Revamping the Underdogs

Traditionally, we’ve considered some network architectures as inherently flawed or limited in dealing with complex tasks. However, these neural networks might be victim to an unfavorable starting point in the parameter space rather than a lack of potential. The researchers found that by pairing these networks with a structured “guide” network for a short while, they could steer the struggling networks towards effective learning.

This technique, unlike knowledge distillation—in which a student model emulates the outputs of a teacher—relies on internal representations. Here, the target network absorbs how the guide network organizes information across its layers instead of mimicking its predictions. Even when the guide network is untrained, this process facilitates a meaningful transfer of knowledge, thereby enhancing the learning ability.

This theory was put to the test using deep fully connected networks (FCNs). The researchers briefly aligned the networks with a guide network using random noise before the actual training. The results were astonishing–notorious for overfitting, these networks became more stable, avoided the usual FCN pitfalls, displayed lower training losses and improved their performances. “It’s impressive that we could use representational similarity to make these traditionally ‘crappy’ networks actually work,” says Vighnesh Subramaniam ’23, MEng ’24, a PhD student in MIT’s Department of Electrical Engineering and Computer Science and lead author of the study.

Game-Changer for Neural Networks

The study reveals that guidance, unlike knowledge distillation, doesn’t falter when using an untrained teacher network. As guidance relies on the network’s internal structure which carries valuable architectural biases. These biases function like a compass, directing the network towards better learning paths.

But the implications of this research don’t stop at performance improvements. It suggests that the success of a network may be more dependent on its starting point in the learning space than the data it’s trained on. By pairing networks with a guide, the impact of architectural design can be isolated from learned experience. This introduction of guidance provides a new perspective for evaluating network structures’ contribution to effective learning. It also gives scientists a way to understand the differences between architectures, helping refine theories about neural network optimization and identifying which components are consequential for learning.

The highlight, however, is that no network is beyond redemption. Even those once branded as ineffective can be brought up to par with modern standards through guidance. Currently, the CSAIL team is studying which architectural elements are major contributors to these improvements, intending to influence future neural network design.

“It’s generally assumed that different neural network architectures have particular strengths and weaknesses,” noted Leyla Isik, an assistant professor of cognitive science at Johns Hopkins University, who was not involved in the study. “This exciting research shows that one type of network can inherit the advantages of another architecture, without losing its original capabilities.”

The research, a collective effort of Subramaniam and his MIT CSAIL collaborators, was supported by organizations such as the Center for Brains, Minds, and Machines, the National Science Foundation, the MIT-IBM Watson AI Lab, and the U.S. Department of the Air Force Artificial Intelligence Accelerator. Their game-changing findings were recently presented at the Conference and Workshop on Neural Information Processing Systems (NeurIPS).

Read the original article from MIT News here: https://news.mit.edu/2025/guided-learning-lets-untrainable-neural-networks-realize-their-potential-1218

Enter Justin Reich, associate professor at MIT’s Comparative Media Studies/Writing program. Reich has written a lifeboat for educators in the form of a guidebook titled “A Guide to AI in Schools: Perspectives for the Perplexed”, through the MIT Teaching Systems Lab. The publication aims to hold the hands of K–12 educators as they sail the unexplored waters of AI in education. “Over my career, I’ve aimed to be a bridge between educational research and practical insights. My goal is to simplify and share insights about complex concepts like AI”, Reich shares.

A Compass, Not a Map

In this guidebook, you won’t find strict rules or ‘best practices’. The content is diverse, featuring thought-provoking experiences and insights from over 100 students and teachers across America. Instead of pushing for a ‘one-size-fits-all’ approach, it stimulates dialogue and encourages a culture of experimentation in classrooms. Reich’s ethos aligns with this approach, advocating a humble acceptance that some strategies might work better than others—an understanding that will become clearer over time.

Important Questions and Conversations

The advent of AI has caused an influx of questions. How do we safeguard academic integrity? How do we protect student data privacy? And how does learning change when students can use AI to bypass the hard work of thinking? Reich states that we’re only starting to wrap our minds around these questions. The scenario, he points out, is akin to us sitting in the teacher’s lounge, conversing about AI, before inviting students, parents and others to join the discussion. Comparing this to the early days of aviation, the guidebook authors note that predicting what’s going to work would be premature at this stage.

Learning from History

Reich is open about the potential problems that lie ahead. “We’re finding our way in the dark,” he candidly admits, drawing parallels with past blunders with classroom technology. Reflecting on flawed strategies such as discouraging students from using Wikipedia or looking for antiquated credibility markers on websites, Reich urges caution. He espouses an evidence-based approach over sweeping assumptions about AI, advocating for trials, observation, and learning from the outcomes.

