To enhance interpretability, we introduce a novel bijective decoding head that explicitly models permutations via the Gumbel-Sinkhorn method, enabling direct extraction of learned cipher mappings. Our architectural innovations and analysis methods are applicable beyond cryptograms and offer new insights into neural network generalization and interpretability. A cryptogram is a type of puzzle in which text is encrypted using a substitution cipher, and the user's task is to recover the original plaintext by inferring the cipher used for the encryption. Users typically solve cryptograms based on prior knowledge about language letter frequency distributions and common words. Originally developed for real encryption purposes, they are now popular in newspapers and puzzle books for entertainment purposes due to their simplicity. This simplicity, however, provides a unique testbed for testing and understanding generalization and reasoning in neural networks. In a one-to-one monoalphabetic substitution cipher, each letter in a fixed alphabet is mapped to a unique substitute character; this cipher represents a bijective mapping over the alphabet. While other ciphers exist (e.g., Vigen ` ere cipher, Playfair cipher), we focus here on one-to-one monoalphabetic substitution ciphers, as the problem space is extremely large but remains structurally simple to interpret. We hereafter mean one-to-one monoalphabetic substitution cipher when we say "cipher", unless otherwise specified. More formally, let Σ be a finite alphabet of size V representing allowable characters (e.g., 26 for the English alphabet).

Due to user demand and recent regulation (GDPR, AI Act), decisions made by AI systems need to be explained. These decisions are often explainable only post hoc, where counterfactual explanations are popular. The question of what constitutes the best counterfactual explanation must consider multiple aspects, where "distance from the sample" is the most common. We argue that this requirement frequently leads to explanations that are unlikely and, therefore, of limited value. Here, we present a system that provides high-likelihood explanations. We show that the search for the most likely explanations satisfying many common desiderata for counterfactual explanations can be modeled using mixed-integer optimization (MIO). In the process, we propose an MIO formulation of a Sum-Product Network (SPN) and use the SPN to estimate the likelihood of a counterfactual, which can be of independent interest. A numerical comparison against several methods for generating counterfactual explanations is provided.

Toxic language detection systems often falsely flag text that contains minority group mentions as toxic, as those groups are often the targets of online hate. Such over-reliance on spurious correlations also causes systems to struggle with detecting implicitly toxic language. To help mitigate these issues, we create ToxiGen, a new large-scale and machine-generated dataset of 274k toxic and benign statements about 13 minority groups. We develop a demonstration-based prompting framework and an adversarial classifier-in-the-loop decoding method to generate subtly toxic and benign text with a massive pretrained language model. Controlling machine generation in this way allows ToxiGen to cover implicitly toxic text at a larger scale, and about more demographic groups, than previous resources of human-written text. We conduct a human evaluation on a challenging subset of ToxiGen and find that annotators struggle to distinguish machine-generated text from human-written language. We also find that 94.5% of toxic examples are labeled as hate speech by human annotators. Using three publicly-available datasets, we show that finetuning a toxicity classifier on our data improves its performance on human-written data substantially. We also demonstrate that ToxiGen can be used to fight machine-generated toxicity as finetuning improves the classifier significantly on our evaluation subset. Our code and data can be found at https://github.com/microsoft/ToxiGen.

Artificial intelligence has "tremendous potential" to support and accelerate a reliable and lowest-cost energy transition, a new report by the World Economic Forum has revealed. Through its high-tech applications, AI can integrate renewable energy resources into the power grid, support an autonomous electricity distribution system and open up new revenue streams for demand-side flexibility, WEF said in its Harnessing Artificial Intelligence to Accelerate Energy Transition report compiled in collaboration with BloombergNEF and Deutsche Energie-Agentur (dena) – the German energy agency. AI can create substantial value for the global energy transition, the report said. Every 1 per cent of additional efficiency in demand will create $1.3 trillion in value between 2020 and 2050 due to reduced investment needs, according to BloombergNEF's net-zero scenario modelling. This could be achieved by enabling greater energy efficiency and flexing demand. "In energy, we are only seeing the beginning ...

Back and forth he walks across the polished hardwood floors of a barge anchored in a fjord off the southwestern coast of Norway. The barge sits alongside one of the world's largest salmon farms. It's November and the sky is cloudless, the mountains are snow-capped, the water is a clear sapphire blue. The control room has the feel of a W Hotel lobby with its elegant lighting and spare Scandinavian design. On one wall are huge monitors streaming video from nine underwater cages nearby. Aarskog scans the footage--masses of salmon swimming in circles like glittering cyclones--and mutters what I take to be Norwegian profanities.

Fish farming is big business - the industry now produces about 100 million tonnes a year - and with salmon prices soaring, producers are turning to lasers, automation and artificial intelligence to boost production and cut costs. How do you know if farmed salmon have had enough to eat? Well, according to Lingalaks fish farms in Norway, which produce nearly three million salmon each year, the fish make less noise once the feeding frenzy is over. The firm knows this thanks to a new hydro-acoustic system it has installed at one of its farms. The system listens to the salmon sloshing loudly about as they feed in a cluster. When the fish have had enough, they swim off and the noise lessens.

Fish farming is big business - the industry now produces about 100 million tonnes a year - and with salmon prices soaring, producers are turning to lasers, automation and artificial intelligence to boost production and cut costs. How do you know if farmed salmon have had enough to eat? Well, according to Lingalaks fish farms in Norway, which produce nearly three million salmon each year, the fish make less noise once the feeding frenzy is over. The firm knows this thanks to a new hydro-acoustic system it has installed at one of its farms. The system listens to the salmon sloshing loudly about as they feed in a cluster. When the fish have had enough, they swim off and the noise lessens.

When you picture laser-wielding robots, equipped with the latest machine vision algorithms, what setting do you imagine them operating in? Currently being employed in the North Sea fjords in Norway, along with a select few lochs in Scotland, a smart underwater drone developed by Stingray Marine Solutions is designed to help deal with the problem of sea lice. Didn't know that salmon had lice? Don't worry, you're not alone. "It's not a problem that's all that well known outside of the salmon farming industry in Norway," John Breivik, general manager at Stingray, told Digital Trends.