A new preprint on arXiv introduces the concept of a "sign code" for trained deep networks. The authors observe that every neuron outputs a real number, and that number can be separated into two parts: its sign (positive or negative) and its magnitude (absolute size). They propose that the collection of all neuron signs—the sign code—may be a key to understanding how these networks represent information.

The idea is that while magnitudes vary continuously, signs are binary. This binary nature could be a fundamental, previously overlooked property of deep networks. The abstract is brief, but it frames the sign code as a potentially important lens for analyzing network behavior, possibly offering a simpler way to interpret what a network has learned.

The full paper is not yet available in the abstract, which cuts off mid-sentence, so the complete claims and evidence are unclear. Still, the sign code concept is a concrete, testable proposal: it reduces each neuron's output to a single bit of information, and the authors argue that this bit-level view deserves attention.