Online recommendations | Your privacy is precious and should not be sacrificed for better internet recommendations. How to combine these two aspects today? | Source: Getty Images
INTERNET RECOMMENDATIONS | Robots know your preferences. They know which song or video to queue in your playlist. They are watching you shopping and may offer you pants that match the shirt you just put in your basket. However, all these recommendations are not free: to benefit from them, you sell your private life.
To give you the best possible suggestions, an algorithmic referral system needs to learn more about you, and this can go beyond what you are willing to share. In other words, bots usually update your information in a centralized database that you cannot control.
Under the guise of improving the future of technology information, you are constantly being asked to disclose more and more personal data. However, do you really want Facebook or Netflix to know all the details of your privacy whether you are using their apps online or offline? Even if you trust them, their huge centralized databases are prime targets for hackers precisely because these databases know a lot about you.
The current recommendation systems do not know you, they categorize you
It is difficult to provide recommendations on access devices, as the usual machine learning algorithms cannot work effectively on these low compute and low power devices. Conversely, the most modern systems absorb huge amounts of data in the cloud . Then, using virtually unlimited computing resources, the deep learning algorithms run on this centralized database to then deliver the best recommendations right to your mobile phone.
However, this type of centralization has a downside: it is too slow to really get to know you. How many times have you received a recommendation for a show or series from an algorithm when you have never watched that series and don’t intend to? The problem is, you are categorized, not understood. You are a single small data point in an ocean of big data for algorithms for social media platforms, retail sites and streaming services.
Focus on IT periphery could ensure confidentiality and real understanding
Recommendation systems based on edge computing could solve this problem. By running the software only on your device, your privacy will not be scrutinized and monetized on a server in cyberspace. This will give you control over your data and more confidence in the system. As a result, the level of understanding will be higher and you will get better recommendations. However, if the answer to the problem is edge computing, how do you bridge the processing gap between cloud computing and edge computing?
The material is an obstacle
The existing solutions have focused on the material , which constitutes the main constraint for information at the periphery. Currently, many companies are developing specialized machine learning accelerators to improve edge computing. However, there is one flaw with these technologies: scalability.
Your phone can only hold a certain amount of data and current computing capacities are limited. The development of integrated circuits specific to (ASIC) applications or programmable door networks in situ (FPGA) is difficult and expensive. These chips are often developed using older processing technologies due to the low production volumes making them slower than general purpose processors.
In Due to their specialization, these chips generally do not have sufficiently developed ecosystems for debugging, optimization and security. However, these features are common on modern generic processors. As a result, machine learning accelerators generally have a low adoption rate. Leveraging their advantages is often too expensive compared to versatile hardware.
You’ve probably heard of the latest chips developed by Apple and their ability to model AI. This is only part of the story. Millions of people have access to the latest hardware from Apple. Billions don’t have access to it.
What if the solution came from the software
For all of these reasons, the solution to machine learning at the edge is in software, not in hardware. Leveraging smart compression and compilation on existing mobile devices can alleviate the traditional strain on machine learning on edge devices. Smart compression reduces the model to what is necessary to achieve the desired result, and smart compilation effectively matches that model to what the processor can actually support. This results in an efficient method of performing computational learning on peripheral devices with very little processing power.
A solution for today
In summary, machine learning on mobile devices is actually possible and works on almost all phones. This is not a solution for the years to come. It can work right now.
With smart compression and compilation, your mobile phone could perform the recommendation algorithm by itself. Your data would remain private on your device, but you could still get the useful information that machine learning can provide. This is the future of machine learning: your phone would become a real brain in your pocket!
Article translated from Forbes US – Author: Yanzhi Wang
To read also: The great novelty in network security