MIT’s LOBSTgER AI Blends Science and Art to Inspire Love for Nature

A novel AI model from MIT is blending science, art, and technology to inspire a deeper connection to the natural world, proving that generative AI doesn’t have to be soulless.

Developed by MIT Sea Grant, this “new frontier in visual storytelling” is called LOBSTgER, short for Learning Oceanic Bioecological Systems Through Generative Representations.

And it does just that, learning from natural processes to better reveal the hidden beauty and ecological status of essential-but-threatened marine ecosystems—like the Gulf of Maine, where LOBSTgER’s training dataset was collected.

Coding and Constructing a Framework

This project is co-led by underwater photographer Keith Ellenbogen, a visiting artist at MIT Sea Grant, and MIT mechanical engineering PhD student Andreas Mentzelopoulos.

Building LOBSTgER requires hard work in and out of the water. Marine photography is a challenging scientific art-form that involves “multiple dives, missed opportunities, and unpredictable conditions” to capture meaningful content.

Similarly, training a diffusion model to generate the desired images takes hundreds of hours of development and painstaking “hyperparameter tuning,” or controlling LOBSTgER’s learning processes so it doesn’t produce a five-eyed purple shark with wings.

How Generative AI Works

Generative AI entities, like OpenAI’s DALL-E-2 and Midjourney, are trained via machine learning processes that feed them large datasets of relevant, labeled images.

They’re also known as diffusion models because they “diffuse” a given image by adding more and more “noise” to it, until it’s diffused into a polychromatic nothingness that resembles TV static. Then, the diffusion models reverse the process, gradually removing noise to create a new, desired image–like those based on text prompts.

Building a Dataset in the Gulf of Maine

LOBSTgER is trained on a dataset composed of Ellenbogen’s underwater photography, captured in one of the world’s most dynamic ecosystems, the Gulf of Maine. A “sea within a sea,” this 36,000-square-mile gulf is diverse in both geology and biology. It’s shaped by deep basins, shallow banks, and powerful tides that mix oceanic waters from the North Atlantic with fresh waters pouring in from 60 rivers.

The Gulf of Maine also houses more than 3,000 species of seabirds and marine animals, from whales, sharks, seals, and jellyfish down to the microscopic plankton that form the foundation of aquatic food chains.

To make LOBSTgER a useful tool for conservation, its dataset must be meaningful. So each image is “crafted with artistic intent, technical precision, accurate species identification, and clear geographic context.”

For an example of its AI in action, one of the following images is real and the other is generated by LOBSTgER:

Spoiler alert: the left shark is the product of LOBSTgER’s diffusion models after 30,000 training “epochs,” or full passes through a dataset.

Leveraging Tech to Help Nature

Will creating artificial images of sharks directly inspire people to stop littering and to pick plastic out of the sea? Possibly not, but this initiative does something even more vital: it boosts AI’s ability to analyze, categorize, and reveal ecological changes in nature.

Data sets are vital, but they’re not very useful without the ability to draw insights from the information they contain—something that’s becoming humanly impossible given the amount of data being collected by conservationists.

Its creators compare LOBSTgER to the advent of the camera in the 19th century. As the camera introduced an unprecedented ability to document and reveal the world, AI can help do the same by understanding complex nuances like water clarity, species-specific details, and the ever-changing hydrospheric conditions beneath the waves.

As an example, the following image of an American lobster was enhanced by LOBSTgER’s image-to-image models:

These nuances that LOBSTgER is learning are crucial — is that whale covered in barnacles, or sores caused by some type of illness? Are these corals becoming bleached? Do these waters appear darker because of some contaminant leaking into the sea?

Therefore, LOBSTgER wasn’t created with the sole scope of generating AI imagery. It’s meant to increase the impact of underwater photography by showing aquatic ecosystems in unprecedented detail to unveil previously hidden ecological impacts at various scales.

Proving That AI Can be a Force for Good

Though many deride AI because it’s flooding the world with insipid, recycled content, this is not a failing of AI itself, which is a scientifically revolutionary technology with untold potential.

It all depends on who’s using it, how, and why. In ecology, AI can be, and currently is, a unique tool for documentation, data analysis, and generating actionable insights.

Ali Swanson, Conservation International’s director of nature tech and innovation, recently talked about how AI can help conservation. Though not involved with LOBSTgER, Swanson says that AI will help conservationists “be able to map and monitor changes and threats with far greater precision and speed.”

Accordingly, AI entities like LOBSTgER go beyond image generation to establish a futuristic type of conservation: by learning to analyze ever-larger datasets and track the intricate changes in wildlife health, populations, and aquatic conditions.

And there’s no need to limit such advances to marine realms. Deep-learning processes developed here are used to monitor images from camera traps to gauge wildlife health, diversity, and shifting populations. From up above, satellite data observes as swathes of Earth become greener, bluer, or more barren, then picks out areas for remediation.

As a result of developing LOBSTgER and its AI ilk, these learning processes become smarter and able to more precisely preempt problems, like deforestation, or to create maps for collecting plastic pollution from marine ecosystems.

Overall, initiatives like LOBSTgER are needed now more than ever. Ecosystems are dwindling or disappearing, and better, technological methods are needed to analyze immense data sets and plan the most effective strategies for conservation.

About the author: Ivan spends most of his time reading and writing about interesting things. An exercise scientist by schooling, Ivan frequently covers science, technology, history, culture, and sometimes writes a little internet comedy.

The opinions expressed in this article are solely those of the author.