home.social

#metalearning — Public Fediverse posts

Live and recent posts from across the Fediverse tagged #metalearning, aggregated by home.social.

fetched live
  1. Ключевые основы автоматического выбора алгоритмов кластеризации мультимодальных данных на основе мета-обучения

    Задача кластеризации относится к классу «обучения без учителя» и является фундаментальным инструментом exploratory data analysis (разведочный анализ данных). В отличие от классификации, здесь отсутствует размеченный набор данных или какая-либо заведомо известная информация о нём. Алгоритм самостоятельно выявляет критерий группировки объектов, и именно в этой самостоятельности кроется ключевая проблема: отсутствие априорной разметки делает оценку результатов крайне субъективной, так как разные алгоритмы обладают уникальным индуктивным смещением (inductive bias), проецируя различные предположения о геометрии и плотности скрытых классов. Универсальный алгоритм, который подходит для всех задач, построить невозможно ( теорема Клейнберга , являющаяся следствием более общей теоремы о «бесплатном обеде»), поэтому алгоритмы кластеризации нужно подбирать и настраивать почти для каждой задачи отдельно. Задача выбора и настройки алгоритма машинного обучения является экспертной, что достаточно затратно по времени, поскольку работа выполняется человеком фактически вручную. Особую сложность представляет кластеризация мультимодальных данных. В прикладных областях, таких как биоинформатика, анализ медиа-контента, медицинская диагностика и мониторинг сложных технических объектов, данные чаще всего представлены в мультимодальном виде. Это означает, что каждый анализируемый объект одновременно описывается гетерогенными источниками информации — структурированными числовыми векторами, неструктурированными текстами, изображениями или временными рядами. Ключевой сложностью при построении систем автоматической кластеризации для мультимодальных данных является разработка эффективного способа слияния (fusion) разнородных признаковых пространств с сохранением уникальной внутренней структуры каждой модальности.

    habr.com/ru/articles/1046942/

    #Автоматическая_кластеризация #метаобучение #мультимодальные_данные #AutoML #выбор_алгоритмов #data_science #metalearning #AutoClustering #обучение_без_учителя #llm

  2. The Unfathered: A Short Story for a Long World

    There is a part of the Book of Genesis that almost no one reads aloud. It sits between the famous scenes, the garden and the flood, and it is only a list. Adam lived a number of years and begat Seth. Seth lived and begat Enosh. The text walks down the page through nine generations of fathers and sons, each man reduced to two facts, that he lived and that he made another like himself, until the line reaches Noah and the world is ready to drown. I used to skip those passages. I read one again last night, on a laptop in an apartment I am about to leave, and I understood for the first time that a genealogy is a horror story told slowly enough to be survived. […]

    bolesblogs.com/2026/06/07/the-

  3. 📚 Zhou & Schapiro show in their recent study that a gradient of complementary #learning systems can emerge through #metalearning. Their model self-organizes into subsystems that differ in timescale and representational abstraction – offering a mechanistic account of distributed #memory architectures across #cortex.

    🌍 biorxiv.org/content/10.1101/20

    #CompNeuro #Neuroscience

  4. Meta-learning lets AI generalize like humans (Lake & Baroni, 2023). Mice map environments through action (Shamash et al., 2023). Both learn—but one is alive. Does AI’s efficiency mask a fundamental gap? What’s missing: biology… or being?
    #AIvsHuman #MetaLearning #Neurobiology

    nature.com/articles/s41586-023

    cell.com/neuron/fulltext/S0896

  5. The real lock-in risk with LLMs isn't just about data, but about learned interaction patterns. Each model builds a unique understanding of how you work, think, and communicate. When using multi-LLM routers, we're essentially distributing this meta-knowledge across models. The challenge? Making this learned context portable and consistent.

    Think of it as teaching multiple assistants your preferences simultaneously vs. deep-learning with just one."

    #AI #LLM #MetaLearning #DigitalTransformation

  6. Bayesian Meta-Learning Is All You Need

    — Why is the deterministic view of meta-learning not sufficient?

    — What is the variational inference?

    — How can we design neural-based Bayesian meta-learning algorithms?

    jameskle.com/writes/bayesian-m

    #machinelearning #bayesian #metalearning #variational

  7. In-context learning is a game changer in #MetaLearning. Instead of engineering exact learning rules, we can just present episodes where learning happens, and let an auto regressive model learn all the types of learning that happen in there.

    This often produces in-context learning rules that are practically optimal in data efficiency.

    But the context length is a limitation. Often we want to learn long-term things, and keep them way past the context length limits.

    Nowadays we have lots of "infinite context length" type architectures, architectures using long-term registries for memory, and so on, which remove or mitigate the context length limit.

    Ultimately it all becomes just about data. We need lots of #data where all sorts of learning and reasoning happens, wide and deep. We need to produce this synthetically and refine it from natural sources.

    #LLMs #AI

  8. Through scaling #DeepNeuralNetworks we have found in two different domains, #ReinforcementLearning and #LanguageModels, that these models learn to learn (#MetaLearning).

    They spontaneously learn internal models with memory and learning capability which are able to exhibit #InContextLearning much faster and much more effectively than any of our standard #backpropagation based deep neural networks can.

    These rather alien #LearningModels embedded inside the deep learning models are emulated by #neuron layers, but aren't necessarily deep learning models themselves.

    I believe it is possible to extract these internal models which have learned to learn, out of the scaled up #DeepLearning #substrate they run on, and run them natively and directly on #hardware.

    This allows those much more efficient learning models to be used either as #LearningAgents themselves, or as a further substrate for further meta-learning.

    I have an #embodiment #research on-going but with a related goal and focus specifically in extracting (or distilling) the models out of the meta-models here:
    github.com/keskival/embodied-e

    It is of course an open research problem how to do this, but I have a lot of ideas!

    If you're inspired by this, or if you think the same, let's chat!

  9. If you are learning from a sequence of observations and actions, and using a probabilistic model, you should be treating the agent's actions as #causal interventions. The authors attribute the lack of causal interventions as the reason for #delusions, and also propose a way to do #MetaLearning. #MachineLearning #Causality #ReinforcementLearning

    arxiv.org/abs/2110.10819

  10. The brilliant Kristopher Torp Jensen arrived on Mastodon! @KrisJensen

    Kris is doing super cool work on complex shapes in neural codes and continual learning / meta-learning in artificial agents. He is doing his PhD with Guillaume Hennequin at CBL in Cambridge.

    #Neuroscience #RL #metalearning #ML