WHAT WE DO

The Evolutionary AI team does basic research and builds applications of cutting-edge AI techniques. We work across disciplines, internally, and with a wide array of skilled research institutions and universities on the common goal of advancing the state of the art in artificial intelligence. With over 50 patents awarded or pending, we combine our unique mix of experienced AI practitioners with the scale of our massively distributed AI platform to change whole disciplines–and the world–for the better.

TECHNOLOGIES

Sentient Answering Critical Questions

EVOLUTIONARY COMPUTATION

Evolutionary computation (EC) mimics the principles of biology. In contrast to deep learning, which focuses on modeling and prediction of known solutions, EC allows AI to create new solutions, by recombining, mutating, adapting (i.e. breeding) better and better ideas. Some of our recent research includes work on novelty search, which incentivizes AI to find creative–or novel–solutions, and an online optimization product, Ascend by Evolv, which develops effective websites and other content for digital marketing.

DEEP LEARNING

Deep learning is another foundation for AI research with the Evolutionary AI team. The main focus of our recent research is on optimization of deep learning architectures and on utilization of multiple datasets through multitask learning. Some of our recent research focuses on massive, unlabelled datasets, such as deep learning applied to raw video.

NEUROEVOLUTION

Neuroevolution is a powerful way to combine evolution and deep learning: evolution is used to automatically optimize deep learning architectures, i.e. the topologies, components, and hyperparameters of neural networks. To put it another way, it is AI designing AI. In this manner, more complex and more powerful deep learning architectures can be discovered, and it is possible to discover them automatically, thus democratizing AI. Our recent work has focused on gated recurrent networks and on multitask networks, improving the state of the art in several machine learning benchmarks, including those in language and vision.

Sentient Answering Critical Questions

SURROGATE OPTIMIZATION

Surrogate optimization is another way to combine the power of deep learning and evolutionary computation. The idea is to learn a model of the domain, use the model as a surrogate to optimize interactions with it, and then apply those interactions to the domain. We have applied surrogate optimization e.g. to the problem of creating optimal growth recipes for plants in computer-controlled greenhouse environments (built by OpenAg at MIT Media Lab). In this manner, it is possible to discover effective recipes that take biologists by surprise.

OUR PHILOSOPHY

AI research is at an exciting stage. With a million-fold increase in computing power, many of the ideas developed over the last three decades now scale up to solve practical problems. Deep Learning is one of those; Evolutionary Computation is another. Evolution also extends the realm of AI from modeling and prediction to creativity and discovery. In that sense, we believe that evolution is the new deep learning, i.e. the next step in building complex deep learning systems, in commercialization of AI, and in solving hard problems.

LEARN MORE

RECENT RESEARCH

After we find the answers to the big questions we put pen to paper and write-up our findings in the form of academic research. Here you can find all the papers we have currently published. Here are a few spotlights:

PDF Research

Flavor-Cyber-Agriculture: Optimization of plant metabolites in an open-source control environment through surrogate modeling

Harper, C. B., Johnson, A. J., Meyerson, E., Savas, T. L., and Miikkulainen, R. (2019). Flavor-Cyber-Agriculture: Optimization of plant metabolites in an open-source control environment through surrogate modeling. PLOS ONE, https://doi.org/10.1371/journal.pone.0213918
PDF Research

A Comparison of the Taguchi Method and Evolutionary Optimization in Multivariate Testing

Jiang, J., Legrand, D., Severn, R., and Miikkulainen, R. (2018). A Comparison of the Taguchi Method and Evolutionary Optimization in Multivariate Testing. arXiv:1808.08347
PDF Research

Evolutionary Architecture Search for Deep Multitask Networks

Citation: Liang, J., Meyerson, E., and Miikkulainen, R. (in press). Evolutionary Architecture Search for Deep Multitask Networks. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2018, Kyoto,…

SEE ALL RESEARCH PUBLICATIONS

RESEARCH APPLICATIONS

AI should be more than just theory. We believe in not just advancing research but applying that research, our theories, and our network to solving complicated problems. We work both internally and externally, with world class institutions like MIT and Oxford, to find those answers. Here are a few of our projects.

MEDIA

We frequently reach out to public media about the research, with the goal of communicating why it is valuable and how it will change lives---as well as what the role of AI will be in the society in the future, and what the challenges and opportunities are. Below you will find a few links to such press interviews, videos, and podcasts: