Learned optimizers are parametric algorithms that can themselves be trained to solve optimization problems. In contrast to baseline optimizers (such as momentum or Adam) that use simple update rules derived from theoretical principles, learned optimizers use flexible, high-dimensional, nonlinear parameterizations. Although this can lead to better performance, their inner workings remain a mystery. How is a given learned optimizer able to outperform a well tuned baseline? Has it learned a sophisticated combination of existing optimization techniques, or is it implementing completely new behavior? In this work, we address these questions by careful analysis and visualization of learned optimizers. We study learned optimizers trained from scratch on four disparate tasks, and discover that they have learned interpretable behavior, including: momentum, gradient clipping, learning rate schedules, and new forms of learning rate adaptation. Moreover, we show how dynamics and mechanisms inside of learned optimizers orchestrate these computations. Our results help elucidate the previously murky understanding of how learned optimizers work, and establish tools for interpreting future learned optimizers.

Learned optimizers are parametric algorithms that can themselves be trained to solve optimization problems. In contrast to baseline optimizers (such as momentum or Adam) that use simple update rules derived from theoretical principles, learned optimizers use flexible, high-dimensional, nonlinear parameterizations. Although this can lead to better performance, their inner workings remain a mystery. How is a given learned optimizer able to outperform a well tuned baseline? Has it learned a sophisticated combination of existing optimization techniques, or is it implementing completely new behavior?

Generative adversarial nets (GAN) has been successfully introduced for generating text to alleviate the exposure bias. However, discriminators in these models only evaluate the entire sequence, which causes feedback sparsity and mode collapse. To tackle these problems, we propose a novel mechanism. It first segments the entire sequence into several sub-sequences. Then these sub-sequences, together with the entire sequence, are evaluated individually by the discriminator. At last these feedback signals are all used to guide the learning of GAN. This mechanism learns the generation of both the entire sequence and the sub-sequences simultaneously. Learning to generate sub-sequences is easy and is helpful in generating an entire sequence. It is easy to improve the existing GAN-based models with this mechanism. We rebuild three previous well-designed models with our mechanism, and the experimental results on benchmark data show these models are improved significantly, the best one outperforms the state-of-the-art model.\footnote[1]{All code and data are available at https://github.com/liyzcj/seggan.git

It's not that often that I can steal the title of a paper and use it for a blog article that people will actually read, but I think "Popcorn-Driven Robotic Actuators" totally works, so credit for that to Steven Ceron at Cornell University who's the first author on this paper, presented at the IEEE International Conference on Robotics and Automation in May. Let's see what else I can steal from it: Popcorn kernels are a natural, edible, and inexpensive material that has the potential to rapidly expand with high force upon application of heat. Although this transition is irreversible, it carries potential for several robotic applications. As kernels can change from regular to (larger) irregular shapes, we examine the change in inter-granular friction and propose their use as granular fluids in jamming actuators, without the need for a vacuum pump. Furthermore, as a proof-of-concept, we also demonstrate the use of popcorn-driven actuation in soft, compliant, and rigidlink grippers.

Natural human-like human-robot interactions require many functional capabilities from a robot that have to be reflected in architectural components in the robotic control architecture.  In particular, various mechanisms for producing social behaviors , goal-oriented cognition , and robust intelligence are required.  In this paper, we present an overview of the most recent version of our DIARC architecture and show how several novel algorithms attempt to address these three areas, leading to more natural interactions with humans, while also extending the overall capability of the integrated system.