In some parts of the deep ocean, it can look like it’s snowing. This “marine snow” is the dust and detritus that organisms slough off as they die and decompose. Marine snow can fall several kilometers to the deepest parts of the ocean, where the particles are buried in the seafloor for millennia.

Now, researchers at MIT and their collaborators have found that as marine snow falls, tiny hitchhikers may limit how deep the particles can sink before dissolving away. The team shows that when bacteria hitch a ride on marine snow particles, the microbes can eat away at calcium carbonate, which is an essential ballast that helps particles sink.

The findings, which appear this week in the Proceedings of the National Academy of Sciences, could explain how calcium carbonate dissolves in shallow layers of the ocean, where scientists had assumed it should remain intact. The results could also change scientists’ understanding of how quickly the ocean can sequester carbon from the atmosphere.

Marine snow is a main vehicle by which the ocean stores carbon. At the ocean’s surface, phytoplankton absorb carbon dioxide from the atmosphere and convert the gas into other forms of carbon, including calcium carbonate — the same stuff that’s found in shells and corals. When they die, bits of phytoplankton drift down through the ocean as marine snow, carrying the carbon with them. If the particles make it to the deep ocean, the carbon they carry can be buried and locked away for hundreds to thousands of years.

But the new study suggests bacteria may be working against the ocean’s ability to sequester carbon. By eroding the particles’ calcium carbonate, bacteria can significantly slow the sinking of marine snow. The more they linger, the more likely the particles are to be respired quickly, releasing carbon dioxide into the shallow ocean, and possibly back into the atmosphere.

“What we’ve shown is that carbon may not sink as deep or as fast as one may expect,” says study co-author Andrew Babbin, an associate professor in the Department of Earth, Atmospheric and Planetary Sciences and a mission director at the Climate Project at MIT. “As humanity tries to design our way out of the problem of having so much CO2 in the atmosphere, we have to take into account these natural microbial mechanisms and feedbacks.”

The study’s primary author is Benedict Borer, a former MIT postdoc who is now an assistant professor of marine and coastal sciences at the Rutgers School of Environmental and Biological Sciences; co-authors include Adam Subhas and Matthew Hayden at the Woods Hole Oceanographic Institution and Ryan Woosley, a principal research scientist at MIT’s Center for Sustainability Science and Strategy.

Losing weight

Marine snow acts as the ocean’s main “biological pump,” the process by which the ocean pulls carbon from the surface down into the deep ocean. Scientists estimate that marine snow is responsible for drawing down billions of tons of carbon each year. Marine snow’s ability to sink comes mainly from minerals such as calcium carbonate embedded within the particles. The mineral is a dense ballast that weighs down the particle. The more calcium carbonate a particle has, the faster it sinks.

Scientists had assumed based on thermodynamics that calcium carbonate should not dissolve within the ocean’s upper layers, given the general temperature and pH conditions in the surface ocean. Any calcium carbonate that is bound up in marine snow should then safely sink to depths greater than 1,000 meters without dissolving along the way.

But oceanographers have long observed signs of dissolved calcium carbonate in the upper layers of the ocean, suggesting that something other than the ocean’s macroscale conditions was dissolving the mineral and slowing down the ocean’s biological pump.

And indeed, the MIT team has found that what is dissolving calcium carbonate in shallow waters is a microscale process that occurs within the immediate environment of an individual particle.

“Most oceanographers think about the macroscale, and in this instance what’s happening in microscopic particles is what is actually controlling bulk seawater chemistry,” Borer says. “Consequences abound for the ocean’s carbon dioxide sequestration capacity.”

A sinking sweetspot

In their new study, the researchers set up an experiment to simulate a sinking particle of marine snow and its interactions at the microscale. The team synthesized particles similar to marine snow that they made from varying concentrations of calcium carbonate and bacteria — organisms that are often found feasting on the particles in the ocean.

“The ocean is a fairly dilute medium with respect to organic matter,” Babbin says. “So organisms like bacteria have to search for food. And particles of marine snow are like cheeseburgers for bacteria.”

