Hi! I’m Nozaki, a Data Scientist intern here at Hacarus. I had the pleasure of attending the latest “Machine Learning Meetup KANSAI” on June 17, 2019 and would like to report on the meeting.
The “Machine Learning Meetup KANSAI” is a reoccurring event in Kansai where students and professionals meet to share ideas and information with each other on the topic of Machine Learning. This past week’s meetup, the fifth in order, was held at Hatena’s awesome Kyoto offices.
The following is a brief overview of an interesting night:
Mackerel’s In-roll Anomaly Detection Design and Operation
Yoshida Yasuhisa (Hatena Co., Ltd.)
The first speaker and returning contributor was Mr. Yoshida from Hatena Co., Ltd. whom presented on the architecture and algorithm function of “in-Role Error Detection”, as part of
Hatena’s server management and monitoring service “Mackerel”.
Until now, when the CPU usage rate exceeds 90%, alerts are sent. However, using machine learning detecting this process can be automated – which reduces server management costs – a key part of Hatena’s services.
Mackerel performs anomaly detection for every role in their service. To meet industry demands, the requirement for anomaly detection is that the Machine Learning system must be lightweight so that it can be resource efficient and learn at a low cost. Mixed Gaussian distribution is an ideal model selected to meet these requirements.
One side effect to operating with a straight forward implementation is when the variance of data at the time of learning is small, a false detection can potentially occur. This would incur an alert with a slight change at the time of operation. Therefore, it is essential that the prior distribution of parameters be incorporated into the model based on human knowledge to overcome this issue.
Mr. Yoshida also explained about a device capable of reducing false positives. He stated it is used to improve the system by manually annotating whether or not the alert that actually occurred as a performance evaluation was appropriate or not.
ML system development / Operation and Basic Design in a Cross-Business Organization
Mr. Shotaro Tanaka (Livesense Inc.)
The second speaker at the meetup was Mr. Tanaka from Livesense Inc. Although Livesense operates multiple web services, the Machine Learning system seems to be independent as a cross organization. Mr. Tanaka’s presentation touched on how to develop and operate the Machine Learning base.
In the past, all Machine Learning systems have been incorporated independently into each service but as several of these implementations are similar, and operations and development benefits have been lost due to different deployments. This was the trigger for changing things around into a Machine Learning base where commonality could be common between systems, and common job management and deployment flow.
By creating a single docker image for each component, such as the recommendation algorithm container and pre-processing container, it is possible to reuse algorithm implementations that had existed separately up to now between systems. Also, the ability to develop individually for each component has lowered the barrier to entry for new members.
He also mentioned that Google Kubernetes Engine and Argo Workflow were introduced to create a more flexible and convenient base.
How to proceed with the Machine Learning process
Kitora Naoki (Hacarus Inc.)
The main session was concluded with a presentation from Hacarus Inc. on the topic of “How to proceed with the Machine Learning process.” Our own CDO, Naoki Kitora, presented on the topic and provided step-by-step instructions on how Hacarus is proceeding with machine learning as well as insightful details on what we are doing at each phase of the project.
Mr. Kitora began with an explanation on the importance of “Understanding the business” phase. This concept emphasizes knowing what problems need to be solved, the purpose of doing what you want to do, understanding the purpose and background of the project, and how much the business is impacted. He then discussed the “Data Understanding” phase and the significance of looking at data. In essence, this is being able to look at as much data as possible to develop an intuition about the data and to potentially catch human errors in its annotation. Lastly, he explained about each phase of data preparation, performance evaluation, actual operation, and what should be taken care of at each phase. Even those who usually only analyze data and evaluate performance was able to appreciate Mr. Kitora’s presentation and felt the importance of understanding and being aware of the Machine Learning process as a whole.
Once the main session was concluded, the after party began with a large number of pizzas and drinks being served (courtesy of Hacarus!). During the party, a new social activity was introduced. People who had questions related to the content of the main session wrote their questions on a Post-it note and placed it up on the whiteboard. I could feel the enthusiasm from the participants as I saw them one by one placing their questions up on the whiteboard. After all the Post-it notes were on the white board, “Human Power Clustering” took place. This is where everyone divided up, found a question of most interest to them and discussed the topic with their newly formed group.
About halfway through the after party, we began the customary Lightning Talk session. This time around, I participated in the Lightning Talk and spoke on the time that I tried to detect anomalies with SST. After my Lightning Talk, I felt a great deal of satisfaction in getting the opportunity to talk with other participants of the MLM-Kansai. There were two additional Lightning Talk presentations both containing very interesting information. The speakers captured the ears of the audience and delivered a few entertaining moments that caused a good laugh.
This was my first time participating in MLM-Kansai. I really enjoyed listening to the speakers and partaking in the social gathering following the presentations. I was able to talk with engineers from various companies and took away new knowledge in how machine learning is developed and operated as well as how it is used in the making of company products. It was an overall great experience and it was very nice to participate.
I think that MLM-Kansai is growing steadily and will continue going forward. If you are interested in learning more about Machine Learning and how it is serving industries today, I recommend becoming a member of MLM-Kansai and participating in the meetups. In becoming a member, you will receive notifications of the next event and other details regarding MLM-Kansai. Please do not miss the opportunity to join and pass up some great experiences. Follow the link here, https://mlm-kansai.connpass.com to MLM-Kansai and become a member today!