Theoretical Physicists Deliberately Fooled Intelligent Machines

When computers independently identify bodies of water and their outlines in satellite images, or beat the world’s best professional players at the board game Go, adaptive algorithms are working in the background. Programmers provide these algorithms with known examples in a training phase: images of bodies of water and land, or sequences of Go movements that have led to successes or failures in tournaments. Just as our brain nerve cells produce new networks during learning processes, special algorithms adapt in the learning phase based on the examples presented to them. This continues until they are able to differentiate bodies of water from land in unfamiliar photos, or successful motion sequences from unsuccessful ones.

Until now, these artificial neural networks were used in machine learning with a known decision-making criterion: we know what a body of water is and what sequences of moves were successful in Go tournaments.

Separating the wheat from the chaff

Now a group of scientists working under Sebastian Huber, Professor of Condensed Matter Theory and Quantum Optics at ETH Zurich, have expanded the applications of these neural networks by developing a method that allows not not only to categorize all data, but also to recognize whether complex data sets contain any categories at all.

Questions of this kind arise in science: for example, the method could be useful for the analysis of measurements coming from particle accelerators or astronomical observations. Physicists could thus filter out the most promising measurements from their often unmanageable amounts of measurement data. Pharmacologists could extract molecules with a certain probability of having a specific pharmaceutical effect or side effect from large molecular databases. And data scientists could sort through huge masses of messy data ripples and get usable insights (data mining).

Find a border

The ETH researchers applied their method to an intensively researched phenomenon in theoretical physics: a many-body system of interacting magnetic dipoles that never reaches a state of equilibrium, even in the long term. Such systems have been described recently, but it is not yet known in detail what quantum physical properties prevent a many-body system from entering a state of equilibrium. In particular, it is unclear exactly where the boundary lies between systems that reach equilibrium and those that do not.

In order to locate this limit, the scientists developed the “act as if” principle: taking data from quantum systems, they established an arbitrary limit based on a parameter and used it to divide the data into two groups. They then trained an artificial neural network by pretending that one group reached a steady state while the other did not. So the researchers acted as if they knew where the boundary was.

Scientists have confused the system

They trained the network countless times overall, with a different limit each time, and tested the network’s ability to sort data after each session. The result was that, in many cases, the network struggled to classify the data as the scientists had. But in some cases, the division into two groups was very precise.

The researchers were able to show that this sorting performance depended on the location of the border. Evert van Nieuwenburg, a PhD student in Huber’s group, explains it this way: “By choosing to train with a boundary far from the real boundary (which I don’t know), I am able to mislead the network. re-train the network incorrectly – and ill-trained networks are very bad at classifying data.” However, if by chance a boundary is chosen close to the actual boundary, a very efficient algorithm is produced. By determining the performance of the algorithm, the researchers were able to draw the boundary between quantum systems that achieve equilibrium and those that do not: the boundary is where the sorting performance of the network is highest.

The researchers also demonstrated the capabilities of their new method using two other theoretical physics questions: topological phase transitions in one-dimensional solids and the Ising model, which describes magnetism inside solids.

Categorization without prior knowledge

The new method can also be illustrated in a simplified form with a thought experiment, where we want to classify the red, reddish, bluish and blue balls into two groups. We assume we have no idea what such a classification might reasonably look like.

If a neural network is trained by telling it that the dividing line is somewhere in the red region, it will confuse the network. “You’re trying to teach the network that blue and red balls are the same and asking it to tell the difference between red and red balls, which it just isn’t able to do,” says Huber.

On the other hand, if you place the boundary in the purple color spectrum, the network learns an actual difference and sorts the balls into red and blue groups. However, it is not necessary to know in advance that the boundary line must be in the purple region. By comparing sorting performance at a variety of chosen boundaries, this boundary can be found without prior knowledge.

Source of the story:

Material provided by ETH Zürich. Note: Content may be edited for style and length.

Comments are closed.