Image recognition is not only the earliest breakthrough in deep learning but also one of the most successful applications. However, as the application deepens, traditional deep learning methods are gradually running out of steam, even causing some industry giants to declare the arrival of a new AI winter. However, recently AI researchers have found inspiration from the laws of physics and have found a way for machines to recognize high-dimensional objects. Furthermore, physicists have discovered that this progress in AI can even make a greater contribution to physics research.
Computers (AI) can now drive, play chess and board games better than world champions, and even write articles. The AI revolution is largely due to the power of a special type of artificial neural network, whose design inspiration comes from the neural connections in the mammalian visual cortex. It has been proven that these "convolutional neural networks" have amazing two-dimensional pattern learning capabilities - this AI is particularly good at computer vision tasks, such as recognizing objects in handwritten text and digital images.
However, this powerful machine learning structure does not perform well when applied to datasets without built-in planar geometry. Around 2016, a new discipline called geometric deep learning emerged, with the goal of freeing CNNs from two-dimensional planes.
Researchers have now proposed a new theoretical framework for building neural networks that can learn any geometric surface pattern. The "equivariantly equivariant convolutional neural network" developed by Taco Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling from the University of Amsterdam and the Qualcomm AI Research Center can not only detect patterns in 2D pixel arrays but also patterns in spherical and asymmetric curved surfaces.
In order to break through the two-dimensional space in deep learning, researchers' solution is closely related to physics. Just like Einstein's general theory of relativity and the standard model of particle physics, the physical theories that describe the world exhibit a so-called "equivariant" property. This means that the quantities and their relationships in this world do not depend on any arbitrary reference frame; they are consistent regardless of whether the observer is moving or stationary, and regardless of how far apart the numbers on the ruler are.
Researchers and their collaborators know that in order to break through Euclidean planes, they need to rethink a basic computational method that allows neural networks to efficiently recognize 2D image patterns. This method is called "convolution", where each layer of the neural network performs mathematical operations on a small piece of input data and then passes the results to the next layer.
Performing convolution calculations on surfaces (geometrically called manifolds) is like holding a small translucent grid paper on the earth and trying to accurately draw the coastline of Greenland. It is impossible to cover Greenland with a square piece of paper without wrinkling it, which means that when you flatten the paper again, the coastline you draw will be distorted. However, using a Mercator projection will also cause distortion. Alternatively, you can simply place the coordinate paper on a flat world map instead of the earth, and then copy these deformations, for example, the entire upper edge of the map actually represents only one point on the earth (the North Pole).
Moreover, if the manifold is not a neat sphere but a more complex or irregular object like a 3D shape or a folded protein, performing convolution calculations on it becomes even more difficult. Changing the properties of the sliding filter in this way can allow CNNs to better "understand" certain geometric relationships.
At the same time, Taco Cohen and his colleagues in Amsterdam started to solve the same problem from the opposite direction. In 2015, when Cohen was still a graduate student, he was not interested in how to free deep learning from planes. Instead, he was interested in an engineering problem that he considered very practical: data efficiency, or in other words, how to reduce the number of samples required for training neural networks.
Researchers say, "Deep learning methods can be considered as very slow learners."
If training CNNs is used to recognize cats, it is almost not a problem. But if you want the network to detect more important content, such as cancerous nodules in lung tissue images, it is not easy to find enough training data (which needs to be accurately and appropriately labeled in medicine without privacy issues). The fewer samples required to train the network, the better.
Researchers know that one of the methods to improve the data efficiency of neural networks is to make specific assumptions about the data in advance - for example, a lung tumor in an image remains a lung tumor even after rotation or mirroring. Typically, convolutional networks must be trained with many samples of the same pattern in different directions to learn this information from scratch.
The road for computers to break free from two dimensions is long, and the current research is just a milestone.
(Source: 36Kr)
Alphabet 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.
Anthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Apple 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.
Google 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.
YouTube 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.
Google 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.
Mistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.