The team designed a small microfluidic chip to contain the particles, and flowed seawater through the chip at various rates to simulate different sinking speeds in the ocean. Their experiments revealed that whenever particles hosted any bacteria, they also rapidly lost some calcium carbonate, which dissolved into the surrounding seawater. As bacteria feed on the particles’ organic material, the microbes excrete acidic waste products that act to dissolve the particles’ inorganic, ballasting calcium carbonate.

The researchers also found that the amount of calcium carbonate that dissolves depends on how fast the particles sink. They flowed seawater around the particles at slow, intermediate, and fast speeds and found that both slow and fast sinking limit the amount of calcium carbonate that’s dissolved. With slow sinking, particles don’t receive as much oxygen from their surroundings, which essentially suffocates any hitchhiking bacteria. When particles sink quickly, bacteria may be sufficiently oxygenated, but any waste products that they produce can be easily flushed away before they can dissolve the particles’ calcium carbonate.

At intermediate speeds, there is a sweet spot: Bacteria are sufficiently oxygenated and can also build up enough waste, enabling the microbes to efficiently dissolve calcium carbonate.

Overall, the work shows that bacteria can have a significant effect on marine snow’s ability to sink and sequester carbon in the deep ocean. Bacteria can be found everywhere, and particularly in the shallower ocean regions. Even if macroscale conditions in these upper layers should not dissolve calcium carbonate, the study finds bacteria working at the microscale most likely do.

The findings could explain oceanographers’ observations of dissolved calcium carbonate in shallow ocean regions. They also illustrate that bacteria and other microbes may be working against the ocean’s natural ability to sequester carbon, by dissolving marine snow’s ballast and slowing its descent into the deep ocean. As humans consider climate solutions that involve enhancing the ocean’s biological pump, the researchers emphasize that bacteria’s role must be taken into account.

“Insights from this work are vital to predict how ecosystems will respond to marine carbon dioxide removal attempts, and overall how the oceans will change in response to future climate scenarios,” says Benedict Borer, who carried out the study’s experiments as a postdoc in MIT’s Department of Earth, Atmospheric and Planetary Sciences.

This work was supported, in part, by the Simons Foundation, the National Science Foundation, and the Climate Project at MIT.

When we learn a new skill, the brain has to decide — cell by cell — what to change. New research from MIT suggests it can do that with surprising precision, sending targeted feedback to individual neurons so each one can adjust its activity in the right direction.

The finding echoes a key idea from modern artificial intelligence. Many AI systems learn by comparing their output to a target, computing an “error” signal, and using it to fine-tune connections within the network. A long-standing question has been whether the brain also uses that kind of individualized feedback. In an open-access study published in the Feb. 25 issue of the journal Nature, MIT researchers report evidence that it does.

A research team led by Mark Harnett, a McGovern Institute for Brain Research investigator and associate professor in the Department of Brain and Cognitive Sciences at MIT, discovered these instructive signals in mice by training animals to control the activity of specific neurons using a brain-computer interface (BCI). Their approach, the researchers say, can be used to further study the relationships between artificial neural networks and real brains, in ways that are expected to both improve understanding of biological learning and enable better brain-inspired artificial intelligence.

The changing brain

Our brains are constantly changing as we interact with the world, modifying their circuitry as we learn and adapt. “We know a lot from 50 years of studies that there are many ways to change the strength of connections between neurons,” Harnett says. “What the field really lacks is a way of understanding how those changes are orchestrated to actually produce efficient learning.”

Some actions — and the neural connections that enable them — are reinforced with the release of neuromodulators like dopamine or norepinephrine in the brain. But those signals are broadcast to large groups of neurons, without discriminating between cells’ individual contributions to a failure or a success. “Reinforcement learning via neuromodulators works, but it’s inefficient, because all the neurons and all the synapses basically get only one signal,” Harnett says.

Machine learning uses an alternative, and extremely powerful, way to learn from mistakes. Using a method called back propagation, artificial neural networks compute an error signal and use it to adjust their individual connections. They do this over and over, learning from experience how to fine-tune their networks for success. “It works really well and it’s computationally very effective,” Harnett says.

