Uber, the American Multinational ride-hailing company recently announced they are going to ease up their AI operations and 3000 people will be laid off from the company. The AI team of Uber was working since last couple of years. Let’s look at some of their works;
AI generating Algorithm proposed by Uber has three pillars;
- Meta-learning architectures.
- Meta-learning the algorithms themselves.
- Generating effective learning environments.
Studies were extensively conducted in the first two pillars, but little has been done in the third one. AI- GA approach could lead to general AI and are worthwhile discoveries, irrespective of their speeds.
Uber AI specialists discovered new techniques to efficiently evolve deep neural networks. Uber discovered that an extremely simple genetic algorithm can train deep convolutional networks. It outperformed evolution strategies (ES) and modern deep reinforcement learning algorithms (RL).
This work on Deep neuroevolution by Uber offered an alternative approach in machine learning. The researchers and specialists introduced new algorithms, combined the optimization power and evolution strategies with the methods that are unique to neuroevolution. These methods promote new exploration in reinforcement learning domains through a number of agents act differently from one another.
POET – Paired Open-Ended Trailblazer
Researchers at Uber developed an algorithm, capable of tackling problems in the two- dimensional landscapes. It pairs the generation of new environmental challenges and optimize the agents to solve this challenge. POET explore through many different paths, through varieties of problems and solutions allow the solutions to transfer between the problems. Open-ended in POET indicates the potential for algorithms to create novel and increasingly complex capabilities. POET can create a diverse range of sophisticated behaviours and solve a wide range of environmental challenges, many of which cannot be solved by direct optimization alone. The ability of POET to transfer solutions from one environment to another demonstrates the unpredictable nature to create new stepping stones.
PLATO Research Dialogue System was introduced by Uber in 2019 and it was integrated with deep learning techniques. The platform helps in building, training and deploying conversational AI agents to conduct the state of the art research in AI and to create prototypes and demonstration systems. This helps in facilitating conversational data collection. It was integrated with Bayesian optimization frameworks and reduces the need to write code.