Telephone: 0114 257 2200

Three Astounding Deep Learning Projects

Deep learning is now a fundamental part of research and teaching at many universities and organisations around the world. Yet as with all kinds of academia, those fastest to findings take the limelight – as well as the next round of research grants.

At the forefront of deep learning is the relatively new IBM Power System S822LC with NVLink technology, codenamed ‘Minsky’ and referred to as such in this post, which was created by IBM and NVIDIA.

As a partner of IBM and NVIDIA via the OpenPOWER Foundation, OCF was able to supply two ‘Minsky’ systems to one of our key clients Queen Mary University of London (QMUL).

In this post, we open up ‘Minsky’ and highlight some of QMUL’s standout projects. We think you’ll agree that they’re fascinating!

Minsky combines POWER8, Tesla P100, and NVLINK

Much of the technology inside Minsky originates from the OpenPOWER Foundation, which is dedicated to innovating new processor, system and software design based on the POWER architecture. IBM, NVIDIA and OCF are all members.

Two key products of the Foundation’s work have been:

  • The IBM POWER8 processor – The first OpenPOWER processor, offering cutting-edge performance for modern workloads like big data and analytics
  • NVIDIA NVLINK – A high-speed CPU-GPU interconnect, enabling 5-12x faster data sharing than PCI Express

These two innovations led to the creation of an incredibly powerful new system for deep learning applications: IBM Power System S822LC with NVLink technology.

Minsky packs two POWER8 processors and four NVIDIA Tesla P100 GPUs, which is a lot of processing power. But more importantly, all those CPUs and GPUs are integrated by super-fast NVLINK inter-connects.

So, while x86 systems suffer from a major bottleneck on CPU-GPU data sharing, Minsky’s CPUs and GPUs work together in parallel.

That makes Minsky a deep learning beast.

Deep learning research at QMUL

But what can these impressive-sounding Minsky systems do in real-world terms?

OCF was already aware of QMUL researchers’ love of novel technology that provides a competitive advantage. The university was the first in Europe to deploy NVIDIA DGX-1 systems, described as “the world’s first AI supercomputer in a box”.

OCF was therefore delighted to also supply QMUL with two ‘Minsky’ systems: one for research, and another for teaching. This made QMUL one of the first universities in Britain to use these powerful deep learning machines.

Results on the research side have been surprising, exciting, and potentially game-changing – as these three examples show.

1. SketchX drawing recognition could revolutionise shopping

First is SketchX, a world-leading research lab on human sketch analytics. SketchX has used deep learning to analyse drawings of objects like clothes and animals, and recognise what they depict. The project has been covered by New Scientist, Digital Trends, and

SketchX is ‘part aware’. This means it can even understand specific ‘fine-grained’ parts of the drawing, such as the height of a shoe heel and whether the shoe has laces.

One potential real-world application of SketchX is in online shopping and search. Instead of finding a product with traditional filters and keywords, you’ll soon be able to simply draw it on your touchscreen!

Try a live SketchX demo

2. Warblr classifies birds by their song, and could help protect endangered species

Co-created by QMUL research fellow Dan Stowell, Warblr is a birdsong recognition app you can download on your Android or iOS device right now. Think Shazam, but for birdsong.

Just record a British bird’s song with your device and Warblr will identify the species with up to 95% accuracy. Its accuracy is greater than other tools because Warblr uses large-scale, unsupervised, automated feature learning from data.

Warblr is also a citizen science project, which makes its recordings and data freely available to research and conservation projects.

You can read more about it at BBC, The Guardian, and Wired.

3. folk-rnn uses neural networks to compose folk music

Finally, QMUL’s Dr Bob Sturm has used Minsky to create a generative model of folk-style music, “by training a recurrent neural network (RNN) with three hidden layers of long short-term memory (LSTM) units”. In other words, an automated folk music composer built on deep learning.

Listen to one of over 3,000 pieces of music that ‘folk-rnn’ has composed so far:

[Embed YouTube example:]

You can also try the Python code over at the folk-rnn Github page.

What could ‘Minsky’ bring to your organisation?

As you can see, ‘Minsky’ or Power System S822LC is enabling truly innovative research projects at QMUL, with compelling outcomes that have made a worldwide impact.

If you’d like to know more about the system and what it can offer your organisation, I’d be happy to have an informal discussion. You can email me directly at, or leave a comment below.