An important goal of ML research is to identify and mitigate unwanted biases intrinsic to datasets and already incorporated into pre-trained models. Previous approaches have identified biases using highly curated validation subsets, that require human knowledge to create in the first place. This limits the ability to automate the discovery of unknown biases in new datasets. We solve this by using interpretable vision-language models, combined with a filtration method using LLMs and known concept hierarchies. More exactly, for a dataset, we use pre-trained CLIP models that have an associated embedding for each class and see how it drifts through learning towards embeddings that disclose hidden biases. We call this approach ConceptDrift and show that it can be scaled to automatically identify biases in datasets like ImageNet without human prior knowledge. We propose two bias identification evaluation protocols to fill the gap in the previous work and show that our method significantly improves over SoTA methods, both using our protocol and classical evaluations. Alongside validating the identified biases, we also show that they can be leveraged to improve the performance of different methods. Our method is not bounded to a single modality, and we empirically validate it both on image (Waterbirds, CelebA, ImageNet), and text datasets (CivilComments).

This work focuses on developing an effective meta-heuristic approach to protect against simultaneous attacks on nodes of a network modeled using a graph. Specifically, we focus on the $k$-strong Roman domination problem, a generalization of the well-known Roman domination problem on graphs. This general problem is about assigning integer weights to nodes that represent the number of field armies stationed at each node in order to satisfy the protection constraints while minimizing the total weights. These constraints concern the protection of a graph against any simultaneous attack consisting of $k \in \mathbb{N}$ nodes. An attack is considered repelled if each node labeled 0 can be defended by borrowing an army from one of its neighboring nodes, ensuring that the neighbor retains at least one army for self-defense. The $k$-SRD problem has practical applications in various areas, such as developing counter-terrorism strategies or managing supply chain disruptions. The solution to this problem is notoriously difficult to find, as even checking the feasibility of the proposed solution requires an exponential number of steps. We propose a variable neighborhood search algorithm in which the feasibility of the solution is checked by introducing the concept of quasi-feasibility, which is realized by careful sampling within the set of all possible attacks. Extensive experimental evaluations show the scalability and robustness of the proposed approach compared to the two exact approaches from the literature. Experiments are conducted with random networks from the literature and newly introduced random wireless networks as well as with real-world networks. A practical application scenario, using real-world networks, involves applying our approach to graphs extracted from GeoJSON files containing geographic features of hundreds of cities or larger regions.

The robotaxi is recording me sitting in the backseat, and I am recording it. Someone in the neighboring car is recording us both. It's an unusually hot day in San Francisco, and I am in a self-driving car named Charcuterie, operated by Cruise. Next to me is William Riggs, a professor at the University of San Francisco who studies self-driving cars. The front seats are both empty, and the wheel silently shifts as the car maneuvers itself along a thoroughfare next to Golden Gate Park.

Cruise will temporarily be deploying fewer autonomous vehicles in San Francisco while investigators are looking into "recent concerning incidents" involving its fleet. According to The New York Times and TechCrunch, the California Department of Motor Vehicles asked the company to cut its fleet in half after an incident wherein one of Cruise's robotaxis collided with a fire truck at an intersection. The fire truck had its sirens and red lights on and was responding to an emergency at the time, while the robotaxi has passengers onboard who sustained non-life-threatening injuries. In another, perhaps less controversial, incident a few days before that, a Cruise vehicle got stuck in wet concrete. The DMV said in a statement that its primary focus is "the safe operation of autonomous vehicles and safety of the public who share the road with these vehicles." It also added that it "reserves the right, following investigation of the facts, to suspend or revoke testing and/or deployment permits" if it determines that a company's vehicles is a threat to public safety.

The driver told local media he turned on his dashcam after seeing the Tesla driver leave the scene of another collision. Clean up for a crash on a northern California freeway turned deadly Saturday after a Tesla plowed into a fire truck that was shielding first responders who were working another accident. The driver of the Tesla Model S was killed when the car hit the truck that was parked on Interstate 680 around 4 a.m., according to the Contra Costa County Fire Protection District. A passenger was also critically injured. Assistant Chief Tracie Dutter said four firefighters who were in the ladder truck were treated for minor injuries.

On an early April morning, around 4 am, a San Francisco Fire Department truck responding to a fire tried to pass a doubled-parked garbage truck by using the opposing lane. But a traveling autonomous vehicle, operated by the General Motors subsidiary Cruise without anyone inside, was blocking its path. While a human might have reversed to clear the lane, the Cruise car stayed put. The fire truck only passed the blockage when the garbage truck driver ran from their work to move their vehicle. "This incident slowed SFFD response to a fire that resulted in property damage and personal injuries," city officials wrote in a filing submitted to the California Public Utilities Commission.

US government regulators are opening an investigation into Tesla's Autopilot system after cars using the feature crashed into stopped emergency vehicles. This story originally appeared on Ars Technica, a trusted source for technology news, tech policy analysis, reviews, and more. Ars is owned by WIRED's parent company, Condé Nast. The National Highway Transportation Safety Administration announced the investigation on Monday, and it encompasses 765,000 Teslas sold in the US, a significant fraction of all of the company's sales in the country. The agency says the probe will cover 11 crashes since 2018; the crashes caused 17 injuries and one death.

The US National Highway Traffic Safety Administration is investigating a crash involving a speeding Tesla that killed two people in a Los Angeles suburb, the agency said on Tuesday. Spokesman Sean Rushton would not say whether the Tesla Model S was on Autopilot when it crashed on 29 December in Gardena. That system is designed to automatically change lanes and keep a safe distance from other vehicles. The black Tesla had left a freeway and was moving at a high rate of speed when it ran a red light and slammed into a Honda Civic at an intersection, police said. A man and woman in the Civic died at the scene.

There are two big unsolved mathematical questions in artificial intelligence (AI): (1) Why is deep learning so successful in classification problems and (2) why are neural nets based on deep learning at the same time universally unstable, where the instabilities make the networks vulnerable to adversarial attacks. We present a solution to these questions that can be summed up in two words; false structures. Indeed, deep learning does not learn the original structures that humans use when recognising images (cats have whiskers, paws, fur, pointy ears, etc), but rather different false structures that correlate with the original structure and hence yield the success. However, the false structure, unlike the original structure, is unstable. The false structure is simpler than the original structure, hence easier to learn with less data and the numerical algorithm used in the training will more easily converge to the neural network that captures the false structure. We formally define the concept of false structures and formulate the solution as a conjecture. Given that trained neural networks always are computed with approximations, this conjecture can only be established through a combination of theoretical and computational results similar to how one establishes a postulate in theoretical physics (e.g. the speed of light is constant). Establishing the conjecture fully will require a vast research program characterising the false structures. We provide the foundations for such a program establishing the existence of the false structures in practice. Finally, we discuss the far reaching consequences the existence of the false structures has on state-of-the-art AI and Smale's 18th problem.

Get YouTube without the ads. Want to watch this again later? Sign in to add this video to a playlist. Report Need to report the video? Sign in to report inappropriate content.