DevConf.IN '18

The fabric8-analytics platform has a use case of showing companion recommendations for a users' input stack (manifest) to help the user make better choices around which packages he should/should not include in the stack.
Earlier this use case was solved using a classical machine learning approach that worked well for the problem size at the time, however, wasn't suited for large-scale ecosystems like NPM. In order to scale our recommendations, I made use of recent advancements in the area of deep learning for recommender systems to apply a state of the art model for serving companion recommendations.
In this session, I'll cover the end to end picture of bringing this machine learning model to production,  all the way from data collection to training the model to serve recommendations.

Determining the number of servers required to process the hits for a particular Adobe Analytics is well calculated. It depends on the what type (class) of the server is used, a number of the demons servers has run and the role of the server. It also depends on the type of the processing needed; first layer of processing or second layer of processing.A first layer of processing only server can process up to x number of hits per day.It has been seen that  if the hardware is allocated based on this calculation then hardware remains underutilized for most of the day as the traffic is distributed. This paper talks about using the Recurrent Neural Network based model for hourly traffic predictions. A customer’s data has been used for the experiment; results have been found to accurate with 5% accurate

This is controversial but key to successfully adopt machine learning. Let us run away from solo focus on speech or image recognition. Let’s start focussing on cases where both cost and error of human involvement is very high.
In 2017, while working on building an image recognition system, we quickly understood that going from a top-5 classification error rate of 4% to 3% need a lot of input control measure. If those input control measures are removed the machines perform much worse than humans.
The big 5 AI giants have already identified these as research problem and are working on it. They have data, money, time and people. So for the majority of us, this is not what will drive adoption of ML products. To know more join me
Outcome
Insights  on smart effort investment in ML

Determining the number of servers required to process the hits for a particular Adobe Analytics is well calculated. It depends on the what type (class) of the server is used, a number of the demons servers has run and the role of the server. It also depends on the type of the processing needed; the first layer of processing or second layer of processing. A first layer of processing the only server can process up to x number of hits per day. It has been seen that if the hardware is allocated based on this calculation then hardware remains underutilized for most of the day as the traffic is distributed. This paper talks about using the Recurrent Neural Network based model for hourly traffic predictions. A customer’s data has been used for the experiment; results have been found to accurate with 5% accurate

In this session, I will be talking about how we can use Machine Learning to have anomaly detection on OpenStack logs, which can be further used to automate actions and helpful in root cause analysis. I will be using an ELK stack to consolidate data from OpenStack into one central machine to be further performed Machine Learning on it.