In 2024, artificial intelligence may be poised to learn right at your fingertips.
Introduction
Ah, the world of AI is largely confined to the realm of cloud computing, with only occasional forays into our pockets via our smartphones. But when you interact with a tool like ChatGPT, the magic happens behind the scenes, in the vast AI data centers constructed by tech giants like Microsoft. These facilities are the unsung heroes of AI, toiling away days, weeks, and months in advance to ensure that the AI systems function seamlessly and accurately. It’s a truly impressive feat of engineering and computational power, and it’s what makes AI-powered tools like ChatGPT so remarkably effective.
2024 might just be the year when the proverbial rubber meets the road when it comes to artificial intelligence! Imagine a world where AI learns at your fingertips, rather than relying on the cloud for processing power. It’s no longer a pipe dream, as there are concerted efforts underway to enable training of neural networks, including sophisticated language models, on your very own personal device. No more relying on external servers or cloud infrastructure; the future of AI is coming to you!
On-device training has some seriously amazing advantages! Let me tell you – it’s all about convenience and personalization. By training AI right on your device, you can skip the lag that comes with connecting to the cloud. It’s like having your own personal AI coach that’s always with you, learning from your every move. You know, like how you tap, scroll, and drag on your phone.
It’s like AI is right there with you, learning from your everyday actions. And the best part? It’s super private! By training on your device, you’re keeping your personal data safe from prying eyes in those big cloud data centers. It’s like AI is learning from the world around you, right there with you, in real-time. The possibilities are endless – AI could become so much more than just a generic, cloud-based solution. It could be tailored to your very own actions and environment, making it feel like a true extension of you! 🤩
Apple engineers have been hard at work on a groundbreaking project to bring advanced neural networks, specifically the “generative” type pioneered by OpenAI’s ChatGPT, to run directly on iPhones. This represents a significant leap forward in the field of artificial intelligence, as these neural networks have the potential to revolutionize the way we interact with our devices.
In related news, Google has been at the forefront of a smaller, more efficient approach to AI called TinyML. This innovative technology allows neural networks to run on devices with minimal power consumption, such as smart sensors on machinery. While this has proven to be a game-changer in terms of energy efficiency, the real challenge lies in enabling these neural networks to not only make predictions but also learn and adapt on the go.
Training a neural network requires significantly more processing power, memory, and bandwidth than simply using it to make predictions. As such, technology companies are facing a formidable challenge in developing hardware and software that can support this local training capability on mobile devices. Nonetheless, the potential rewards are well worth the effort, as the ability to train and learn on the go has the potential to transform the way we interact with AI in the years to come.
Attempts have been underway to conquer the complex realm of computing in the field of artificial intelligence by employing various strategies, such as selectively updating specific portions of the neural network or using transfer learning to refine already mostly-trained models. One notable example of this is MIT’s TinyTL, which leverages transfer learning to improve the performance of a neural network on small tasks like facial recognition. However, the current state of the art is shifting towards tackling larger and more complex AI models, such as large language models (LLMs) developed by OpenAI, including GPT-4. These models boast hundreds of billions of neural weights that must be stored in memory and updated in real-time as new data becomes available.
This unprecedented training challenge requires a significant increase in computational power and memory capacity, as highlighted in a recent research report by STMicroelectronics. The authors, Danilo Pietro Pau and Fabrizio Maria Aymone, emphasize that simply performing inference on mobile devices is insufficient, as the AI models’ performance deteriorates over time due to a phenomenon known as concept drift. The solution lies in updating the models with new training data to keep them fresh and accurate.
The authors aim to simplify neural network training on memory-restricted devices by streamlining the training process. To achieve this, they experiment with eliminating backpropagation, the computationally intensive component of training in large language models (LLMs). By replacing backpropagation with simpler math, they found that the amount of on-device memory required for neural weights can be reduced by a remarkable 94%.
Another approach to tackling this challenge is through federated learning, where the training task is split across multiple client devices. This decentralized approach allows for more efficient use of memory and computing resources, making it possible to train models on devices with limited capacity.
