In the last few weeks, a few papers containing evolutionary techniques applied in the context of deep neural networks have been published. For someone with a background on evolutionary computing and interested in everything that is bio-inspired, these are great news! Recently we’ve seen: Evolving Deep Neural Networks , Genetic CNN, Large-Scale Evolution of Image Classifiers and PathNet: Evolution Channels Gradient Descent in Super Neural Networks.
These recent papers are not the first ones on the topic (and won’t be the last) since many different applications of evolutionary techniques to neural networks, including deep ones, have been published in the past. However, it “feels” that finally the field is catching up and paying attention to the very fast developments in neural networks. Especially when organizations like DeepMind and the Google Brain Team are investing in the topic.
The research and development of evolutionary techniques for deep nets is, in my opinion, very important. These methods have achieved many “human competitive results” and thus have the potential to present innovative solutions and at the same time, reduce human intervention in the process of design and optimization of a deep model. It can also be used to produce new insights that can latter be used to develop new methods, by looking and analyzing the proposed solutions. Some people may simply criticize these approaches by claiming it’s playing LEGOs, “forgetting” that humans are already playing it. Or for the amount of resources required. This is a valid point which just means that more understanding and development is required.
Since there are no free lunches, you need to understand when it makes sense to apply this type of methods as well as how to design them. Unfortunately, it’s very common to see direct applications of concepts that are already outdated. There’s no value in using the standard genetic algorithm from John Holland or the Koza-style genetic programming. Even though they are easy to apply and understand, they are outdated! Another example is not knowing how to analyze an evolutionary algorithm through their representation properties. This means that most likely you will have an inefficient approach.
Recently, the inspiration for most of the new advances in deep learning are coming from the math/game theory/etc., side. And not so much from the biological side, e.g., neuroscience. I would expect more coming from this area of inspiration (an attempt can be found here for example: Towards an integration of deep learning and neuroscience) since ultimately, the human brain is used as the main example of what kind of AI we want/would like to build. However, the brain is not a final product, and it was not designed in a single step. It’s the product of a long evolutionary process (which still goes on)! It means that we need to study and understand more these two so that we can effectively use them for the artificial variants. Deep Neuroevolution should be a path to pursue.
For the next days I will be attending PPSN 2010, in Krakow, Poland. I was here for the venue two years ago and it was a conference that I enjoyed very much. Mostly because of the model which is different from the standard ones (poster only format, absence of parallel sessions and a summary presentation of the papers by a senior session chair).
I will be talking about my lastest work which ties together two techniques that I’ve always wanted to do some serious work with them: Ant Systems and Genetic Programming. In this first paper, I am using GP to evolve an Ant System component. A key issue in AS research is how to design the communication mechanism between ants that allows them to effectively solve a problem. We propose in this work to evolve the current pheromone trail update methods. We tested with the TSP and initial results show that the evolved strategies perform well and exhibit a good generalization capability when applied to larger instances.
Doing this work was also very fun and it was all made in Lisp. I just need to improve the code a bit and package it nicely before making it free available.
Anyway, I wish PPSN 2010 will be an interesting conference!
Just got back from PPSN 2008 and I enjoyed it very much. It was my first time at this conference and I believe the poster only format, absence of parallel sessions and a summary presentation of the papers by the session chair works very well. You are given a quick overview of the papers, having a bit more of information on what you want to see and afterwards, the interactions between attendants and presenters is much stronger. Of course some posters are flooded with people whilst others not but that’s not important. I belive the conversations between researchers are always interesting.
Although there is a bias towards theory papers and Evolutionary Strategies, you have a good diversity of approaches and type of papers, from applications to new techniques for example. For me it is hard to pinpoint a best work from all that I’ve seen but I was impressed with Peter Merz‘s new approach to very large TSP problems.
Finally, I guess the organization and the venue of the conference were very good. No complaints here. I wish I can attend in 2010 too.
And the second paper related to the work done in INRIA is accepetd, at the 8th International Conference on Hybrid Intelligent Systems. The event will take place in Barcelona, Spain, from September 10th until the 12th. From their website:
“The objectives of HIS 2008 are: to increase the awareness of the research community of the broad spectrum of hybrid techniques, to bring together AI researchers from around the world to present their cutting-edge results, to discuss the current trends in HIS research, to develop a collective vision of future opportunities, to establish international collaborative opportunities, and as a result to advance the state of the art of the field.”
I got accepted my first paper related to the work done in INRIA, at the 10th International Conference on
Parallel Problem Solving From Nature. The event will take place in Dortmund, Germany, from September 13th until the 17th. The interesting fact about PPSN is the conference format. From the website:
“all accepted papers will be presented during small poster sessions of about 16 papers. Each session will contain papers from a wide variety of topics, and will begin by a plenary quick overview of all papers in that session by a major researcher in the field. Past experiences have shown that such presentation format led to more interactions between participants and to a deeper understanding of the papers.”
I agree that this might motivate the interaction between participants and thus, everyone will benefit much more from the event. I’m curious and I wish it will be a nice event since I’ve never been to PPSN before.
The 2nd Natural Computing Research and Applications Group workshop was a success and very interesting. I got the chance to see first hand the work developed at NCRA and I liked it. I found the work of Erik Hemberg on Meta-Grammars & Grammatical Evolution interesting, as well as Sébastien Piccand’s presentation on Optimisation of PSO topology. On a side note, the slides of my talk can be found in slideshare.