Researchers have crafted an artificial intelligence (AI) system capable of deciphering fragments of ancient Babylonian texts. Dubbed the “Fragmentarium,” the algorithm holds the potential to piece together some of the oldest stories ever written by humans, including the Epic of Gilgamesh.

The work comes from a team at Ludwig Maximilian University in Germany who have been attempting to digitize every surviving Babylonian cuneiform tablet since 2018.

The problem with understanding Babylonian texts is that the narratives are written on clay tablets, which today exist only in countless fragments. The fragments are stored at facilities that are continents away from each other, such as the British Museum in London and the Iraq Museum in Baghdad.

Humans are innately able to reason about the behaviors of different physical objects in their surroundings. These physical reasoning skills are incredibly valuable for solving everyday problems, as they can help us to choose more effective actions to achieve specific goals.

Some computer scientists have been trying to replicate these reasoning abilities in artificial intelligence (AI) agents, to improve their performance on specific tasks. So far, however, a reliable approach to train and assess the physical reasoning capabilities of AI algorithms has been lacking.

Cheng Xue, Vimukthini Pinto, Chathura Gamage, and colleagues, a team of researchers at the Australian National University, recently introduced Phy-Q, a new testbed designed to fill this gap in the literature. Their testbed, introduced in a paper in Nature Machine Intelligence, includes a series of scenarios that specifically assess an AI agent’s physical reasoning capabilities.

“Liquid” neural nets, based on a worm’s nervous system, can transform their underlying algorithms on the fly, giving them unprecedented speed and adaptability.

Diffusion models have recently produced outstanding results on various generating tasks, including the creation of images, 3D point clouds, and molecular conformers. Ito stochastic differential equations (SDE) are a unified framework that can incorporate these models. The models acquire knowledge of time-dependent score fields through score-matching, which later directs the reverse SDE during generative sampling. Variance-exploding (VE) and variance-preserving (VP) SDE are common diffusion models. EDM offers the finest performance to date by expanding on these compositions. The existing training method for diffusion models can still be enhanced, despite achieving outstanding empirical results.

The Stable Target Field (STF) objective is a generalized variation of the denoising score-matching objective. Particularly, the high volatility of the denoising score matching (DSM) objective’s training targets can result in subpar performance. They divide the score field into three regimes to comprehend the cause of this volatility better. According to their investigation, the phenomenon mostly occurs in the intermediate regime, defined by various modes or data points having a similar impact on the scores. In other words, under this regime, it is still being determined where the noisy samples produced throughout the forward process originated. Figure 1(a) illustrates the differences between the DSM and their proposed STF objectives.

Figure 1: Examples of the DSM objective’s and our suggested STF objective’s contrasts.