Insights

About the author(s)

About the writer

Kenny Chan is studying Information Engineering in The City University of Hong Kong who is interested in Data Management and Data Analysis.

MI_banner.jpeg

<h3>The good stuff</h3><p>On-device Machine Learning has recently become a hot topic in tech. In comparison to its cloud based counterpart,&nbsp; On-device Machine Learning has made the development of applications significantly easier, while offering a number of key advantages.</p><p>Firstly, on-device Machine Learning is simply faster. As uploading information such as images, audio, and text to the cloud or server is not required, there's no necessity to wait for a response from Google. The real impact of this comes when using features like pose or face detection, as a shorter delay translates to a better user experience.</p><p>Another stand out advantage is the lack of need for programmers to construct remote calls to access Google’s cloud or servers, as they would need to do with cloud-based ML. When all the execution processes are performed within the device, Machine Learning can be used anywhere. No network coverage doesn’t mean Machine Learning processes can’t be executed. Moreover, a Machine Learning model can be applied to different projects after building the model.</p><p>Information security is another bonus to on-device Machine Learning. All that private data that you definitely do not want to fall into the wrong hands is kept on your device. The risk of your confidential information leaking is reduced.</p><h3>The limitations</h3><p>There are certain drawbacks to choosing On-device Machine Learning over its cloud-based cousin. The first of which is its longer processing time, although this can be influenced by the performance capabilities of the device itself. Limited storage and power consumption can increase the processing time.</p><p>Accuracy is also a factor to consider. If a model is bundled into the application for example, its size needs to be limited due to the app stores policies. Also, as in mobile devices, the network usage also needs to be managed. Models, therefore, models cannot be too large, less data for training the model results in lower accuracy.</p><h3>Professional tools</h3><p>Google ML Kit is a powerful and easy-to-use package that includes a lot of pre-trained sounds, information, and face detection Machine Learning models. The Google ML Kit can be used in mobile, web, Android and IOS development. You can also consider TensorFlow Hub also provides custom models if you require a more targeted model.</p><h3>How Google ML Kit solves real-life problems</h3><p>In a nutshell, having repetitive work done by machines frees time and energy up for more meaningful work. Take a highway Surveyor tasked with counting the number of vehicles passing through a particular section of highway. Using object detection (included in the Google ML kit) frees up their time for other tasks.</p><p><img src="https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/MI_banner_53413c61bb.jpeg" alt="MI_banner.jpeg" srcset="https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/thumbnail_MI_banner_53413c61bb.jpeg 245w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/small_MI_banner_53413c61bb.jpeg 500w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/medium_MI_banner_53413c61bb.jpeg 750w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/large_MI_banner_53413c61bb.jpeg 1000w," sizes="100vw"></p><p style="text-align:center;">Figure 1. Demonstration of counting vehicles with the original model.</p><p>Take figure 1 for example, a car is detected and counted after being captured by the camera. Machine Learning helps the camera to recognize the object captured as a ‘car’ thus distinguishing the object from being classified as another class of object, like an animal or a running human!</p><p><img src="https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/banner_af1a0bdc6f.jpg" alt="banner.jpg" srcset="https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/thumbnail_banner_af1a0bdc6f.jpg 245w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/small_banner_af1a0bdc6f.jpg 500w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/medium_banner_af1a0bdc6f.jpg 750w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/large_banner_af1a0bdc6f.jpg 1000w," sizes="100vw"></p><p style="text-align:center;">Figure 2. Limitation of Object detection with the original model.</p><p>In figure 2, although the camera captured and the object was counted ‘successfully’, it recognized the vehicle as a “fashion good” instead of being a ‘vehicle’. Limitations can occur if only the default object detection model is used for Machine Learning processing, although it can be solved by using custom models.</p><p><img src="https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/157a4eb0_81e0dd0737.jpg" alt="157a4eb0.jpg" srcset="https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/thumbnail_157a4eb0_81e0dd0737.jpg 245w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/small_157a4eb0_81e0dd0737.jpg 500w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/medium_157a4eb0_81e0dd0737.jpg 750w," sizes="100vw"></p><p style="text-align:center;">Figure 3. Home page of TensorFlow Hub.</p><p>There are some custom models on the TensorFlow Hub. The advanced object detection model can be applied in your projects which will help the camera to recognize the vehicle correctly.</p><h3>How to train new models (if you cannot find a suitable model online)</h3><p><img src="https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/fdb95cdc_51add77655.jpg" alt="fdb95cdc.jpg" srcset="https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/thumbnail_fdb95cdc_51add77655.jpg 245w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/small_fdb95cdc_51add77655.jpg 500w,https://gt-terraform-web-s3-prod.s3.ap-east-1.amazonaws.com/uploads/medium_fdb95cdc_51add77655.jpg 750w," sizes="100vw"></p><p style="text-align:center;">Figure 4. Results after auto Machine Learning.</p><p>Google vision is an option to carry out automated Machine Learning model training. To do so, a large number of images are needed to be prepared and uploaded to the model. In figure 3, the details of the model, such as its precision, will be shown immediately after the training process. However, more research may be needed to be completed in order to optimize the model’s performance and accuracy.</p><h3>Conclusion</h3><p>This is just an example of using some of the Google ML kit features that help solve simple problems. More complicated scenarios can be constructed by mix-and-matching different models and pre-built functions in the Google ML kit. So have a try and have some fun while doing it!</p>

On-device Machine Learning has recently become a hot topic in tech. In comparison to its cloud based counterpart,  On-device Machine Learning has made the development of applications significantly easier, while offering a number of key advantages. 

Wingy has the merit of being both a dedicated technical expert and an approachable manager. He is expert in various fields due to his passionate on keep learning. Also, he is always willing to give much possible support to his teammates to explore their strong point in different projects.

On-device Machine Learning: Tools

Cloud in Fintech: Impact of Cloud Services on Fintech Innovation

Exploring the Key Features and Benefits of Google Cloud Platform

What is a Cloud Service?