Jonas Degrave

Journal: PLXtrum - realtime machine learning for predicting note onset

Sat, 28 Jan 2017 18:00:00 +0000

This system detects in real time when the note is coming by using a piezo and accelerometer on the plectrum. In this demonstration, we play noise between the prediction and the actual onset of the note to show you how far ahead we knew the next note was coming. The controller was optimized to predict 50ms ahead. The prediction is done with reservoir computing for the sensor fusion (also known as echo state networks).

In order to have everything running in realtime on the Axoloti board, the weight matrices had to be kept sparse, posing additional constraints on the system. After all it had to predict an onset happening in 50ms on a microcontroller. The computations to come to that conclusion therefore need to be fast as well, otherwise you will need to predict even sooner as you lose time computing. Evenmore, everything had to be done on the 168MHz STM32F427 microcontroller.

The Axoloti board

Paper available here: http://dx.doi.org/10.1080/09298215.2017.1288747

Patent pending

Competition: 489th on Facebook hackercup

Sat, 21 Jan 2017 23:00:00 +0000

Facebook’s hackercup just ended. It is the first time I competed, and I made it to through the qualifications and round 1 into round 2. I unfortunately got stuck on the second question with an algorithm which was not fast enough. I still ended in the top 500, so I got a facebook hackercup T-shirt!

Research: Backpropagation through camera

Mon, 16 Jan 2017 00:00:00 +0000

I finally implemented the backpropagation through a camera.

We have optimized a controller for the pendulum-cart system, by backpropagating through the physics, vision and controller. By treating the whole system as a differentiable function, we could update the controller parameters, and have it solve this problem. In only 2420 update steps (or trials in the simulator), it managed to completely learn the vision and the control task, and solve the pendulum-cart system. I hope the paper for ICLR will be accepted soon, such that I can publish the code and see what other people will do with it.

Next step: evaluating this method on the real setup!

Control the pendulum-cart setup with only visual perception. Optimized deep controller with backpropagation through time, camera and physics pic.twitter.com/6gu7GHpQBv
— 317070 (@317070) January 16, 2017

Research: poster on Deep Reinforcement Learning Workshop at NIPS

Fri, 09 Dec 2016 00:00:00 +0000

The poster presented at NIPS

Speaker: Critical Utopias - Cultural Discourse in the Digital Age

Sat, 03 Dec 2016 00:00:00 +0000

Artificial intelligence might replace taxi-drivers in the future, but how will the future of cultural journalists look like in light of future technological revolutions?

The Finnish Cultural Institute for the Benelux invited me to give a talk in the design musem in Helsinki on the topic of a future of art and art critics, in an age of rapid evolution of technology.

Unfortunately, the video failed, so there is nothing to show on this blog.

Paper: A Differentiable Physics Engine for Deep Learning in Robotics

Thu, 03 Nov 2016 00:00:00 +0000

Abstract:

One of the most important fields in robotics is the optimization of controllers. Currently, robots are often treated as a black box in this optimization process, which is the reason why derivative-free optimization methods such as evolutionary algorithms or reinforcement learning are omnipresent. When gradient-based methods are used, models are kept small or rely on finite difference approximations for the Jacobian. This method quickly grows expensive with increasing numbers of parameters, such as found in deep learning. We propose an implementation of a modern physics engine, which can differentiate control parameters. This engine is implemented for both CPU and GPU. Firstly, this paper shows how such an engine speeds up the optimization process, even for small problems. Furthermore, it explains why this is an alternative approach to deep Q-learning, for using deep learning in robotics. Finally, we argue that this is a big step for deep learning in robotics, as it opens up new possibilities to optimize robots, both in hardware and software.

You may find the paper here where it has been submitted for ICLR 2017.

Gait found by gradient descent through physics. Only 500 model evaluations required, starting from standstill. https://t.co/oAuH0vOjCB pic.twitter.com/ld9iovX31P
— 317070 (@317070) November 4, 2016

Award: Best poster award at Benelearn conference

Tue, 13 Sep 2016 00:00:00 +0000

The poster presented at Benelearn

The award

Activism: Pirate Party polls at 1.8%

Fri, 09 Sep 2016 00:00:00 +0000

Polling results according to the RTBF

I have been the co-president of Pirate Party Belgium for a couple of years, and have been been on lists for elections twice. So I am following them closely and am glad that they are holding on. It is still a long way to go for the 5% required for a first seat.

Speaker: the 3rd Data Science Ghent Meetup - Jonas Degrave

Thu, 16 Jun 2016 00:00:00 +0000

I explained how to use deep learning effectively, what is needed in order to get your models to work. These are the steps, which I often find missing by a lot of people.

A video of the presentation underneath:

Edit: I received this mail afterwards.

Hoi Jonas,

Ik wou je graag nog even bedanken voor de inspirerende presentatie van vandaag tijdens de data science meetup. Het was veruit de beste presentatie die ik me kan herinneren en ik heb enkele concrete ideeën opgedaan om toe te passen in de Kaggle competitie waar ik momenteel deel aan neem. Na 5 weken sta ik er trouwens eerste en mocht dat zo blijven is dat grotendeels te danken aan jouw schitterende vertoning, bedankt!

Groeten,

Tom

Edit2: Apparently, Tom of the mail above came second in another Kaggle competition and is working at Google Deepmind now. Congratulations!

Activism: Camping on roof of an arms factory

Fri, 20 May 2016 00:00:00 +0000