WHAT WE DID
Tin AI is a leading provider of embedded artificial intelligence solutions for smart devices and IoT applications. With a focus on efficiency, accuracy, and scalability, Tin AI offers a range of AI models and services to enable intelligent edge computing and enhance user experiences.
Building a decentralized training cluster
The embedded models required for Tin AI's solutions are trained on a large corpus of data to ensure high accuracy and performance. To meet the computational demands of training these models, Outroll built a decentralized training cluster that leveraged the power of distributed computing to accelerate the training process. By distributing the workload across multiple nodes, Outroll was able to reduce training times by up to 50%, enabling Tin AI to iterate on model development more quickly and efficiently.
Optimizing model architecture for x86 and ARM processors
Tin was built with the focus to run on standard CPUs without the need for specialized GPUs and TPUs. This required additional research and development to build models that are performant on x86 and ARM processors. Outroll worked closely with the Tin AI team to optimize the model architecture for both processor types, ensuring that the models could run efficiently on a wide range of devices without sacrificing performance or accuracy. The team leveraged a custom Go implementation with optimized assembly components to achieve the desired performance levels, resulting in models that could be deployed on a variety of devices with minimal overhead.
Implementing a custom quantization pipeline
To further enhance the efficiency of the models, Outroll developed a custom quantization pipeline that reduced the size of the models by up to 50% without sacrificing accuracy. By quantizing the model weights and activations to 2-bit formats, Outroll was able to significantly reduce the memory footprint of the models, making them more suitable for deployment on resource-constrained devices. The custom quantization pipeline also included techniques for fine-tuning the quantized models to maintain high accuracy levels, ensuring that the models performed optimally in real-world scenarios.
The next step: Enabling on-device training and inference
Outroll is now working on new ways to enable on-device training and inference for Tin AI's models, allowing users to customize and fine-tune the models directly on their devices. Once this stage is complete, Tin will open its Beta to a wider audience, enabling developers and researchers to experiment with the models and provide feedback for further improvements. The goal is to democratize AI development and make it accessible to a broader audience, empowering users to create intelligent applications that meet their specific needs and requirements.
