Author: Yury Kh

In the span of the last 10 years, the term “neural networks” has gone beyond the scientific and professional environment. The theory of neural network organization emerged in the middle of the last century, but only by 2012 the computer power has reached sufficient values to train neural networks. Thanks to this their widespread use began. 

Neural networks are increasingly being used in mobile application development. The Deloitte report indicates that more than 60% of the applications installed by adults in developed countries use neural networks. According to statistics, Android has been ahead of its competitors in popularity for several years.

Neural networks are used:

• to recognize and process voices (modern voice assistants), 
• to recognize and process objects (computer vision), 
• to recognize and process natural languages (natural language processing),
• to find malicious programs, 
• to automate apps and make them more efficient. For example, there are healthcare applications that detect diabetic retinopathy by analyzing retinal scans.

What are neural networks and how do they work?

Mankind has adopted the idea of neural networks from nature. Scientists took the animal and human nervous systems as an example. A natural neuron consists of a nucleus, dendrites, and an axon. The axon transitions into several branches (dendrites), forming synapses (connections) with other neuronal dendrites.

The artificial neuron has a similar structure. It consists of a nucleus (processing unit), several dendrites (similar to inputs), and one axon (similar to outputs), as shown in the following picture:

Connections of several neurons form layers, and connections of layers form a neural network. There are three main types of neurons: input (receives information), hidden (processes information), and output (presents results of calculations). Take a look at the picture.

Neurons on different levels are connected through synapses. During the passage through a synapse, the signal can either strengthen or weaken. The parameter of a synapse is a weight – some coefficient can be any real number, due to which the information can change. Numbers (signals) are input, then they are multiplied by weights (each signal has its own weight) and summed. The activation function calculates the output signal and sends it to the output (see the picture).

Imagine the situation: you have touched a hot iron. Depending on the signal that comes from your finger through the nerve endings to the brain, it will make a decision: to pass the signal on through the neural connections to pull your finger away, or not to pass the signal if the iron is cold and you can leave the finger on it. The mathematical analog of the activation function has the same purpose. The activation function allows signals to pass or fail to pass from neuron to neuron depending on the information they pass. If the information is important, the function passes it through, and if the information is little or unreliable, the activation function does not allow it to pass on.

How to prepare neural networks for usage?

Work with neural nets goes through several stages:

1. Preparation of a neural network, which includes the choice of architecture (how neurons are organized), topology (the structure of their location relative to each other and the outside world), the learning algorithm, etc. 
2. Loading the input data into a neural network.
3. Training a neural network. This is a very important stage, without which the neural network is useless. This is where all the magic happens: along with the input data volume fed in, the neuronet receives information about the expected result. The result obtained in the output layer of the neural network is compared with the expected one. If they do not coincide, the neural network determines which neurons affected the final value to a greater extent and adjusts weights on connections with these neurons (so-called error backpropagation algorithm). This is a very simplified explanation. We suggest reading this article to dive deeper into neural network training. Neural network training is a very resource-intensive process, so it is not done on smartphones. The training time depends on the task, architecture, and input data volume. 
4. Checking training adequacy. A network does not always learn exactly what its creator wanted it to learn. There was a case where the network was trained to recognize images of tanks from photos. But since all the tanks were on the same background, the neural network learned to recognize this type of background, not the tanks. The quality of neural network training must be tested on examples that were not involved in its training. 
5. Using a neural network – developers integrate the trained model into the application.

Limitations of neural networks on mobile devices

RAM limitations 

Most mid-range and low-end mobile devices available on the market have between 2 and 4 GB of RAM. And usually, 1/3 of this capacity is reserved by the operating system. The system can “kill” applications with neural networks as they run when the RAM limit approaches.

The size of the application

Complex deep neural networks often weigh several gigabytes. When integrating a neural network into mobile software there is some compression, but it is still not enough to work comfortably. The main recommendation for the developers is to minimize the size of the application as much as possible on any platform to improve the UX.

Runtime

Simple neural networks often return results almost instantly and are suitable for real-time applications. However, deep neural networks can take dozens of seconds to process a single set of input data. Modern mobile processors are not yet as powerful as server processors, so processing results on a mobile device can take several hours.

To develop a mobile app with neural networks, you first need to create and train a neural network on a server or PC, and then implement it in the mobile app using off-the-shelf frameworks.

Working with a single app on multiple devices

As an example, a facial recognition app is installed on the user’s phone and tablet. It won’t be able to transfer data to other devices, so neural network training will happen separately on each of them.

