Diverse Datasets for Equitable AI Models

How do we ensure that the future of AI is safe for everyone? Listen to women. Specifically, the brilliant women of color researchers who, like Timnit Gebru, Dr. Rumman Chowdhury, Safiya Noble, Ph.D., Seeta Pena Gangadharan, and Dr. Joy Buolamwini, have been sounding the alarm about the societal discrimination and biases that AI can magnify. An analysis of data sources that feed GPT-2 revealed that less than 15% of Wikipedia contributors were women or girls, only 34% of Twitter users were women, and 67% of Redditors were men. These sources are where large language models (LLMs) get their training data (aka the data you use to train a machine learning algorithm or model). Even more disheartening, Gebru’s research proves that white supremacist and misogynistic views are prevalent in the training data. Buolamwini’s project also revealed that darker-skinned women were 34.7% more likely to be misclassified than white men at 0.8%. This resulted from the datasets being simply not diverse enough, as the systems were not given enough Black and brown faces to learn how to understand what they look like. We must be aware of the consequences of bias in the automated systems used by 99% of Fortune 500 companies for hiring practices. AI-powered discrimination is a pressing issue affecting real lives. As artificial intelligence continues gaining traction, it’s time for us to take responsibility for our decisions about how these technologies are trained and where the data is coming from. By including different perspectives, we can uncover blind spots, mitigate biases, and ensure that AI benefits everyone.

Brilliant and timely article by Tom Godden 1. Treat data as a product, not a byproduct. Adopt practices like versioning, dedicated resources, and clear governance structures to manage data like a valuable product aligned with business objectives. 2. Curate diverse datasets. Ensure your training data encompasses a broad range of demographics, industries, and experiences to build fair, inclusive, and unbiased AI systems that accurately reflect all stakeholders. 3. Prioritize data quality. Implement robust validation pipelines, human review processes, and continuous monitoring to detect anomalies, errors, and biases in your data. Poor quality data leads to nonsensical AI outputs and significant deficiencies. #AI #generativeAI #data

In recent reflections on the surge of AI within the edtech landscape, an alarming trend becomes evident: the barrier to entry is significantly lower than it was during the Web 2.0 rush of the late 2000s. Unlike the previous era where products were often built from scratch, requiring substantial innovation and development, the core components of AI—such as large language models (LLMs) and comprehensive data sets—are already widely available. This shift means that companies can more easily package these technologies, adorning them with flashy branding and aggressive public relations campaigns, without necessarily contributing foundational innovations to the field. This context magnifies the importance of Ken Shelton's critical questions, as they bring into focus not just the what and how of AI in education, but also the who and why behind these technologies. It urges educators and stakeholders to: 1️⃣ Examine Data Sets and Supervision: What data sets does your organization use? Do these data sets bear labels, and are they supervised? The integrity and bias of data sets underpin the outcomes AI technologies produce, making transparency around these elements non-negotiable. 2️⃣ Scrutinize Diversity in Design: How does your design and decision-making team's diversity reflect the multifaceted identities of our student body? The perspectives and experiences of those creating AI solutions must resonate with, and reflect, the diversity of those impacted by these technologies. Understanding to what extent these teams' lived experiences align with our students' realities is crucial in creating equitable educational tools. 3️⃣ Question Impact and Transformation Goals: What impact does your organization aim to achieve within the education sector? Beyond mere functional contributions, how do your efforts seek to challenge and transform existing norms? A critical examination of how these technologies plan to dismantle historical and institutional barriers is imperative. In the fast-evolving AI landscape, the ease of entry underscores the necessity for vigilance, not mere skepticism. Ken Shelton's critical questions serve as essential due diligence, ensuring we embrace new AI technologies with informed enthusiasm. These inquiries help us discern genuine educational advancements from mere novelties, guiding us towards solutions that are equitable, inclusive, and truly transformative. By demanding clarity on data integrity, team diversity, and impact, we advocate for a future where technology aligns with our educational values and goals. #ai #aiethics #edtech #education #innovation #vigilance #educationalequity #criticalthinking #digitalcitizenship

The buzzword at BioIT World 2024 was AI. Dan Stanzione set the tone, providing a magical tour of the evolution of AI, covering hardware, software, AI/ML, Simulation and Analytics. Against this context, I was privileged to lead a discussion on "Data Readiness for AI'. Jesse Johnson, Shameer Khader, Ph.D, MPH, Siping Wang, Gian Prakash, Jay Schuren and I, unpacked the topic using the framework "Acquire>Manage>Govern>Use".  Acquire: Design and execute experiments with AI in mind. Anticipate broad and long term use of data by AI Plan for diversity within and across datasets through data acquisitions or patient selection. Lack of diversity leads to overfitted models with no real world applicability   Manage Data Ecosystem - A single data store will never be an answer - find ways to work with data in all formats and locations Start somewhere - Even if you don’t have a clear idea of the path forward. Data in any form can be transformed so start with storing them mindfully. Timely updates to data - so models are current and relevant. Plan for access and integration to data so models are not just built on PowerPoint Think of how data will be used multimodally when you plan for management of select data domains   Governance A judicious combination of policies and processes yields good data quality, a must for good AI models Pragmatic not onerous approaches especially for personal and secure data, generate better adoption of governance. David Hewlett in his plenary remarked that we are often driven by fear. And good governance is one that allows for compliance but is not overly driven by fear.   Use Deliberate empathy: Psychology of data producers and consumers differs across the value chain - bring all relevant constituents, patients, scientists, data scientists, technologists, decision makers into discussions. Literacy and Trust: Create awareness, education, upskilling in data, digital and AI. Speak trilingual across business, data and technology. Build confidence of teams in AI, for good data governance and management Change management "starts at the very beginning" and everyone must think of themselves as agents of change.   Operating Model Business led discussions make for good data practices Old technology investments don't have to be replaced, but new technologies must be adopted only when they show value Evolve support processes e.g. contracting, procurement etc. to meet new needs Include data as a key player in digitization conversations   Lessons from other industries Pay attention to data at point of acquisition - use prompts that alert to quality Invest in lineage in a shareable way Practice embedding as a way to make data accessible to ML models Summarize and annotate data in a "privacy agnostic way" to help improve awareness of datasets Use LLMs to turn free text into structure for managing data quality, Thank you Cindy Crowninshield, RDN, LDN, HHC and BioIT World for an energetic meet.   #data #AI #BioITWorld