MIT Professor Paris Smaragdis, a musician and computer scientist, is on a mission to enhance machine hearing. He aims to develop machines with superior hearing capabilities, enabling them to locate sounds such as traffic accidents or industrial drill malfunctions. Smaragdis, a Greek native, discovered his fascination with sound and technology at a young age through classical guitar and synthesi...
Meta's Privacy Aware Infrastructure (PAI) introduces PrivacyLib, a common library used for enforcing and monitoring data flows in the final stage of PAI workflow. Previously, new privacy requirements meant integrating privacy enforcement separately into each data and processing system, leading to a lengthy process. PrivacyLib allows system integration once, with new privacy capabilities as modu...
In this post, we share the story of Property Finder, a digital real estate application, where 96.8% of their evaluation labels were found to be inherited from a single model's predictions rather than verified by human experts, leading to systemic bias and inaccurate comparisons of alternative models. Property Finder, which processes thousands of property images daily across the Middle East and ...
Climate Week NYC, an annual event bringing together over 100,000 people in more than 1,000 events across New York City to discuss climate challenges and explore solutions, saw a strong presence from the MIT community this year. MIT was represented multiple days with appearances by Evelyn Wang, vice president for energy and climate, and numerous sessions organized by MIT departments, labs, cente...
This text discusses the benefits and evolution of enterprise agentic retrieval using Amazon Bedrock Knowledge Bases on the AWS Cloud Development Kit (CDK). Traditional Retrieval Augmented Generation (RAG) systems have limitations, such as handling questions that span multiple corpora or require multiple lookups. To overcome these challenges, the text introduces agentic retrieval, where a reason...
In this text, the author explains the concept of an "AI agent" and how it faces the same issue as the character in the movie "Groundhog Day," where it must re-read its entire story before performing an action. This leads to high costs due to repeated inputs. The author then introduces "prompt caching," a solution where previously processed inputs are saved and reused, reducing costs and maintai...
The author employed AI to enhance the performance of their existing C# gradient boosting regression implementation, resulting in significant improvements. Gradient boosting regression is a tree-based method used to predict a single numeric value, such as predicting a maximum safe loan amount based on various factors. This process involves creating a collection of decision trees, where each tree...
This text discusses the challenge of managing computer vision models across multiple cameras in industrial quality inspection systems. Each camera captures the world differently, resulting in unique training data. To address this, individual fine-tuned models are needed for each camera, leading to the need to train, evaluate, version, and ship thousands of models without human intervention. The...
The author implemented a decision tree regression system using JavaScript and used an AI system to analyze its correctness. The AI discovered edge cases where the code might fail and suggested improvements for enhanced performance. After extensive efforts, a revised version of the JavaScript decision tree regression was developed, leading to a synthetic dataset with 220 items as an example. Key...
When updating Amazon SageMaker AI real-time endpoint configuration, changing the instance type, instance count, or environment variables through AWS CloudFormation or AWS CDK can result in deployment failure. This is due to SageMaker AI requiring the scalable target to be deregistered before updating the endpoint. However, deregistering the scalable target causes CloudFormation to lose track of...
Amazon Quick Sight's geospatial maps allow businesses to visualize patterns connected to location through point maps, filled maps, and now, layer maps. With the new point and LineString geometry types from GeoJSON files, users can mark specific locations and trace routes. This post covers dataset and GeoJSON structure, prerequisites, and advanced techniques for creating layer maps in Amazon Qui...
NVIDIA founder and CEO Jensen Huang recently visited Astana, Kazakhstan, where he discussed the country's push to build a strong ecosystem for artificial intelligence (AI). Kazakhstan's goal is to create AI that reflects its language and culture, utilizing its abundant energy resources to support homegrown AI and transforming energy into intelligence for export. The conversation also touched on...
Multi-Turn Reinforcement Learning (MTRL) is a training method used to teach models to complete multi-step tasks with feedback only after the entire sequence. In contrast to fine-tuning, MTRL doesn't rely on a single correct sequence or label. Agents interact with an environment, making decisions, and reporting rewards after each sequence ends. With MTRL, a model can be made proficient in a spec...
Bayer Crop Science uses machine learning models to analyze data from genomics, field trials, imagery, and environmental conditions for plant breeding. To manage and make repeatable and governable these ML model configurations, Bayer developed the Connected Decision Engine (CDE). As the usage increased, understanding the results became a challenge, leading to the creation of the Agentic Decision...
MIT's School of Humanities, Arts, and Social Sciences was celebrated at its 75th Anniversary Symposium for its value and contributions to teaching, research, and society. The event, held at the Linde Music Building, drew scholars, students, staff, and alumni from across the school. MIT leadership, including President Sally Kornbluth and Provost Anantha Chandrakasan, emphasized the importance of...