Modern mobile apps and energy consumption
How to put a neural network on a strict diet
The first thing an experienced engineer does is send the neural network to the ‘gym’ to shed all the excess weight. In the industry, this is known as quantisation and pruning, but in layman’s terms, it’s simply getting rid of the fat. Most mobile tasks don’t require mathematical precision down to the hundredth decimal place, so the model’s ‘weight’ is compressed, turning cumbersome calculations into simple operations. It’s a bit like packing your things into a vacuum bag before a trip: the volume is reduced by a factor of three, whilst the contents remain the same. There’s also a brilliant trick called distillation, where a huge ‘smart’ neural network trains a smaller version of itself to do the same job. The result is a compact algorithm that doesn’t make your phone ‘sweat’ every time you run it.
The first thing an experienced engineer does is send the neural network to the ‘gym’ to shed all the excess weight. In the industry, this is known as quantisation and pruning, but in layman’s terms, it’s simply getting rid of the fat. Most mobile tasks don’t require mathematical precision down to the hundredth decimal place, so the model’s ‘weight’ is compressed, turning cumbersome calculations into simple operations. It’s a bit like packing your things into a vacuum bag before a trip: the volume is reduced by a factor of three, whilst the contents remain the same. There’s also a brilliant trick called distillation, where a huge ‘smart’ neural network trains a smaller version of itself to do the same job. The result is a compact algorithm that doesn’t make your phone ‘sweat’ every time you run it.
Many people have come across this problem: you launch a cool new app on your phone with augmented reality features or instant voice translation, marvelling at the ‘magic’, only to realise fifteen minutes later that your phone has got hot and the battery is draining before your very eyes. This is the main irony of modern mobile development, according to experts at Magenta Favorita IT company. Users want smartphones to be incredibly smart and to process everything locally (i.e. on the device itself) for the sake of privacy, but in doing so they forget that complex neural networks are voracious creatures that love to ‘gorge’ on processor resources. If a developer simply takes a heavy server-side model and ports it to mobile code, they are signing the death warrant for the user experience. At Magenta Favorita Portugal, which specialises in mobile app development amongst other things, they emphasise that the real art today lies not in making a neural network work, but in making it imperceptible to the battery.
Secret weapons and the right tools
The days when the main processor had to do everything on its own are long gone, according to experts at Magenta Favorita. Nowadays, every decent smartphone contains a dedicated chip for neural computing, which uses tens of times less energy whilst performing the same tasks. Using the right frameworks, such as CoreML or TensorFlow Lite, allows the system to decide for itself where to send a task in order to conserve battery life. It’s like a smart home that switches off the lights in rooms where no one is present: if a neural network can process a task on an energy-efficient core in milliseconds, it shouldn’t need to bring out the ‘heavy artillery’ and cause the device to overheat.
The days when the main processor had to do everything on its own are long gone, according to experts at Magenta Favorita. Nowadays, every decent smartphone contains a dedicated chip for neural computing, which uses tens of times less energy whilst performing the same tasks. Using the right frameworks, such as CoreML or TensorFlow Lite, allows the system to decide for itself where to send a task in order to conserve battery life. It’s like a smart home that switches off the lights in rooms where no one is present: if a neural network can process a task on an energy-efficient core in milliseconds, it shouldn’t need to bring out the ‘heavy artillery’ and cause the device to overheat.
The balance between intelligence and common sense
Ultimately, an app’s success depends on the developer’s empathy towards the user. Even the smartest neural network in the world is useless if it makes people afraid to unlock their screen unnecessarily. Good software must know when to hold back: for example, by reducing the frame rate if the phone has overheated, or by disabling resource-intensive functions altogether when the battery is down to 20 per cent. We are moving towards an era when AI will become as much a basic standard as mobile internet once was, but this path requires meticulous fine-tuning, according to experts at Magenta Favorita IT company. It is important to remember that users are not buying the number of features in your model, but a user-friendly tool that won’t let them down in the middle of the working day. It is only through this balance of technology and careful attention to the ‘hardware’ that truly great products are born.
Ultimately, an app’s success depends on the developer’s empathy towards the user. Even the smartest neural network in the world is useless if it makes people afraid to unlock their screen unnecessarily. Good software must know when to hold back: for example, by reducing the frame rate if the phone has overheated, or by disabling resource-intensive functions altogether when the battery is down to 20 per cent. We are moving towards an era when AI will become as much a basic standard as mobile internet once was, but this path requires meticulous fine-tuning, according to experts at Magenta Favorita IT company. It is important to remember that users are not buying the number of features in your model, but a user-friendly tool that won’t let them down in the middle of the working day. It is only through this balance of technology and careful attention to the ‘hardware’ that truly great products are born.
Other company cases