Thus, Reich’s guidebook and corresponding podcast series “The Homework Machine”, a collaboration with journalist Jesse Dukes, provide vital resources for educators navigating AI in the classroom. Reich encourages a diverse conversation around AI, involving voices from various stakeholders. This dialog has further been amplified through an hour-long radio special that aired on public radio stations nationwide, thereby reaching a wider audience.

While the demands of navigating new technology like AI can be high, Reich underlines the need for caution, collaboration, and a collective willingness to learn from our past experiences—as, unlike previous technologies, AI has seeped into students’ lives, not through traditional channels but through their smartphones. This has placed even more pressure on educators to respond without clear directions or tried-and-tested practices. But this doesn’t mean they should be denied a voice in shaping its use. As Reich says, “AI is already here. Teachers don’t get to choose whether it exists, but they should have strong input on shaping its future.”

For more insights and to join this ongoing conversation, you can visit the original article from MIT News: Helping K-12 Schools Navigate the Complex World of AI.

Researchers from MIT, Lawrence Berkeley National Lab, and the University of Orléans are on the case, and their latest findings bring much-needed clarity. In a study recently published in PNAS, the team managed to create computer models that mirror real-world experiments with impressive accuracy. That might sound technical, but it’s a big deal: having models that we can trust makes designing safer storage methods a whole lot easier.

What makes the research groundbreaking is its connection to the Mont Terri laboratory in Switzerland, a site that’s been at the forefront of nuclear waste studies since the 1990s. The scientists focused on Opalinus clay, a natural material thought to be ideal for locking away radioactive material. By running experiments at Mont Terri and then feeding that data into sophisticated computer simulations, they were able to see how artificial barriers (like cement) interact with clay deep underground, not just for months or years, but potentially for centuries.

This progress owes a lot to new tools. The latest computer model, called CrunchODiTi, is a notable upgrade over older programs. Unlike previous versions, CrunchODiTi can account for the tiny electrical charges in clay minerals. Those details matter because they help predict how radioactive particles might move—or stay put—over very long periods of time.

The team zeroed in on a tiny, crucial “skin” layer, just a centimeter thick, where cement and clay meet. This interface plays an outsized role in the slow migration of radioactive elements through the subsurface. Data from a 13-year experiment at Mont Terri gave the researchers a rare, long-term look at changes in this skin zone. Their computer model matched the physical observations, lending real confidence that they’re on the right track.

Dauren Sarsenbayev, the study’s lead author, finds it remarkable to observe these changes unfold. Seeing how the intersection between cement and clay evolves over time helps bridge the gap between theory and reality. The team’s observations also support long-standing ideas about how minerals build up and how the material’s porosity shifts, both of which matter for waste containment over the long term.

What does all this mean for nuclear waste management? For starters, these more accurate models could replace the outdated simulations currently used to judge the safety of nuclear repositories. That’s critical if the U.S.—or any country—wants to move forward with building a permanent disposal site. The models also have the flexibility to assess different rock types, including salt formations, stretching their usefulness far into the future.

Looking ahead, the researchers plan to refine their simulations even more, possibly integrating machine learning to speed up the process. They’re also eager to gather more data from ongoing experiments, inching closer to a time when storing nuclear waste is a solved problem rather than a looming risk.

For Sarsenbayev and the team, the ultimate goal is straightforward: create a scientifically solid, trustworthy solution for nuclear waste that decision-makers—and the public—can have faith in. As he puts it, their work sits at the intersection of science, systems, and society: a genuinely collaborative approach to one of the most daunting engineering challenges of our era.

Read the full story at MIT News.

When people talk about Artificial General Intelligence (AGI), it’s easy to imagine it as a distant, high-tech concept only relevant to the world’s most advanced cities. But in reality, AGI—and its younger sibling, everyday AI technology—has the potential to make the biggest difference where resources have traditionally lagged behind. My time working alongside iCog Labs in Ethiopia, a company co-founded by Ben Goertzel and Getnet Aseffa, has shown me just how much promise and how many unique hurdles there are when nurturing AI projects in developing nations. There’s enormous potential here, but also a lot to learn.

How AI Can Help Close Education Gaps

AI could be a real game changer for education—especially in places where shortages of teachers, textbooks, and even basic infrastructure are everyday challenges. And if we dig deeper, the obstacles aren’t just about missing blackboards: many children in countries like Ethiopia are expected to learn in languages they barely understand, using materials that don’t reflect their real lives or cultures. The answer isn’t just technology for technology’s sake; instead, it’s about solutions born out of local insights, creativity, and hands-on partnerships.

UNESCO estimates that about 40% of children worldwide go to school in a language they don’t fully grasp, which means they’re missing out not just on words, but on all kinds of opportunities. This is where AI-powered language tools shine. While there’s often little profit motivation for global tech giants to support small or under-resourced language communities, targeted projects fill the gap. Alongside Curious Learning, we’ve built AI-backed reading apps catering to local Ethiopian languages, now used by over 85,000 people. They’re simple, low-resource tools taming the hardest linguistic challenges faced by so many kids.

