Researchers in Switzerland are hoping to reduce LLM compute costs by 85%. YouTube will soon tell you if you’re being a hater in the comments, Google brings no-code machine learning to Sheets with SimpleML and Intel’s photorealistic simulation platform claims to accelerate the training and validation of embodied AI systems in indoor domains. Let’s dive in!
Research Highlights
Researchers from Georgia Institute of Technology presented the Trust, but Verify (TbV) dataset. The dataset was created by mining thousands of hours of data from autonomous vehicle fleet operations over a nine-month period. With over 7.8 million images, the dataset, which consists of maps and logs collected in six North American cities, claims to be one of the largest AV datasets to date.
Google researchers proposed sparse upcycling as a method for reusing sunk training costs by starting a sparsely activated Mixture-of-Experts model from a dense checkpoint. Training large, deep neural networks to convergence can be prohibitively expensive and sparsely activated models are becoming an attractive alternative to dense models.
Researchers in Switzerland presented the idea of Language Model Programming (LMP). Their paper argues that even though LLMs are continuing to advance, they still require ad-hoc interaction with complex task- and model-specific programs when it comes to such tasks as adapting language models for specific tasks. LMP claims to retain or increase the accuracy on several downstream tasks, while also significantly reducing the required amount of computation or cost in the case of pay-to-use APIs (13-85% cost savings).
ML Engineering Highlights
YouTube announced that its automated detection systems and machine learning models will soon notify you if a comment you left on a video is deemed to be too abusive in accordance with the website's community guidelines. If a user receives a notification but continues to post abusive comments, YouTube will silence them with a 24-hour commenting ban.
UltraSense, a consumer electronics company that provides sensor solutions, announced the development of new touch-control sensors for automobiles that are more accurate due to the use of machine learning. Their new In-Plane sensing automotive technology enables multi-mode sensing and human-machine interface (HMI) control in the SmartSurface plane (or A-Surface). They claim that this method greatly reduces the size and weight of the sensors, as well as the number of parts and complexity.
Google Sheets now includes Simple ML, a tool for applying machine learning to data prediction and sorting tasks. The tool comes with two functions by default: 'predict missing values' and'spot abnormal values'. For more complex applications, it can even predict values within large data sets - Google cited a study in which researchers used Simple ML to predict a person's age based on their DNA.
Open-Source Highlights
Intel Labs introduced an open-source simulator for AI. Their Simulator for Photorealistic Embodied AI Research (SPEAR), claims to be a highly realistic, open-source simulation platform that accelerates the training and validation of embodied AI systems in indoor domains. The solution can be downloaded under an open-source MIT license.
Google released a new free tool that makes it easier for open-source developers to find vulnerability information relevant to their projects. The Go-based OSV-Scanner tool matches a developer's code and dependencies against lists of known vulnerabilities and provides instant feedback if patches or updates are required.
What are tensors? And which deep learning library are we going to use? We’re glad you asked! Check out Unit 2 of Sebastian Raschka’s Deep Learning Fundamentals course to learn more about tensors and familiarize yourself with PyTorch, the most widely used deep learning library.
This week, we’re highlighting a few recent PRs that fostered some great conversation between members of our team and the Lightning community. Kudos to everyone involved!
#16005, about the datamodule(s) implementation in Bolts, reminds us that there’s always a tradeoff between “blackbox” or “magic” solutions and allowing users full flexibility over their work. We want to prevent users from having to mess around with the Lightning Trainer internals to make certain steps work, but it’s equally important to build with a variety of users in mind — particularly those who might not need the flexibility.
#16023, a personal favorite of this newsletter contributor, is about documentation. Often overlooked and sometimes swept under the rug in pursuit of innovation, documentation can be considered the living, breathing heart of an open-source project like ours. Cheers to Nikhil Shenoy for your contribution!
#15915 improves the performance and reduces the overheard of MLFlowLogger API calls. Shoutout to our community member Jake Schmidt, who first opened this issue and landed this PR. Awesome stuff!
Don’t Miss the Submission Deadline
IJCAI 2023: the 32nd International Joint Conference on Artificial Intelligence. Aug 19-25th 2023. (Cape Town, South Africa) Abstract due: January 11, 2023.Full paper submission deadline:January 18, 2023
ACL 2023: The 61st Annual Meeting of the Association for Computational Linguistics. July 9-14 2023 (Toronto, Canada). Full paper submission deadline: January 20, 2023
ICML 2023: Fortieth International Conference on Machine Learning. Jul 23-29 (Honolulu, Hawaii). Full paper submission deadline: January 26, 2023 08:00 PM UTC
IROS 2023 : International Conference on Intelligent Robots and Systems. Oct 1 – 5, 2023 (Detroit, Michigan). Full paper submission deadline: March 1, 2023
ICCV 2023: International Conference on Computer Vision. Oct 2 - 6, 2023. (Paris, France). 1. Full paper submission deadline: March 8, 2023 23:59 GMT
Want to learn more from Lightning AI? “Subscribe” to make sure you don’t miss the latest flashes of inspiration, news, tutorials, educational courses, and other AI-driven resources from around the industry. Thanks for reading!
Realtor Associate @ Next Trend Realty LLC | HAR REALTOR, IRS Tax Preparer
2yLove this.