More Gemini Features Coming to Google TV
Google TV integrates Gemini AI for image/video generation, enhances Google Photos, and adds a dedicated YouTube Shorts feed to the homepage.
Read moreNeural network-based machine learning models are the driving force behind many recent technological advancements, including highly accurate translation of text and autonomous driving cars. Researchers are also increasingly using them to help solve physical problems. Neural networks have been identified as a new phase of matter, encompassing everything from detecting interesting anomalies in high-energy physics experiment data to finding astronomical objects called gravitational lenses in graphs. However, despite the surge in results obtained through neural networks, the inner workings of this tool remain elusive, and it is often unclear how the network processes information to solve problems. Now, a team at the Swiss Federal Institute of Technology in Zurich (ETH) has demonstrated a method for finding this information. Scientists can use their method to view problems and approaches to solving them in a completely new way.
A neural network is a computational tool that roughly simulates the operations of the human brain. The network typically consists of multiple layers of interconnected artificial neurons, which perform calculations and weight the connections between neurons. These weights, ranging from millions to billions, form the adjustable parameters of the network. The advantage of neural networks is that they can solve tasks without being programmed. Instead, they learn how to adjust the parameters to improve the solutions they output over time. For example, to train a neural network to recognize faces, the network is given many different images of the same person. The network then learns to recognize the face, changing the weights of the connections until it produces reliable results.
While neural networks can learn to solve a wide range of tasks, the inner workings of these models are often a black box. One way to understand the knowledge learned by the network is to examine its weights. However, this is often challenging due to their large number. The lack of understanding of how neural networks operate is particularly unsatisfactory in physics: the tool can solve challenging problems, but we do not know the rules and principles it uses to generate solutions. This is where the new results from Raban Iten, Tony Metger, and their colleagues come in.
The team started with a standard neural network consisting of seven layers. They then modified the network in two key ways. First, they changed the middle layers of the network to have fewer neurons than the other layers, creating what is known as an information bottleneck. For example, in one case, they reduced the number of neurons in that layer from 100 to 2 (Figure 1). (In this case, both the input and output layers also had two neurons.) In other cases, the modified layers had more neurons, but the number was always less than 10. Secondly, they encoded the modified layers so that each neuron contained independent information. They used a technique called variational decoupling, which comes from the field of computer vision. Going back to the face recognition example, this modification means that one neuron in the altered layer might contain information about the shape of the mouth, while another might contain information about the size of the eyes. These two modifications together mean that the final neural network (SciNet) has several parameter layers, with each neuron containing information about independent properties of the problem being solved, making the network more amenable to study. It is through this layer that the team investigates the inner workings of the network.
To study the neural network, the research group asked SciNet to solve various physical problems, with astronomy being the most representative. For this problem, SciNet was given the angular coordinates of Mars and the Sun, which are measured from Earth relative to certain fixed stars. The neural network was then asked to predict the future positions of these celestial bodies. After the training process, the research group examined the weights of the two neurons in the fourth layer of the network to gain deeper insights into how SciNet solves the task. By analyzing the output of the intermediate layer neurons, they discovered that SciNet performs a coordinate transformation, altering the angle of Mars, so it appears to be measured from the Sun rather than from Earth.
The demonstration by the ETH team allowed them to understand how neural networks solve various specific tasks. But the results have more significant implications. By understanding the inner workings of neural networks, physicists can use them to gain new insights and conceptual understanding of problems, rather than just the final answers. For example, SciNet can be trained to predict measurement results of quantum systems to understand how the network relates the mathematical theory of quantum mechanics to reality. Another exciting opportunity with SciNet is studying the rotation curves of galaxies. Physicists still do not understand why stars in the Milky Way rotate faster than predicted by visible mass, leading to the hypothesis of the existence of unknown "dark matter" mass in the universe. Understanding whether SciNet solves this problem by adding hidden dark matter mass, modifying the laws of gravity, or using completely different representations that astrophysicists can explain would be fantastic. The answers from SciNet can help researchers identify new directions to solve this long-standing important problem.
This work takes a step towards using machine learning models as a source of scientific inspiration, helping researchers find new perspectives and enhance human creativity in understanding physical problems. Hopefully, in a few years, when these methods are better understood and applied to unsolved scientific problems, they will lead to new conceptual understandings and accelerate the progress of physics itself.
Google TV integrates Gemini AI for image/video generation, enhances Google Photos, and adds a dedicated YouTube Shorts feed to the homepage.
Read moreAlphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Read moreAnthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Read moreApple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Read moreGoogle launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
Read moreYouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Read moreGoogle grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Read moreMistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.
Read moreCustom subliminal audio maker

AI Social Media Manager

AI Search Visibility

Workout and Calorie Tracker
Stock research platform

AI interview assistant

Scientific figure generator

AI courses with real code