IMC: Difference between revisions
| Line 36: | Line 36: | ||
We are entering a new era in computing, where the services provided are becoming more and more dependent on machine learning and artificial intelligence, with deep learning at the forefront of this new gold-rush. The introduction of deep learning architectures were only possible due to the use of GPUs, and they have done a great job training more and more complex models. However, FPGAs offer huge benefits over GPUs in terms of power savings, which is crucial when web-scale applications are considered. There are many exciting developments in this field, but unfortunately, contrary to GPUs, there are close to no public tooling available for FPGAs. We would like to develop a code generator, which takes as input a commonly used neural-network definition file, and spits out fpga code. This project could potentially help fuel the next generation systems for which all of us interact with daily. | We are entering a new era in computing, where the services provided are becoming more and more dependent on machine learning and artificial intelligence, with deep learning at the forefront of this new gold-rush. The introduction of deep learning architectures were only possible due to the use of GPUs, and they have done a great job training more and more complex models. However, FPGAs offer huge benefits over GPUs in terms of power savings, which is crucial when web-scale applications are considered. There are many exciting developments in this field, but unfortunately, contrary to GPUs, there are close to no public tooling available for FPGAs. We would like to develop a code generator, which takes as input a commonly used neural-network definition file, and spits out fpga code. This project could potentially help fuel the next generation systems for which all of us interact with daily. | ||
Feedback: | |||
I've discussed with a couple of the staff who teach our hardware course, and | |||
they confirm that the dev boards the students use do have FPGA that might be | |||
used here. | |||
I wasn't suggesting online learning - rather that they might see how fast | |||
they can train a simple image classifier. | |||
Our research team have recently implemented something along these lines. | |||
Here are some relevant papers: | |||
Neural nets in custom hardware: | |||
http://www.cl.cam.ac.uk/~atm26/papers/fccm2012-bluehive.pdf | |||
Custom hardware v vector processing | |||
http://www.cl.cam.ac.uk/~atm26/pubs/FPL2013-BlueVec.pdf | |||
Video and background | |||
http://www.cl.cam.ac.uk/research/comparch/research/bimpa.html | |||
An undergrad group could perhaps work with their code, rather than build | |||
something from scratch? | |||
==2016 projects== | ==2016 projects== | ||
Revision as of 07:33, 26 October 2016
Contact: Liselot de Jonge <liselot.dejonge@imc.nl>
2017 proposals
Confirmed:
More ideas:
Tamagotchi Brief
Jan Kis : jan.kis@imc.com
(‘Creative’) Pokemon Go took the world by storm. Now it’s your turn! Remember those cute, egg-shaped devices with the dancing critters you all played with when you were younger? Well, this task will involve the exciting opportunity of creating an interactive online Tamagotchi world! Your aim is to entice users to explore the fascinating concept of optimal parameter selection, critical to many real world problems such as Machine Learning and automated trading, through their Tamagotchi’s. Develop a mobile application allowing users to view their Tamagotchi and exchange limited resources with each other by agreeing a fair value for their exchange. Once acquired, users should be able to modify the appearance of their Tamagotchi using these resources. Users will thus need to carefully optimize their basket of resources to build their ultimate Tamagotchi. They want to recreate that incredible wig they saw at last night’s bop but only have two bundles of cloth? Well, they trade six spindles of thread for two bundles of cloth, apply some suspect sowing skills and ta dah!
(Formal) Optimal selection of parameters is an important aspect of many exciting real world problems from Machine Learning to automated trading in world markets. This project aims to get users to explore this concept. The task is to build an interactive online mobile application to enable users to view and build their ultimate Tamagotchi. Users will need to exchange limited resources with each other by coming to an agreement about a fair value. These resources can then be used to alter the appearance of their Tamagotchi. Consequently, users will need to carefully select their optimal basket of resources.
24000 words per second
Ben Catterall : ben.catterall@imc.com
Many fields in machine learning, from image recognition to machine translation have recently received a tremendous boost using deep learning. Deep architectures have sparked a renewed interest in artificial intelligence, and resulted in a lot of cool applications. It has also arrived together with a new wave of peer-reviewed research, where people share and publish all of their code online. Most big companies race to provide their pre-trained models online for free. In this project we will focus on automatic video captioning, and the aim is to build a prototype system for a real-time captioning system, using already published research. The resulting product could be used by visually impaired people, or to create automatic tags on instagram.
Feedback:
Quite a lot of our students enjoy playing around with pre-trained deep neural nets, so this is certainly feasible. However, it seems at present like a one-person project, and would have to be expanded to suit work from a team. We already have some projects planned for this year that involve training a deep net, but this involves some ingenuity in identifying suitably labelled training data, as well as getting hold of a machine suitable for running the GPU-intensive deep learning frameworks that are currently popular.
Do you want to provide a reference to the specific piece of published research that you thought might be applied, and we can see where we might go from there?
Neural Networks in FPGA
Taylan.Toygarlar@imc.com
We are entering a new era in computing, where the services provided are becoming more and more dependent on machine learning and artificial intelligence, with deep learning at the forefront of this new gold-rush. The introduction of deep learning architectures were only possible due to the use of GPUs, and they have done a great job training more and more complex models. However, FPGAs offer huge benefits over GPUs in terms of power savings, which is crucial when web-scale applications are considered. There are many exciting developments in this field, but unfortunately, contrary to GPUs, there are close to no public tooling available for FPGAs. We would like to develop a code generator, which takes as input a commonly used neural-network definition file, and spits out fpga code. This project could potentially help fuel the next generation systems for which all of us interact with daily.
Feedback:
I've discussed with a couple of the staff who teach our hardware course, and they confirm that the dev boards the students use do have FPGA that might be used here.
I wasn't suggesting online learning - rather that they might see how fast they can train a simple image classifier.
Our research team have recently implemented something along these lines. Here are some relevant papers:
Neural nets in custom hardware: http://www.cl.cam.ac.uk/~atm26/papers/fccm2012-bluehive.pdf
Custom hardware v vector processing http://www.cl.cam.ac.uk/~atm26/pubs/FPL2013-BlueVec.pdf
Video and background http://www.cl.cam.ac.uk/research/comparch/research/bimpa.html
An undergrad group could perhaps work with their code, rather than build something from scratch?
2016 projects
earlier ideas
Maksym Korotkiy Maksym Korotkiy <Maksym.Korotkiy@imc.nl>
Prototype a 2D visualization for an execution of a genetic algorithm (GA) applied to a multi-dimensional search problem. The visualization should provide insights into all stages of GA (selection, crossover, mutation) as well as into evolution of candidate solutions. We can assume that number of dimensions is between 10 and 50, number of candidate solutions (population size) is 500 and number of generations is around 100. The visualization should make it easy to understand internal workings of GA and to show an impact of different selection and crossover strategies or mutation rates. Students can use any general GA implementation and can apply it to any multi-dimensional search problem.
Taylan Toygarlar <Taylan.Toygarlar@imc.nl>
Radmilo Racic <radmilo.racic@imc.nl>
Visualization techniques for large set of financial markets data This project develops techniques for visualizing multiple data sets of financial data, including ticker states of global futures and significant stocks, bonds and currencies. The end goal is not only to unearth hidden relationships and correlations between global markets but also to convey trader sentiment and pin point market moving trades. As global market landscape is quite complex and correlated, we will be using Oculus Rift as the principal display tool. Students will be provided with data from Eurostoxx, DAX, CAC, KOSPI, Nikkei, ES, EUR/USD, T-Note, GBL, etc.