To dig even deeper, we’ve launched Leyu—a decentralized data collection platform—focused on gathering language samples directly from far-flung villages. The data we collect trains AI translation systems, so another under-represented language doesn’t get left out of tomorrow’s digital conversation.

Making Learning Exciting and Relevant

Of course, language is only part of the equation. True educational progress depends on content that means something locally. AI has a role here, too: it can generate science lessons about crops local kids see every day or math problems built around the rhythms of their own markets. This is education that clicks, because it’s finally about them.

Some of our proudest work has rolled out on the road with Digitruck—a full-blown mobile classroom built into a truck. Thanks, in part, to support from SingularityNET, Digitruck is decked out with computers and devices, and rolls into some of Ethiopia’s most remote spots. Kids test out coding and even get hands-on experiences with AI—all while staying close to their roots.

Looking back, there have also been joyful, unforgettable moments. From 2015 to 2019, the RoboSapiens program had university students programming humanoid robots to play soccer—a blend of culture and high-tech excitement that created instant buzz from Ethiopia to Kenya and Nigeria. Sadly, practical obstacles like expensive import tariffs on electronics put the program on hold, a reminder that technical ambition needs support from strong policy and infrastructure.

Rooting AI in Real Community Needs

All this energy must be matched by caution and care. As AI makes its way into classrooms and daily life, it’s crucial not to let these tools replace essential skills or undermine teachers. Rather, the aim is to help educators do what they do best—foster curiosity and independent thinking—while supporting them with new resources.

It’s not about one-size-fits-all rules. Instead, the communities themselves need to guide what AI does for them. That kind of meaningful, grassroots engagement is what will ensure AI enhances education, rather than becoming just another barrier.

Equitable, responsible AI in education isn’t simple or quick. It means working side-by-side with local leaders, keeping cultural nuances front and center, and planning for ethical risks. But it’s worth every bit of effort. When a child in rural Ethiopia codes for the very first time on a Digitruck laptop, that’s when the big themes of equity and empowerment stop being buzzwords—and become real, lived experiences.

Original article on Unite.AI

The Unseen Shortcomings in Healthcare Data and the Role of AI Education

Wonderfully highlighted by Leo Anthony Celi, an accomplished physician and senior research scientist at MIT’s Institute for Medical Engineering and Science, this oversight in AI education may create issues down the line. In a recent article, he explains how data bias, especially in clinical data primarily collected from white males, might lead to inefficiencies in AI systems when applied to more diverse populations. For instance, pulse oximeters often overestimate oxygen saturation levels in people of color due to their underrepresentation in clinical trials. This is only the tip of the iceberg, with countless more cases where medical equipment and data systems overlook population diversity, producing skewed results and potentially harmful decisions.

Another crucial issue orbits around the use of electronic health records (EHRs) as a basis for AI models. Although an essential part of medicine, EHRs were never intended to serve as learning systems, and they are ripe with inconsistencies and biases. Celi, however, is not all doom and gloom, advocating for ingenious ways of responsibly using the existing data instead of replacing the entire EHR infrastructure—something currently not feasible. Thankfully, innovative approaches such as transformer models are being explored to understand correlations between lab results, vital signs, and treatments better. This fascinating approach could help lessen the impact of missing or biased data, often affected by social determinants of health and implicit provider biases.

Addressing the Imperfections and Maximizing Learning

The challenges become apparent when you consider Celi’s experiences teaching AI in healthcare. Since starting in 2016, his MIT team realized that students were being taught to optimize models in terms of statistical performance rather than question the data’s integrity. A review of 11 online courses revealed the scale of this problem; only five courses talked about data bias and a mere two offered substantial discussions on the subject. As AI continues to establish its footprint in healthcare, the onus lies with educators to ensure students can not only build models but also scrutinize the data fueling them. Bridging this divide will require a shift in focus from solely model building to also understanding the data – an area Celi believes should account for at least half the course content.

One initiative helping to solve this puzzle is the MIT Critical Data consortium. They’ve been hosting international datathons since 2014. These sessions unite clinicians, data scientists, and healthcare professionals to collaboratively examine local datasets, aiming to understand health and disease within the unique cultural and systematic context of each region. These collaborations inspire an environment where critical thinking organically thrives.

Embracing the imperfections in data can also be a step towards improvement, albeit a challenging one. A good example is the MIMIC database, which took over 10 years to formulate a usable schema, largely due to users acknowledging and pointing out its flaws. An insightful reminder from Celi at this point is that even without all the answers, inspiring people to start asking the right questions can be a game-changer. As students and researchers engage with AI development in healthcare, they need to remain aware of its transformative potential and the ethical responsibilities that come with that.

For a more in-depth discussion on this subject with Leo Anthony Celi, visit MIT News.