On-Device Training

Introduction

Mirai is pioneering the field of on-device AI, focusing on training and fine-tuning neural networks directly on user devices. Our goal is to provide a flexible, powerful solution that reduces reliance on expensive server-side GPUs and brings AI capabilities closer to the end-user.

We aim to create a comprehensive ecosystem for on-device AI that includes:

1. Pre-installed models for immediate use

2. Tools for creating and converting custom models

3. A flexible SDK for on-device training and fine-tuning

While our current focus is on iOS and macOS, we plan to expand to other platforms in the future.

---

Mirai vs. Existing Solutions

CoreML and MLX

While solutions like CoreML and MLX exist, they have limitations:

• CoreML is primarily for inference and optimizing pre-trained networks

• MLX allows training but is largely limited to macOS

Mirai aims to overcome these limitations by offering a more flexible solution that works across all iOS and macOS devices.

---

Core Components

Our on-device AI solution consists of two main parts:

1. Model Registry

2. On-Device NN SDK

Model Registry

The Model Registry contains pre-installed models ready for on-device use:

1. Preinstalled-mirai: Models created by the Mirai Team

   • Currently focused on text processing

   • Future plans include image, video, and user behavior models

2. Preinstalled-community: State-of-the-art models and community contributions

On-Device NN SDK

Our SDK provides tools for creating, training, and fine-tuning neural networks on-device:

1. Data Tools:

   • Focus on textual data processing

   • Includes tokenizers, lemmatizers, embedding models, and text analysis tools

2. Operators:

   • Library of neural network building blocks (e.g., fully connected layers, normalization, activation functions)

   • Optimized for on-device execution

   • Supports various data types and parallel processing

3. Optimizers:

   • Implements backpropagation algorithm

   • Includes popular loss functions and gradient descent methods

4. Pipelines:

   • Framework for building data processing and training pipelines

   • Similar to Hugging Face pipelines

5. Metrics:

   • Includes essential evaluation metrics (e.g., Accuracy, Precision, Recall)

6. Converters:

   • Tools for converting PyTorch and TensorFlow models

   • Integrates existing tools (CoreML, TFLite) and custom converters

---

Future Plans

• Expand device and OS support beyond iOS/macOS

• Broaden model types to include image, video, and user behavior

• Integrate with Hugging Face ecosystem

• Continuously improve and expand SDK capabilities

By leveraging Mirai's on-device AI solutions, developers can create more responsive, private, and efficient AI-powered applications that run directly on users' devices.