{
  "video_id": "ch6K00NORb8",
  "channel_slug": "mitcsail",
  "channel_handle": "mitcsail",
  "title": "Meet the Mind:  MIT Professor Sara Beery",
  "duration_seconds": 503.0,
  "url": "https://www.youtube.com/watch?v=ch6K00NORb8",
  "upload_date": "",
  "transcript": "(air whooshing)\n(dramatic music)\n- I'm Sara Beery.\nI am a new assistant professor at MIT\nand I'm also a PI at the Computer\nScience and AI Laboratory.\nAnd my new research\ngroup is really focused\nat this intersection between\nAI, biodiversity, and ecology.\nWe're interested in\ndeveloping AI techniques\nthat help us better understand\nhow species are changing\nat a global scale.\nAnd we do this across\ndifferent data modalities.\nEverything from images\nto acoustics, to sonar,\nto remote sensing data.\nAnd we combine these different modalities\nand try to interpret them,\nso that we can understand\nhow species are changing\nat a global scale.\nSo actually, I've always\nloved the natural world.\nI grew up camping with my\nfamily in the Pacific Northwest,\nbut actually my first passion was dance.\nI started training as a\nballerina very seriously.\nI got my first job offer\nwith the Atlanta Ballet\nwhen I was only 16,\nand I moved from Seattle\nto Atlanta by myself\nto join a ballet company.\nI lived in sort of like\na not super great area\nthat was kind of close to Georgia Tech.\nThey would put posters\non the telephone poles\nfor scientific talk that\nalways advertised free food,\nso I would go.\nI almost immediately found myself\nhooked on the science,\nand not specifically science writ large,\nbut it was eyeopening to me\nin terms of how science and technology\ncan be used as a force for public good.\nAnd that planted the seed.\nSo, I went on and I continued\nto do professional ballet\nfor six years.\nI loved it, but I always started,\nI started to think about\nhow I might instead go back\nafter I finished, after\nI retired from ballet,\nand become a scientist.\nAnd I ended up doing that.\nI went back.\nI retired from ballet\nand I started a career in\nelectrical engineering.\nSo one of the projects I\nactually started in my PhD\nand have brought to my lab\nhere at MIT is a project\nlooking at improving the sustainability\nof the salmon fisheries from Alaska\nall the way down through\nCanada to Northern California.\n- This project is\nworking with stakeholders\nin fishery management and conservation,\nwho are interested in monitoring\nthe number of salmon migrating every year.\nAnd to do this, they\ndeploy sonar video cameras\nat rivers around the world\nand they have technicians deployed there\nthroughout the migration\nseason, counting how many salmon\nare moving by the cameras every day.\nWe're working on developing\ncomputer vision algorithms\nto make the process more efficient\nand more accurate for them.\nSo most of these sonar camera deployments\nare out in the field with\nlimited connectivity and power.\nA solution that we've been working on\nfor deploying these algorithms\nuses edge computing.\nIt's basically a low power,\nlow cost Linux computer\nthat can be deployed out in\nthe field and powered by solar.\nAnd we can deploy the\nmodel on this device.\nAnd we have a software\napplication that runs,\nthat allows technicians to\nanalyze all of their video.\nThe way that the algorithm works\nis that each individual fish\nhas different colored box drawn around it.\nIt has a length estimation.\nAnd then at the bottom,\nwe say how many fish\nhave gone left or right,\nand they, based on the\norientation of the camera,\nthen change this to upstream\nor downstream counts.\n- So our interests are really,\nfrom a research perspective,\nhow do you effectively use human time,\nhuman decisions throughout this process,\nso that you can maintain the\nreliability of the system,\nbut also make it reasonably deployable.\nThe next project I wanna talk about\nis a little more open-ended.\nAnd here, what we're interested in\nis as we start collecting\nthese large repositories of data\nfrom different data modalities\nlike the fish sonar,\nbut actually specifically in this project,\ndata collected from passionate\nvolunteers around the world\nwho will take photos of plants and animals\nthat then become species\noccurrence records.\nAnd the way that they do that,\none mechanism they do that,\nis through a platform called iNaturalist.\n- We're using embedding\nmodels, which are AI models\nthat are trained to find\nthe semantic relationships\nbetween images and text.\nSo, we're able to pre-compute\nthese sort of semantically\nmeaningful embeddings\non all of our 5 million images.\nSo when you're trying to search through\nand you give us a piece of text\ndescribing what you're looking for,\nwe can, in less than\na second, look through\nand compare with all the\nmillions of images we have\nto give you exactly the closest\nnatural you're looking for.