Past Meetup

Data stream mining & Multi-task learning in Tensorflow and Keras

This Meetup is past

61 people went


GDG Reading and Thames Valley is part of the ML community and to kick off that relationship we have an ML special for you!

Two great speakers, the first a Doctoral student from the University of Reading talking about local research in data stream mining and the second speaker a researcher at Seldon, London. Seldon hosts the Tensorflow meetups in London and we are thrilled to have Giovanni over to speak to us.


18:30 doors open - food and mingling

19:00 Talk: Giovanni Vacanti from Seldon, London

20:00 Short drinks break

20:10 Talk: Thien Le, PhD student from University of Reading

21:00 Close

Computationally Efficient Rule-Based Classification for Continuous Streaming Data

Speaker: Thien Le (

Summary Data Stream Mining is the analysis of infinite and potentially fast flows of data that require real-time adaptive algorithms that are computationally efficient.

The field of Data Stream Mining necessitated through advances in software and hardware that allow generating and recording data streams such as stock market data, sensor networks, network traffic, etc. An important method of Data Stream Mining is the classification of previously unseen instances.

Traditional data mining algorithms need to be trained on already labelled data and are then applied on previously unseen and unlabelled data. Whereas traditional classifiers typically use several iterations over the same data this is not possible on data streams, as they are potentially very fast and infinite. Hence algorithms that only need one pass through the data are required.

This is an overview about a novel algorithm for classification for approaches in dealing with data in real-time streaming environment.

Multi-task learning for recommendations systems in tensorflow and keras

Speaker: Giovanni Vacanti (, Machine Learning Engineer @ Seldon (

Giovanni is a theoretical quantum physicist and data science expert with eight years experience in academic research. He received his Ph.D from the National University of Singapore, and has also carried out work in Italy, France, and the United Kingdom. Giovanni is highly skilled in mathematical modeling, numerical analysis of complex systems, analytic calculations, and problem solving in general.


Introduction: Machine learning concepts

Multi task learning:

• Learning many tasks,

• Curriculum training,

• The problem of catastrophic forgetting,

• How to address it,

• Something unseen: adiabatic training

Going deep into recommendation systems:

• What’s a recurrent neural network,

• Collaborative filtering and latent vectors,

• Content based recommendations,

• Putting them together



As we all know, cool kids play with machine learning nowadays.

Even cooler kids play with deep learning using Tensorflow and keras. Due to their versatility, these tools represent an extremely useful resource in many contexts and they are an ideal playground for development and experimentation.

In this talk, we focus on multi-task learning (MLT), introducing various techniques that allow a machine learning model to minimize different cost functions (each representing a different task) together.

In addition, we present the implementation in tensorflow and keras of an hybrid recommendation system based on a recent proposal ( Bansal et al., ( (2016) ) , which put together deep learning techniques (specifically recurrent neural networks), matrix factorization and multi-task learning.


This event is sponsored by Priocept. Check out