Overview of neural network development libraries for Android

TensorFlow

TensorFlow is an open-source library from Google that creates and trains deep neural networks. With this library, we store a neural network and use it in an application.

The library can train and run deep neural networks to classify handwritten numbers, recognize images, embed words, and process natural languages. It works on Ubuntu, macOS, Android, iOS, and Windows. 

To make learning TensorFlow easier, the development team has produced additional tutorials and improved getting started guides. Some enthusiasts have created their own TensorFlow tutorials (including InfoWorld). You can read several books on TensorFlow or take online courses. 

We mobile developers should take a look at TensorFlow Lite, a lightweight TensorFlow solution for mobile and embedded devices. It allows you to do machine learning inference on the device (but not training) with low latency and small binary size. TensorFlow Lite also supports hardware acceleration using the Android neural network API. TensorFlow Lite models are compact enough to run on mobile devices and can be used offline.

TensorFlow Lite runs fairly small neural network models on Android and iOS devices, even if they are disabled. 

The basic idea behind TensorFlow Lite is to train a TensorFlow model and convert it to the TensorFlow Lite format. The converted file can then be used in a mobile app.

TensorFlow Lite consists of two main components:

• TensorFlow Lite interpreter – runs specially optimized models on cell phones, embedded Linux devices, and microcontrollers.
• TensorFlow Lite converter – converts TensorFlow models into an efficient form for usage by the interpreter, and can make optimizations to improve performance and binary file size.

TensorFlow Lite is designed to simplify machine learning on mobile devices themselves instead of sending data back and forth from the server. For developers, machine learning on the device offers the following benefits:

• response time: the request is not sent to the server, but is processed on the device
• privacy: the data does not leave the device
• Internet connection is not required
• the device consumes less energy because it does not send requests to the server

Firebase ML Kit

TensorFlow Lite makes it easier to implement and use neural networks in applications. However, developing and training models still requires a lot of time and effort. To make life easier for developers, the Firebase ML Kit library was created.

The library uses already trained deep neural networks in applications with minimal code. Most of the models offered are available both locally and on Google Cloud. Developers can use models for computer vision (character recognition, barcode scanning, object detection). The library is quite popular. For example, it is used in:

• Yandex.Money (a Russian e-commerce system) to recognize QR codes;
• FitNow, a fitness application that recognizes texts from food labels for calorie counting;
• TutboTax, a payment application that recognizes document barcodes.

ML Kit also has:

• language detection of written text;
• translation of texts on the device;
• smart message response (generating a reply sentence based on the entire conversation).

In addition to methods out of the box, there is support for custom models.

What’s important is that you don’t need to use any services, APIs, or backend for this. Everything can be done directly on the device – no user traffic is loaded and developers don’t need to handle errors in case there is no internet connection. Moreover, it works faster on the device. The downside is the increased power consumption.

Developers don’t need to publish the app every time after updates, as ML Kit will dynamically update the model when it goes online.

The ML Kit team decided to invest in model compression. They are experimenting with a feature that allows you to upload a full TensorFlow model along with training data and get a compressed TensorFlow Lite model in return. Developers are looking for partners to try out the technology and get feedback from them. If you’re interested, sign up here.

Since this library is available through Firebase, you can also take advantage of other services on that platform. For example, Remote Config and A/B testing make it possible to experiment with multiple user models. If you already have a trained neural network loaded into your application, you can add another one without republishing it to switch between them or use two at once for the sake of experimentation – the user won’t notice.

Problems of using neural networks in mobile development

Developing Android apps that use neural networks is still a challenge for mobile developers. Training neural networks can take weeks or months since the input information can consist of millions of elements. Such a serious workload is still out of reach for many smartphones. 

Check to see if you can’t avoid having a neural network in a mobile app if:

• there are no specialists in your company who are familiar with neural networks;
• your task is quite non-trivial, and to solve it you need to develop your own model, i.e. you cannot use ready-made solutions from Google, because this will take a lot of time;
• the customer needs a quick result – training neural networks can take a very long time;
• the application will be used on devices with an old version of Android (below 9). Such devices do not have enough power.

Conclusion

Neural networks became popular a few years ago, and more and more companies are using this technology in their applications. Mobile devices impose their own limitations on neural network operation. If you decide to use them, the best choice would be a ready-made solution from Google (ML Kit) or the development and implementation of your own neural network with TensorFlow Lite.