemplified by the staggering cost associated with training models like ChatGPT-4, are a significant bottleneck in AI development. GreenLightningAI's innovative approach breaks free from the high costs and energy consumption plaguing conventional methods. The result is not just efficiency, but a transformative shift towards a more sustainable and economically viable future for large language models.


Try GreenLightingAI in your applications


We invite you to explore the possibilities with GreenLightningAI. If you're eager to witness the transformative impact of our technology in your applications, reach out to us. Together, let's shape the future of AI—one that is powerful, responsible, and environmentally conscious.

Thank you for reaching out to Qsimov! Our team will be in touch shortly.
Oops! Something went wrong while submitting the form.

Get in touch

Reach out to us and transform your ideas into reality.
Thank you for reaching out to Qsimov! Our team will be in touch shortly.
Oops! Something went wrong while submitting the form.