\nWe hope that our dataset\nwill provide the foundation\nand the motivation for\nresearchers in the field\nto actually work on building models\nthat are not just generally good,\nbut actually useful for scientists.\nOne great example is California condors.\nThis is a species that\nalmost went extinct in 1982.\nThere were only 22 individuals left.\nBut because of a successful\nbreeding program,\nnow there's a stable\npopulation of hundreds.\nAnd scientists still want to track them\nto monitor the population\nand understand where they\nare and what they're doing.\nSo, all California\ncondors have tags on them\nwith a color and a few symbols\nthat identify who they are.\nAnd so, scientists can't\nbe everywhere all at once.\nSo, it's been really difficult\nto be able to successfully monitor them\nand see where they're going.\nBut now with citizen\nscience and regular people\ntaking pictures of these\nbirds they find interesting\nand uploading them with geotag locations,\nwe can automatically mine this information\nwith these vision language\nmodels that understand\nboth text and images to\npinpoint all of their locations\nand track them in a way\nthat we've never been able to do before.\n- So we can try to design systems\nthat enable lots of scientists\nto do diverse science efficiently\nand hopefully opening up whole new avenues\nfor scientific discovery.\nWhen I was in grad school,\nI was passionately working\non developing these tools\nfor colleges in the field,\nbut I realized I hadn't\nhad that experience\nof actually trying to collect data\nand deploy monitoring\nsystems in the field myself.\nSo, I went out of my way\ntowards the end of my PhD\nto actually start curating\nsome of that experience.\nI really found it invaluable.\nMy first version of\nthis, I planned a survey\nusing static cameras in\nLaikipia County and Kenya,\nsort of north of Mount Kenya,\nwhere we were trying to study\nan endangered species of Grevy's zebra.\nAnd I went out and I placed\nall of those cameras myself,\nhiking out sometimes miles into the forest\nwith a ranger with a gun,\nbecause you're dealing\nwith very dangerous large\nspecies out in the Savannah.\nAnd really understanding\nhow much time and effort\ngoes into collecting\nevery single data point,\nit really gives you a different mentality\nabout how to value that data.\nBut not only that,\nthere are things that you don't\neven realize are challenges\nif you're just sitting behind\nyour computer writing code\nthat actually really\nreframe the priorities\nI think I have in my research.\nI found that experience\nincredibly rewarding\nbut also incredibly informative,\nand it's now something I\nreally seek to continue to do\nin my lab here.\nI want to require all of my\ngrad students to go to the field\nto be working much more hands\non with specific projects.\nMaybe not for every project,\nbut I think it's really valuable context.\nAnd it does actually\nmean that often our group\nand my own research, we've\nbrought new challenges\nto the machine learning\ncommunity, framed problems\nin ways that potentially they\nhadn't before seeing them.\nAnd I think it's really that\napplication driven innovation\nthat really helps us move the needle on\nin terms of what machine\nlearning needs to do.\nOne of the things I'm really\nexcited about moving forward\nis actually starting to bridge the gaps\nbetween these different data modalities\nthat are being collected globally.\nCombining these\nhyperspectral imaging systems\nfrom space with drones,\ncapturing individual animals\nmoving through the environment,\nwith static networks of\nbioacoustic sensors photos\ncollected by passionate human volunteers,\nactually sharing information\nacross these modalities of data.\nSo that we can get a\nmuch more nuanced picture\nof the tree of life.\nI think if I had any advice\nfor anyone who's excited to\nget started in this field,\nit would be to just talk to people.\nAnd there is increasingly\nthese amazing communities resources\nthat are coming together.\nI run a Slack channel\nthat has 2,500 global researchers on it,\njust working on AI for conservation.\nGetting onto these\nplatforms, meeting people,\nasking questions, putting\nyourself out there,\nbut learning about sort of all\nof the different diverse ways\nthat people are sort of\ntackling these problems.\nI would love to see this continue to grow.",
  "transcript_chars": 8712,
  "ingested_at": "2026-05-15T10:56:26.738618+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 11142,
    "like_count": 335,
    "channel_id": "UCBpxspUNl1Th33XbugiHJzw",
    "categories": [
      "Science & Technology"
    ],
    "tags": [
      "MIT",
      "Massachusetts Institute of Technology",
      "CSAIL"
    ]
  }
}