Large scale training requires massive parallelism to finish the training within a reasonable amount of time. To support massive parallelism, large batch training is the key enabler but often at the cost of generalization performance. Existing works explore adaptive batching or hand-tuned static large batching, in order to strike a balance between the computational efficiency and the performance. However, these methods can provide only coarse-grained adaption (e.g., at a epoch level) due to the intrinsic expensive calculation or hand tuning requirements. In this paper, we propose a fully automated and lightweight adaptive batching methodology to enable fine-grained batch size adaption (e.g., at a mini-batch level) that can achieve stateof-the-art performance with record breaking batch sizes. The core component of our method is a lightweight yet efficient representation of the critical gradient noise information. We open-source the proposed methodology by providing a plugin tool that supports mainstream machine learning frameworks. Extensive evaluations on popular benchmarks (e.g., CIFAR10, ImageNet, and BERT-Large) demonstrate that the proposed methodology outperforms state-of-the-art methodologies using adaptive batching approaches or hand-tuned static strategies in both performance and batch size. Particularly, we achieve a new state-of-the-art batch size of 78k in BERT-Large pretraining with SQuAD score 90.69 compared to 90.58 reported in previous state-of-the-art with 59k batch size.

As soon as Echo Park Eats opened on the corner of Sunset Boulevard and Douglas Street in the fall of 2023, Sandy Romero said her neighborhood became overrun with delivery drivers. "The first day that they opened business it was chaotic, unorganized and it's just such a nuisance now," she said. Echo Park Eats is a ghost kitchen, a meal preparation hub for app-based delivery orders. It rents its kitchens to 26 different food vendors. The facility is part of CloudKitchens, led by Travis Kalanick, co-founder of Uber Technologies, which has kitchen locations across the nation including 11 in Los Angeles County.

Waymo, the autonomous ride-hailing company that launched its services in Los Angeles late last year, is recalling more than 1,200 vehicles due to a software defect, the National Highway Traffic Safety Assn. said Wednesday. The recall comes after a series of minor crashes with gates, chains and other obstacles in the road that did not result in any injuries, the Mountain View, Calif.-based company said in a filing with the NHTSA. The recall applies to 1,212 driverless vehicles operating on Waymo's fifth-generation automated driving software. Waymo released a software update to resolve the issue, and that update has already been rolled out in all affected vehicles, the recall notice said. The company operates more than 1,500 vehicles across Los Angeles, San Francisco, Phoenix and Austin.

When survivors from January's wildfires in Los Angeles County apply to rebuild their homes, their first interaction might be with a robot. Artificial intelligence will aid city and county building officials in reviewing permit requests, an effort to speed up a process already being criticized as too slow. "The current pace of issuing permits locally is not meeting the magnitude of the challenge we face," Gov. Gavin Newsom said when announcing the AI deal in late April. Some 13,000 homes were lost or severely damaged in the Eaton and Palisades fires, and many families are eager to return as fast as they can. Just eight days after the fire began and while it was still burning, the city received its first home rebuilding application in Pacific Palisades.

Treasury Secretary Scott Bessent kicked off Michael Milken's annual financial bash in Beverly Hills by doubling down on President Trump's economic policy of trade reform, tax cuts and deregulation -- promising the "America First" agenda would be "the blueprint for a more abundant world." The former hedge fund manager, in a brief speech Monday that opened the Milken Institute Global Conference, said that all three elements of the policy must be taken together in order to be understood. "They are interlocking parts of an engine designed to drive long-term investment in the American economy," he said, in remarks at the Beverly Hilton. "Tariffs are engineered to encourage companies like yours to invest directly in the United States. Hire your workers here, build your factories here, make your products here. You'll be glad you did, not only because we have the most productive work force in the world, but because we will soon have the most favorable tax and regulatory environment as well," he said.

Cameras were first installed on the windshields of some Metro buses last year, but the first tickets were issued in mid-February. Initially, the only buses to have cameras were along line 212, from Hollywood/Vine to Hawthorne/Lennox stations via La Brea Avenue, and line 720, from Santa Monica to downtown L.A. via Wilshire Boulevard. Line 70, which services Olive Street and Grand Avenue, and lines 910 and 950 that serve Metro's J Line have since been included. The AI-powered cameras scan for illegally parked cars and compile a video of each violation, a photo of the license plate and the time and location, according to the Los Angeles County Metropolitan Transportation Authority. Each citation is reviewed by a human.

Tricia Helfer, who played a humanoid robot Cylon on "Battlestar Galactica," says the show's look at the conflict between humans and AI still resonates today. "We did warn against AI while we were shooting it," Helfer told Fox News Digital at the Beverly Hills Film Festival this week. She continued, "It was 20 years ago, and I've recently re-watched it and went, 'Oh my gosh, it's even more relevant now.' So I think we just really need to be careful. It's a slippery slope between using it to our advantage and having it maybe be able to control us a little bit." "I think we're a little bit far off from the humanoid Cylons yet and humanoid robots, but I don't know, they're coming," Helfer added.

Imagine observing someone scratching their arm; to understand why, additional context would be necessary. However, spotting a mosquito nearby would immediately offer a likely explanation for the person's discomfort, thereby alleviating the need for further information. This example illustrates how subtle visual cues can challenge our cognitive skills and demonstrates the complexity of interpreting visual scenarios. To study these skills, we present Visual Riddles, a benchmark aimed to test vision and language models on visual riddles requiring commonsense and world knowledge. The benchmark comprises 400 visual riddles, each featuring a unique image created by a variety of text-to-image models, question, groundtruth answer, textual hint, and attribution. Human evaluation reveals that existing models lag significantly behind human performance, which is at 82% accuracy, with Gemini-Pro-1.5 leading with 40% accuracy. Our benchmark comes with automatic evaluation tasks to make assessment scalable. These findings underscore the potential of Visual Riddles as a valuable resource for enhancing vision and language models' capabilities in interpreting complex visual scenarios.

This paper presents Incremental Vision-Language Object Detection (IVLOD), a novel learning task designed to incrementally adapt pre-trained Vision-Language Object Detection Models (VLODMs) to various specialized domains, while simultaneously preserving their zero-shot generalization capabilities for the generalized domain. To address this new challenge, we present the Zero-interference Reparameterizable Adaptation (ZiRa), a novel method that introduces Zero-interference Loss and reparameterization techniques to tackle IVLOD without incurring a significant increase in memory usage. Comprehensive experiments on COCO and ODinW-13 datasets demonstrate that ZiRa effectively safeguards the zeroshot generalization ability of VLODMs while continuously adapting to new tasks. Specifically, after training on ODinW-13 datasets, ZiRa exhibits superior performance compared to CL-DETR and iDETR, boosting zero-shot generalizability by substantial 13.91 and 8.74 AP, respectively.