It seemed likely that brains might use similar error signals for learning. But neuroscientists were skeptical that brains would have the precision to send tailored signals to individual neurons, due to the constraints imposed by using living cells and circuits instead of software and equations. A major problem for testing this idea was how to find the signals that provide personalized instructions to neurons, which are called vectorized instructive signals. The challenge, explains Valerio Francioni, first author of the Nature paper and a former postdoc in Harnett’s lab, is that scientists don’t know how individual neurons contribute to specific behaviors.

“If I was recording your brain activity while you were learning to play piano,” Francioni explains, “I would learn that there is a correlation between the changes happening in your brain and you learning piano. But if you asked me to make you a better piano player by manipulating your brain activity, I would not be able to do that, because we don’t know how the activity of individual neurons map to that ultimate performance.”

Without knowing which neurons need to become more active and which ones should be reined in, it is impossible to look for signals directing those changes.

Understanding neuron function

To get around this problem, Harnett’s team developed a brain-computer interface task to directly link neural activity and reward outcome — akin to linking the keys of the piano directly to the activity of single neurons. To succeed at the task, certain neurons needed to increase their activity, whereas others were required to decrease their activity.

They set up a BCI to directly link activity in those neurons — just eight to 10 of the millions of neurons in a mouse’s brain — to a visual readout, providing sensory feedback to the mice about their performance. Success was accompanied by delivery of a sugary reward.

“Now if you ask me, ‘How does the mouse get more rewards? Which neuron do you have to activate and which neuron do you have to inhibit?’ I know exactly what the answer to that question is,” says Francioni, whose work was supported by a Y. Eva Tan Fellowship from the Yang Tan Collective at MIT.

The scientists didn’t know the exact function of the particular neurons they linked to the BCI, but the cells were active enough that mice received occasional rewards whenever the signals happened to be right. Within a week, mice learned to switch on the right neurons while leaving the other set of neurons inactive, earning themselves more rewards.

Francioni monitored the target neurons daily during this learning process using a powerful microscope to visualize fluorescent indicators of neural activity. He zeroed in on the neurons’ branching dendrites, where the appropriate feedback signals have long been suspected to arrive. At the same time, he tracked activity in the parent cell bodies of those neurons. The team used these data to examine the relationship between signals received at a neuron’s dendrites and its activity, as well as how these changed when mice were rewarded for activating the right neurons or when they failed at their task.

Vectorized neural signals

They concluded that the two groups of neurons whose activity controlled the BCI in opposite ways, also received opposing error signals at their dendrites as the mice learned. Some were told to ramp up their activity during the task, while others were instructed to dial it down. What’s more, when the team manipulated the dendrites to inhibit these instructive signals, mice failed to learn the task. “This is the first biological evidence that vectorized [neuron-specific] signal-based instructive learning is taking place in the cortex,” Harnett says.

The discovery of vectorized signals in the brain — and the team’s ability to find them — should promote more back-and-forth between neuroscientists and machine learning researchers, says postdoc Vincent Tang. “It provides further incentive for the machine learning community to keep developing models and proposing new hypotheses along this direction,” he says. “Then we can come back and test them.”

The researchers say they are just as excited about applying their approach to future experiments as they are about their current discovery.

“Machine learning offers a robust, mathematically tractable way to really study learning. The fact that we can now translate at least some of this directly into the brain is very powerful,” Francioni says.

Harnett says the approach opens new opportunities to investigate possible parallels between the brain and machine learning. “Now we can go after figuring out, how does cortex learn? How do other brain regions learn? How similar or how different is it to this particular algorithm? Can we figure out how to build better, more brain-inspired models from what we learn from the biology?” he says. “This feels like a really big new beginning.” 

In high-stakes settings like medical diagnostics, users often want to know what led a computer vision model to make a certain prediction, so they can determine whether to trust its output.

Concept bottleneck modeling is one method that enables artificial intelligence systems to explain their decision-making process. These methods force a deep-learning model to use a set of concepts, which can be understood by humans, to make a prediction. In new research, MIT computer scientists developed a method that coaxes the model to achieve better accuracy and clearer, more concise explanations.

The concepts the model uses are usually defined in advance by human experts. For instance, a clinician could suggest the use of concepts like “clustered brown dots” and “variegated pigmentation” to predict that a medical image shows melanoma.

