A head-on collision of two objects moving at a high rate of speed can be a spectacular event to watch, particularly if you see it in slow motion. The debris field of thousands of pieces flying in every direction is mesmerizing. Now imagine watching two objects collide at nearly the speed of light, can you imagine what that debris field would look like? At the European Union Laboratory CERN, scientists are doing just that at the Large Hadron Collider (LHC). Although, capturing the debris field in slow motion is extremely difficult because of how fast it appears and then vanishes. This is where different types of machine learning could help us understand these collisions and potentially provide us insight into new dimensions.
CERN scientists are sending beams of high-energy particles in opposite directions at nearly the speed of light and then making them collide. The debris fields of these collisions are helping them better understand everything in our physical world to the structure of the Universe. Most recently they are using these tests to prove the existence of dimensions outside our known four (3 spatial and 1 temporal).
How we may discover a new dimension
What has perked the interest of many scientists throughout the ages is why gravity is so much weaker than all of the other forces in nature. Gravity is not just a little weaker, the magnitude is staggering at 1040 times weaker than electromagnetic force. To illustrate gravity’s relative weakness, you can pick up a pile of paper clips using just a small bar magnet. This mind-blowing difference has scientist thinking that perhaps gravity is not as weak as we think and it is actually seeping through to another dimension, so in a sense we’re being fooled and simply not seeing the full picture.
Event including two high-energy photons whose energy (depicted by red towers). Image courtesy CERN.
Getting back to the debris field that CERN scientists are watching, an example of how there could be clues of the existence of another dimension is how the debris field behaves. Some scientists believe that at an extremely high resolution an abundance of a specific theoretical particle referred to as a “graviton” could show up in the debris field. As CERN explains, it is believed that gravitons are associated with gravity in the same way as the photon is associated with the electromagnetic force. If gravitons appear in the collision they should behave much like any other particle with a balanced rate or momentum of expansion and energy. Now if they don’t behave the same in the expanding debris field there could be an obvious empty zone. Imagine you snapped a picture of the Amtrak train colliding through the snow after one of the recent snowstorms here in New York and you saw a strange void in the expansion of the snow debris. If this were to happen after a collision in the LHC, it is suggested that this would be strong evidence that the gravitons were seeping through to another dimension.
The huge data problem
One of the biggest problems CERN encounters during and after these controlled collision events is the magnitude of data that must be analyzed. The following is a highly simplified and abbreviated explanation on how big the data collection is for any given experiment at CERN. As the collisions are happening a 60-megapixel camera is taking 40 million pictures per second. There are about 600 million events/collisions per second, each event equating to 1 megabyte of data. As the data is being captured an algorithm is filtering down to only 100 or 200 events of interest per second. All of this raw data is recorded onto servers at CERN. The final data is equivalent to around 1.5 CDs per second.
How machine learning is being used today
My belief is that scientist could make better use of the rapidly improving machine learning methods to identify, analyze, and classify particles faster. One of the most common ways to train an algorithm today is referred to as supervised learning with the use of what is called a convolutional neural network, or CNN (modeled after how humans process visual perception through our neurons). Supervised learning means that you give the answer or labels (the final preferred output) associated with a very large data set of inputs to train it. The large input (for instance images of a specific particle) helps the computer predict/learn or to correctly classify similar images without further training of labels. The CNN is made up of layers of individual nodes (neurons) that identify patterns and points of interest by assigning weights layer-by-layer (identifying patterns of patterns). With the help of humans, the CNN is trained by validating correct activations of nodes by tweaking weights. It is a trial-and-error process, which eventually results in an accurate output. CERN could use a CNN to identify known particles as they are appearing and classifying/learning all that are unknown.
CERN has begun to use machine learning algorithms, but training algorithms is a difficult and laborious task, so the use of machine learning may not be fully utilized (details of why it is difficult is beyond the scope of this article, but if you are interested, look up “training” neural networks and the problem of the vanishing gradient). This is where the use of unsupervised deep (multi-layered) learning could become quite useful to CERN.
As unsupervised suggests, algorithms train with very large unlabeled data sets without the help of humans to arrive at a preferred output. The use of unsupervised methodologies will become increasingly important and likely the only way to best identify behaviors of known and new particles. It is estimated that near future upgrades to the LHC will result in 10x more data than current day. It is clear that these levels of data are far beyond what even the brightest minds on the planet can process and comprehend, hence this is why we need to rely on more advanced unsupervised deep machine learning.
How a computer could imagine a new dimension
One way CERN could better utilize unsupervised learning is to write descriptions/concepts of theoretical particles and the computer could then create representations from its imagination. From its imagination it could identify new particles faster as they appear, or through abstraction it could extrapolate. Yes, this may sound crazy that a computer could come up with images from its imagination, but in 2015 this was done at machine learning firm Indico in Boston and Facebook’s AI research lab in New York with a network called a generative adversarial network. Essentially a generative adversarial network is a system that creates fake data in one part the network in the attempt to fool the other part (mistaken as training data). The result has been more precise output on its own. The team gave pictures of smiling women and then told the system to remove pictures of women with neutral expressions, and add men with neutral expressions. The team hoped that the result would be that the system would understand the concept of smiling and combine with man to ‘imagine’ a ‘smiling’ man. The neural network successfully generated pictures of smiling men without human assistance (for more on the concept of generative adversarial networks and the experiments, check out this article by Facebook).
We may not be able to see into other dimensions, but with the use of advanced deep machine learning we may finally be able to prove the existence of them. Most scientists would agree that we are at the precipice of significant discoveries in the high-energy subatomic field and much of these discoveries will be made possible using machine learning. No longer can scientists simply rely on hand-coding algorithms and manually training neural networks. Around the same time as Indico and Facebook’s generative adversarial network experiments in 2015 several high-energy research facilities began to employ CNNs to their experiments. Scientists at Fermilab’s NOvA neutrino experiment have started to use CNNs and CERN has started to explore deep learning to make their experiments more autonomous.
Who knows, perhaps the computer will be the first to ‘imagine’ what a new dimension looks or behaves like!
Brian Rogers (firstname.lastname@example.org) – Founder neurothink.io and organizer for the Machine Learning Society New York Chapter, is true technologist with 20+ years experience in technology management and software development, having started several companies, authored patents in advance video and television. He holds a BA in Economics from Denison University and an MS in Technology Management from Columbia University.