Machine Learning (Without CODE)

Machine learning is very fascinating for data science practitioners and everyone and there's a continuous effort different companies are putting in to make it available for mass to use.

Related to this we spoke about "AutoML" in one of my previous articles https://www.dhirubhai.net/pulse/automl-first-glance-raja-saurabh-tiwari/, which talks about writing lesser code for achieving machine learning goals. Let's discuss this further today.

There's been continuous effort in the direction to make the machine learning and AI capabilities available for Life science, medical and agriculture sectors. And if you really want to make it 'general' and 'approachable' thing you need to lessen the dependency on data science expertise or coding background as much as possible.

Further to making it easy for non-coders/experts to use, Microsoft has recently released a tool called Lobe. It's beta version is released. Lobe was acquired by Microsoft couple of years back. This was in-lined with their vision to make another step towards AI adaption.

Microsoft says Lobe uses

“open-source machine learning architectures and transfer learning to train custom machine learning models on the user's own machine”

Lobe gives you image classification model in few minutes with NO CODE at all.

From Lobe's -

Lobe has everything you need to bring your machine learning ideas to life. Just show it examples of what you want it to learn, and it automatically trains a custom machine learning model that can be shipped in your app."

I thought to try this out myself. Trust me it was fun :)

Being a bird photography aspirant I thought its best to use my photographs to classify whether they are birds or not. This is what I did,

1.Download Lobe desktop app from https://lobe.ai/

2. Once successfully installed open it up. It's has very simple and intuitive UI. It just needs 5 images to train the model.

3. Import images : import images from the right top corner buttons.

4. "Label" the image : You can label the images 1 by 1 or first import all of them and then label it

Statistics about images/labels will start updating on the left side of the screen.

5. Import some images for the other class. I'm keeping other class where there's no bird. I call it "No Object".

6. Repeat #5 and #6 for as my images you want. Ideally you should keep a balance between the classes, otherwise you'll face "class imbalance" problem.

You'll notice that on the background the "Train" is running. This means that on the background the model is getting trained.

7. Once you are done uploading images you can hit "Train" model. This is an explicit call to train the model. This should not take more than few minutes.

8. Once training is done, hit the "Play" button. Now it's like testing your model which you just created. Add new images and see if your model is able to correctly classify your image. As you upload image, it'll try to guess it, label it and also it'll show statistics on the left.

9. If its you not correctly predicted the class, you can always change/update the label. This way you are telling model that "hey you did not do well here, correct this".

The app will note that down and improve it’s model to predict better. You can do this kind of test and correct the label few times and then you are just done.

10. Now you can export your model with all the model details with TensorFlow, TensorFlow light and few more options.

There are other ways to use the tool by clicking images on go. An illustration can be watched on Youtube - Introducing Lobe | Build your first machine learning model in ten minutes.

This is a fantastic tool for the people who want to do something with AI and Machine Learning but not an expert in data science or programming. Also this gives access to AI like things to common people in easy to use way.

You can use it for variety of things for example detection of a person wearing mask or not. If a person sleeping or not etc.

Microsoft has released beta version of the tool , which is right now limited to image classification problems. It'll be interesting to see how it progresses and how it removes dependency on complexity of data science and brings it to more common. 

Raja Saurabh Tiwari