🤯MIT's AI Debiasing Breakthrough

MIT researchers tackle AI bias, boost accuracy with new breakthrough technique

Key Points 

• MIT’s technique removes specific data points causing bias, improving AI performance without sacrificing overall accuracy.

• The method helps identify hidden biases in unlabeled data, making AI models fairer in high-stakes applications.

📖Context of the news - Machine-learning models often struggle to make good predictions for groups that aren’t well-represented in their training data. For example, a model trained mostly on data from male patients might not work well for female patients when used in a hospital.  

One way to fix this is by balancing the dataset, which involves removing data until all groups are equally represented. However, this can mean losing a lot of data, which can hurt the model’s overall performance.  

👨🏻‍🎓News - MIT researchers have come up with a new way to solve this problem. Their method finds and removes only the specific data points that cause the model to fail for underrepresented groups. This means they can use much less data than before while still keeping the model accurate overall.  

What’s interesting is that this technique also works on datasets without labels, which are common in many fields. It could help make AI models fairer in areas like healthcare, ensuring that minority groups aren’t overlooked or misdiagnosed.  

👨🏻‍🏫What’s more? Most machine-learning models are trained on huge datasets collected from all over the internet. These datasets are too big to manually check, so they often include poor-quality data.  

The researchers combined two ideas: identifying bad data and understanding its effect on a model. Their goal was to fix what’s called “worst-group error,” where a model performs badly for smaller, underrepresented groups.  

🤓What was the result? By pinpointing the exact data points that lower accuracy for these groups, the team removed only those samples and retrained the model. This way, the model stayed accurate overall while performing much better for underrepresented groups.

Google’s Gemini gets a major upgrade with Deep Research feature

Key Points 

• Google Gemini’s Deep Research helps users conduct in-depth research and create detailed reports.

• Deep Research is available for Google One AI Premium Plan users and will expand to mobile apps in 2025.

🤖News – Google is upgrading its Gemini chatbot platform with a new feature called "Deep Research" that helps users work through research problems and create detailed reports. 

With Deep Research, Gemini can analyze information from across the web to help answer questions. It organizes the findings into research briefs, which can be viewed in the Gemini app or exported to Google Docs for further editing. Google says the new feature uses advanced reasoning and can handle longer context to generate thorough reports.

👨🏻‍💻How does it work? To get started, a user asks a question, and Deep Research creates a multi-step plan to gather relevant information. Once the user approves, it refines its search, saving useful details and using them to guide further research. This process repeats until a final report is generated, summarizing the most important findings.

Currently, Deep Research is available exclusively to Google One AI Premium Plan users, who pay $20 a month for access to Gemini Advanced. The feature is available in English on desktop and mobile web, with plans to expand to mobile apps in early 2025.