DeepMind has released Gato, the most general machine learning (ML) model. If you're familiar with arguments about the possible dangers of robust AI systems, you'll see that the term "generic" has significant implications. 

Today's ML systems are rapidly improving; nonetheless, even the finest systems we've seen are limited in the jobs they can perform. DALL-E, for example, makes visuals that rival human ingenuity; unfortunately, it accomplishes nothing else. Similarly, substantial language models like GPT-3 excel at some text-based tasks, such as sentence completion, but struggle with others, such as arithmetic.

If future AI systems are as intelligent as humans, they will need to use different skills and pieces of information to complete various tasks in different situations. In other words, they need general intelligence as humans do. This type of system is called "artificial general intelligence" (AGI). AGI systems could lead to many great new ideas, but they could also become more intelligent than humans and "superintelligent." If researchers didn't align a superintelligent system, it could be hard or even impossible to control and predict its behaviour, leaving humans vulnerable.

Image source: Deepmind

Gato is a single neural network that can do many different things. DeepMind says it can "play Atari, caption images, chat, stack blocks with a real robot arm, and do much more, deciding based on its context whether to output text, joint torques, button presses, or other tokens." It doesn't have intelligence as humans do yet, but it does have some general abilities.

How did they build Gato?

Gato's training methodology differs slightly from those of other well-known AI agents. For example, the AI system AlphaGo, which beat the world champion Go player Lee Sedol in 2016, was trained mainly by a form of trial and error called reinforcement learning (RL). Expert Go players showed AlphaGo how to play the game when researchers first trained it, but in the next version, AlphaGo Zero. AlphaGo Zero learned how to play games only by playing them itself.

Evaluation

Gato was on various control tasks by calculating the mean of 50 results. Researchers compared these averages to the results of trained and fine-tuned expert agents for each given control task. It is essential to note that Gato was also on language, vision, and robotics data, represented in the model. 

The following photos demonstrate how the pre-trained Gato model with the same weights can perform numerous tasks, including image captioning, interactive discussion, and robot arm movement.

Image source: DeepMind

Analysis

A critical part of intelligence is learning new things quickly by using what you already know and have done before. With this in mind, DeepMind hypothesized that "it is possible to train an agent that is generally good at a large number of tasks, and that this general agent can be changed with a small amount of extra data to be good at a greater number of tasks."

DeepMind used a trained Gato model and a limited collection of demos from new tasks that weren't in its training set to evaluate this. Then, they compared Gato's performance to a "blank slate" model that had been randomly set up and trained only on these demonstrations. If the new activities are similar to Gato's previous tasks, faster learning can occur. For example, a Gato model trained on continuous control tasks learned more quickly on new control tasks than a model trained on text and graphics.

Conclusion

Gato was on three different model sizes, the biggest of which was small compared to the most recent advanced models. It seems likely that bigger versions of Gato will work much better than we've described here. There are limits, though. Scaling alone wouldn't be enough for Gato to perform better than experts on a wide range of tasks since it is to copy the experts rather than try new things and act in new ways. Moreover, it's still unclear how hard it will be to train generalist agents like Gato, who can do a better job than specialized ones.

Image source: Unsplash