Build Your Machine Learning Model in Minutes with AutoML
By Haoran Lyu and Xingyu Chen
We may notice that more and more Internet companies are applying machine learning in their products. AI seems to be a powerful tool to solve many different problems. When talking about machine learning and artificial intelligence, what will come to mind? Arcane theoretical knowledge? Tedious mathematical proofs? What if we want to use such a powerful tool with little or even no background? AutoML gives such a chance that even someone has no background in machine learning, a model can be trained in several minutes.
The problem that the tool addresses
Google AutoML provides a powerful machine learning tool for the users even if the users do not have much knowledge of machine learning. With the help of this tool, the users will be able to create their own custom machine learning model which can be implemented in the business field. Also, it can be integrated into the user’s own applications and websites.
There are several components for different tasks.
In the AutoML Natural Language part, users can classify content, extract the entity and do sentiment analysis using the tool.
In the AutoML Table component, users can build and deploy state-of-the-art machine learning models on structured data at massively increased speed and scale. And it can help create clean, effective training data by providing information about missing data, correlation, cardinality, and distribution for each of the features. Moreover, since there is no charge for importing the data and viewing the information, there will be no cost until starting training the model.
In AutoML Translation, users can create their own, custom translation models so that translation queries return results specific to the domain. And users do not need a custom model solution. APIs are provided for over 100 languages.
In the AutoML Vision Classification and Intelligence Classification part, users can train the machine learning model with the help of the components to classify shots, segments in the videos, and images according to the labels defined by the user.
Recently, those have been embedded into a new platform called AI Platform. In our movie dataset, we use AutoML Table to have a quick tour.
A quick tour of AutoML Table with movie rating data
The following snapshots are a quick tour. Since the AutoML has been embedded into the AI platform, we tried both two ways. They seem similar to each other.
First, we need to create a new dataset and import our data.
After importing data, we can see some analyzed metadata about our dataset. For example, missing rate, type of the data, etc.
The next step is training. Auto ML has already defined how to train the model. All we need to do is setting some hyperparameters like how much cost we expect to spend on training, which columns are needed. AutoML will finish the remaining steps for us.
We can also use customized training steps, which require us to provide detailed training steps. For our training data, we may hope to use collaborative filtering. Thus we can provide our own training files. This step is prepared for specialists and can only be found in the AI platform, rather than AutoML. So we did not give a further attempt.
In the movie streaming scenario, we provided user’s ratings towards different movies, expecting the model to give us a suitable result. However, after trying, we found that the tool is not suitable for our training dataset. Basically, what we need to do is collaborative filtering, however, AutoML Table can only provide certain classification and regression models, which exclude CF. In other words, if we have proper data, like user profile and movie profile, AutoML Table will become a powerful tool.
But still, it is so convenient. With the above steps, a machine learning model is just built. Choosing those parameters only took us several minutes and the model just started training. For those who are not major in computer science or related subjects, this codeless tool is suitable. The only thing that needs to do is to prepare data.
The strength and limitations of the tool
Strength
After a quick tour, the strength of the tool is obvious. It is really convenient. Once the users provide proper data, AutoML can do everything left for them, including training the model, evaluating the model, and deploying it. Users do not need to have a strong background and knowledge about machine learning. It will make machine learning more suitable for everyone.
From the official website, we can see many different companies have built successful machine learning models and use them in production. Those tasks used to require very professional data scientists. However, now they just need AutoML.
Limitation
It can only be used in some specific cases mentioned above. In other cases, since the component is not provided by Google, users will not be able to work on it. Since Google prepared “everything” for users, this tool may not be suitable for those technics, who may want to have more operations to adjust.
Also, the target of this product is those who have little or no background knowledge about machine learning. Then after training, AutoML does not provide an explicit evaluation of the model. Although it provides lots of curves and numbers, without a baseline, its users are still hard to tell whether the model is good or not. In our quick tour, the model basically randomly guesses some numbers for the result given a user id and movie id. However, some results seem to be good, like a pretty low false-positive rate, which is meaningless. With a recall of about 20%, this model is useless. But for those with little experience in machine learning, the results may be not too explicit to tell.
Conclusion
AutoML is a powerful tool for starters of machine learning. Quick and convenient are the most obvious strengths. With prepared data, users can set up training a complex model within five minutes. Also, no code is required. Only some basic ideas of machine learning can help developers to build a model. This is great for some startups or primer individuals. However, for more advanced experience, AutoML may not be a good tool. If users want to apply a more specified training process, the AI platform also provided by Google may be a good choice.