Researchers at Kyung Hee University, led by Chu Myaet Thwal, have recently developed an innovative approach to train a large language model (LLM) across 50 workstations, each equipped with a single Nvidia GPU gaming card. This feat was accomplished by adapting the standard LLM used for image recognition, which resulted in reduced memory usage on the devices without compromising accuracy.
Meanwhile, some experts emphasize the need to optimize network communications to improve the performance of federated learning, a technique that enables mobile devices to collaborate in training the LLM. By fine-tuning the network communications, these devices can communicate more effectively and enhance the overall performance of the LLM.
Scholars at the Institute for Electrical and Electronic Engineering have proposed an innovative communication network utilizing the upcoming 6G standard, where the majority of large language model (LLM) training takes place in a data center before being fine-tuned by a network of cloud-connected client devices. This “federated fine-tuning” approach enables each device to contribute to the LLM training process by leveraging its local data, reducing the overall processing power required on battery-powered devices compared to full training.
To further optimize the process, researchers are exploring various techniques to minimize the memory and processing required for each neural weight. One promising approach is the use of “binary neural networks,” where each weight is represented by a simple one or zero, significantly reducing the on-device storage needs. This novel approach has the potential to revolutionize the way LLMs are trained and deployed, enabling more efficient and effective language processing on a wide range of devices.
the technicalities of neural nets may seem complex, but let’s dive into the real-world applications of training these models locally. Take the example of a team at Nanyang Technological University in Singapore, who used on-device learning to create a tailored “intrusion-detection system” (IDS) for cybersecurity. By training the IDS locally on individual devices, they could fine-tune the model to better detect threats specific to their network environment, without the need for constant communication with a central server. This not only enhances security but also prevents sensitive data from being transmitted across the network, where it could fall into malicious hands.
And it’s not just in the realm of cybersecurity where on-device learning is being explored. Apple, for instance, has been hinting at integrating more AI capabilities into their iOS devices. In a recent paper, Apple scientists described a “Never-ending UI Learner” program that can automatically learn the functionalities of mobile apps by interacting with them on a smartphone. This on-device learning approach allows the app to adapt to the user’s preferences without relying on remote servers, further enhancing the user experience.
Apple researchers are exploring a novel approach to train AI models on mobile devices without relying on human annotators. By leveraging the power of federated learning, they aim to automate the process of training AI models on individual devices, rather than relying on a team of workers manually annotating app functions. In a controlled environment, the team tested the approach and found it to be effective. However, if the trial were to be conducted in the real world using actual customer iPhones, a privacy-preserving approach would be necessary, such as on-device training, to protect sensitive user data.
In another study published in 2022, Apple scientists delved into the realm of speech recognition AI on mobile devices. They investigated the possibility of training large-vocabulary neural language models using the federated learning approach, which enables the training of AI models on resource-constrained devices without compromising on accuracy. By harnessing the power of distributed learning, the team was able to achieve impressive results, paving the way for more efficient and effective AI training on mobile devices.
Researchers are working tirelessly to develop innovative ways to train neural networks on battery-powered devices, such as smartphones, by leveraging samples of interactions with virtual assistants like Siri. The goal is to compress and distribute the training process across multiple devices to make it feasible for mobile devices with limited memory and processing power. While it’s unclear whether this effort will lead to breakthroughs in 2024, one thing is clear: the training of neural networks is on the cusp of a major transformation, with the potential to move away from cloud-based computing and into the palm of your hand.
2024 marks a significant milestone in the evolution of artificial intelligence, as it transitions to a more accessible and intimate form, literally fitting in the palm of our hands. The convergence of advanced algorithms, computing power, and compact mobile devices has empowered individuals with unprecedented opportunities to engage with AI on a personal level.
From virtual assistants enhancing daily productivity to immersive educational tools revolutionizing learning, the potential for AI to enrich our lives is boundless. However, as we embrace this transformative technology, it’s crucial to remain vigilant about ethical considerations, privacy concerns, and the equitable distribution of its benefits. As we navigate this exciting frontier, let us harness the power of AI responsibly, ensuring that it continues to serve as a force for progress, innovation, and human empowerment in the years to come.