Evolutionary AI: Making AI Creative

Cognizant AI Labs conducts basic research and builds applications with modern AI techniques such as evolutionary computation and deep learning. We work across disciplines, both internally and in collaboration with other research institutions. Our work has resulted in a unique evolutionary computation platform called LEAF (Learning and Evolutionary AI Framework), and over 50 publications and 70 awarded or pending patents. The core team originated at Sentient Technologies and continues (since 2019) as Cognizant AI Labs. This site reviews the Evolutionary AI perspective, technology, and results, and also hosts a series of expert interviews and podcasts on the future of AI. Explore the site to find out how Evolutionary AI is making a difference.

Technologies

Evolutionary Computation

In contrast to deep learning, which focuses on modeling known behaviors, Evolutionary Computation creates new solutions, by repeatedly recombining, mutating, and adapting a population of them. Our recent research includes discovering decision-making strategies in design, business, and games, incentivizing AI to find creative and novel solutions, utilizing multiple objectives, and creating AI that can explain its decisions in terms of rules.

Deep Learning

Deep learning is another foundation for AI research with the Evolutionary AI team. The main focus of our recent research is on optimizing deep learning architectures for performance and computational cost, utilizing of multiple datasets through multitask learning, and discovering useful neural network building blocks.

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, hyperparameters, and weight parameters of neural networks. To put it another way, it is AI designing AI. Our recent work has focused on neural architecture search, improving the state of the art in several machine learning benchmarks.

Surrogate Optimization

The idea is to first build a model of the domain with e.g. Deep Learning, and then use the model as a surrogate to optimize the interactions with it using Evolutionary Computation. We have used this approach to optimize e.g. growth recipes for agriculture, behaviors for game agents, and non-pharmaceutical interventions in the COVID-19 pandemic. In this manner, it is possible to discover creative and effective decision strategies in a safe and efficient manner.

Metalearning

Modern deep learning models have many aspects that need to be fine-tuned by hand, a laborious process that requires special expertise. Metalearning is a family of techniques that allows for architectures, loss functions, hyperparameters, activation functions, and more to be automatically optimized, resulting in higher-performance models.

Trustworthy AI

In order to trust predictions and prescriptions of an AI system, it needs to indicate how confident it is, it needs to allow exploring alternative solutions, and in some cases, it needs to explain its behavior through explicit rules. The LEAF platform includes technologies developed specifically towards these three goals.

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 can now be scaled up to solve practical problems. Deep learning is one of those ideas; evolutionary computation is another. In particular, evolution extends the realm of AI from modeling and prediction to creativity and prescription. In that sense, we believe that evolutionary computation is the new deep learning: It is the next step in using AI in the real world.

Recent Research

All of our basic research and many of the applications are reported in scientific papers. Here are a few recent spotlights:

Research Paper

From Prediction to Prescription: Evolutionary Optimization of Non-Pharmaceutical Interventions in the COVID-19 Pandemic

Research Paper

Effective Reinforcement Learning through Evolutionary Surrogate-Assisted Prescription

Research Paper

Quantifying Point-Prediction Uncertainty in Neural Networks via Residual Estimation with an I/O Kernel

See All Research Publications

Research Applications

AI is more than theory. Even the most fundamental research is often motivated by practical problems, and it is usually possible to apply the results to the real world. We often work with experts in other disciplines to build such applications---in turn, such collaborations help us develop better AI methods. Here are a few recent application examples:

Media

AI is changing society and affecting many lives. It is important to reach out to public media to communicate why the research is valuable, what role of AI can play in the society in the future, and what the challenges and opportunities are. Here are links to selected recent such outreach:


See more media