But previously defined concepts could be irrelevant or lack sufficient detail for a specific task, reducing the model’s accuracy. The new method extracts concepts the model has already learned while it was trained to perform that particular task, and forces the model to use those, producing better explanations than standard concept bottleneck models.

The approach utilizes a pair of specialized machine-learning models that automatically extract knowledge from a target model and translate it into plain-language concepts. In the end, their technique can convert any pretrained computer vision model into one that can use concepts to explain its reasoning.

“In a sense, we want to be able to read the minds of these computer vision models. A concept bottleneck model is one way for users to tell what the model is thinking and why it made a certain prediction. Because our method uses better concepts, it can lead to higher accuracy and ultimately improve the accountability of black-box AI models,” says lead author Antonio De Santis, a graduate student at Polytechnic University of Milan who completed this research while a visiting graduate student in the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT.

He is joined on a paper about the work by Schrasing Tong SM ’20, PhD ’26; Marco Brambilla, professor of computer science and engineering at Polytechnic University of Milan; and senior author Lalana Kagal, a principal research scientist in CSAIL. The research will be presented at the International Conference on Learning Representations.

Building a better bottleneck

Concept bottleneck models (CBMs) are a popular approach for improving AI explainability. These techniques add an intermediate step by forcing a computer vision model to predict the concepts present in an image, then use those concepts to make a final prediction.

This intermediate step, or “bottleneck,” helps users understand the model’s reasoning.

For example, a model that identifies bird species could select concepts like “yellow legs” and “blue wings” before predicting a barn swallow.

But because these concepts are often generated in advance by humans or large language models (LLMs), they might not fit the specific task. In addition, even if given a set of pre-defined concepts, the model sometimes utilizes undesirable learned information anyway, which is a problem known as information leakage.

“These models are trained to maximize performance, so the model might secretly use concepts we are unaware of,” De Santis explains.

The MIT researchers had a different idea: Since the model has been trained on a vast amount of data, it may have learned the concepts needed to generate accurate predictions for the particular task at hand. They sought to build a CBM by extracting this existing knowledge and converting it into text a human can understand.

In the first step of their method, a specialized deep-learning model called a sparse autoencoder selectively takes the most relevant features the model learned and reconstructs them into a handful of concepts. Then, a multimodal LLM describes each concept in plain language.

This multimodal LLM also annotates images in the dataset by identifying which concepts are present and absent in each image. The researchers use this annotated dataset to train a concept bottleneck module to recognize the concepts.

They incorporate this module into the target model, forcing it to make predictions using only the set of learned concepts the researchers extracted.

Controlling the concepts

They overcame many challenges as they developed this method, from ensuring the LLM annotated concepts correctly to determining whether the sparse autoencoder had identified human-understandable concepts.

To prevent the model from using unknown or unwanted concepts, they restrict it to use only five concepts for each prediction. This also forces the model to choose the most relevant concepts and makes the explanations more understandable.

When they compared their approach to state-of-the-art CBMs on tasks like predicting bird species and identifying skin lesions in medical images, their method achieved the highest accuracy while providing more precise explanations.

Their approach also generated concepts that were more applicable to the images in the dataset. 

“We’ve shown that extracting concepts from the original model can outperform other CBMs, but there is still a tradeoff between interpretability and accuracy that needs to be addressed. Black-box models that are not interpretable still outperform ours,” De Santis says.

In the future, the researchers want to study potential solutions to the information leakage problem, perhaps by adding additional concept bottleneck modules so unwanted concepts can’t leak through. They also plan to scale up their method by using a larger multimodal LLM to annotate a bigger training dataset, which could boost performance.

“I’m excited by this work because it pushes interpretable AI in a very promising direction and creates a natural bridge to symbolic AI and knowledge graphs,” says Andreas Hotho, professor and head of the Data Science Chair at the University of Würzburg, who was not involved with this work. “By deriving concept bottlenecks from the model’s own internal mechanisms rather than only from human-defined concepts, it offers a path toward explanations that are more faithful to the model and opens many opportunities for follow-up work with structured knowledge.”

This research was supported by the Progetto Rocca Doctoral Fellowship, the Italian Ministry of University and Research under the National Recovery and Resilience Plan, Thales Alenia Space, and the European Union under the NextGenerationEU project.