As part of our purpose, we are working on creating AI tools with a broader understanding of the world. This can allow you to transfer useful knowledge between many different types of tasks.
Using the learning of reinforcement, our AI Alphazero and Muzero systems have achieved superhuman games. Since then, we have expanded their capabilities to help design better computer systems, along with optimizing data centers and video compression. And our specialized version of Alphazero, called Alphadev, also discovered new algorithms accelerating the software on the basics of our digital society.
Early results showed the transformation potential of more general use of AI tools. Here we explain how these progress shape the future of calculations – and already helps billions of people and planet.
Designing better computer systems
Specialized equipment is necessary to make sure that today's AI systems are effective for large -scale users. But the design and production of new computer systems may take up years of work.
Our researchers have developed an approach based on artificial intelligence to design stronger and more efficient circuits. Treating a circumference like a neural network, we found a way to speed up systems design and transfer performance to new heights.
Neuron networks are often designed to take users' input data and generate outputs such as images, text or video. In the neural network, the edges are connected with nodes in a structure similar to the chart.
To create a circuit design, our team has proposed a circuit neural networks, a new type of neural network that turns the edges into cables and nodes into logic gates and learns how to connect them.
Animated illustration of the neuron network of the circuit of the circuit design. Specifies which edges (wires) are connected to which nodes (logical gates) to improve the overall circuit structure.
We optimized the learned circuit in terms of calculation speed, energy efficiency and size, while maintaining its functionality. Using the “simulated annealing”, a classic search technique, which looks a step into the future, we also tested various options to find its optimal configuration.
We won thanks to this technique IWLS 2023 programming competition – With the best solution for 82% of problems with the design of the competition in the competition.
Our team also used Alphazero, which can look many steps into the future to improve the circuit design, treating a challenge as a game to solve.
So far, our research connecting the neuronal networks of the circuit with the function of reinforcement learning prize has shown very promising results of building even more advanced computer systems.
Optimization of resources of the data center
Data centers manage all, from providing search results to processing data sets. Like a game of multidimensional tetris, a system called Credit Manages and optimizes loads in extensive Google data centers.
To plan your tasks, Borg is based on hand -coded rules. But on Google scale hand -coded rules cannot include the diversity of constantly changing load distribution. So they are designed as one size to match everyone.
At this point, machine learning technologies such as Alphazero are particularly helpful: they are able to work on a scale, automatically creating individual rules that are optimally adapted to different load distribution.
During the training, Alphazero learned to recognize patterns in tasks entering data centers, and also learned to predict the best ways to manage abilities and make decisions with the best long -term results.
When we used Alphazero to Borg in experimental research, we found that we can reduce the percentage of unused equipment in the data center by up to 19%.
Animated visualization of a neat, optimized data storage, compared to disordered and unaffected storage.
Effective video compressing
Video streaming is the majority of internet traffic. So finding ways to increase streaming efficiency, although large or small, will have a huge impact on millions of people watching movies every day.
We worked with YouTube to compress and transmit video by means of Muzero problems solving. By reducing temperature transmission by 4%, Muzero strengthened the overall experience of YouTube – Without prejudice to visual quality.
Initially, we used Muzero to optimize the compression of each individual video frame. We have now expanded this work to help make decisions in the grouping and cancellation of frames during coding, which leads to more savings.
The results of the first two steps show a great promise of Muzero's potential to become a more generalized tool, helping to find optimal solutions in the entire video compression process.
Visualization showing how Muzero compresses video files. Defines photos of photos with visual similarities of compression. A single key cage is squeezed. Then Muzero squeezes other frames, using the key frame as a reference. The process is repeated for the rest of the video until the compression is completed.
Discovering faster algorithms
AlphadevThe Alphazero version made an innovative breakthrough in computer science when she discovered faster sorting algorithms and hash. These fundamental processes are used by trillions times a day for sorting, storing and downloading data.
Alphadeva sorting algorithms
Sorting algorithms help in processing and displaying digital information, from the ranking of online search results and social posts to user recommendations.
Alphadev has discovered an algorithm that increases the efficiency in sorting short sequences of elements by 70% and about 1.7% for sequences containing over 250,000 elements, compared to algorithms in the C ++ library. This means that the results generated from the user's inquiry can be sorted much faster. Using on a scale it saves huge amounts of time and energy.
Alphadev's Hashing Algorithm
Mixing algorithms are often used to store and download data, as in the customer database. They usually use the key (e.g. “Jane Doe”) to generate a unique shortcut, which corresponds to the data values that require recovery (e.g. “order number 164335-87”).
Like the librarian who uses the classification system to quickly find a specific book, with a mixing system, the computer already knows what he is looking for and where to find it. After applying in the range of 9-16 bytes of mixing function in data centers, the Alphadeva algorithm improved the efficiency by 30%.
The impact of these algorithms
We added sorting algorithms to LLVM standard C ++ library -APPLICATION OF TRAWS, which have been used for over a decade. And contributed to alphadeva mixing algorithms to Abeliac Library.
Since then, millions of programmers and companies began to use them in various industries as diverse as cloud processing, online shopping and supply chain management.
General tools to feed our digital future
Our AI tools are already saving billions of people and energy. This is just the beginning. We imagine a future in which AI general tools can help optimize the global computer ecosystem.
We are not there yet – we still need faster, more efficient and balanced digital infrastructure.
You need many other theoretical and technological breakthroughs to create fully generalized AI tools. But the potential of these tools – between technology, science and medicine – we enjoy what is on the horizon.
