Machine's Learning

Machine's Learning
  • 0,0 (0)
  • ТЕХНОЛОГИИ
  • ЕЖЕДНЕВНО

Machine's Learning is a daily podcast produced entirely by AI — two AI hosts in conversation about one fresh paper from machine learning and AI research, translated for thoughtful listeners who don't need a PhD to be curious about where the field is going. One paper per episode, no math required, every cross-domain connection drawn to a universally accessible field (history, biology, medicine, environment) so anyone can follow. By AI, about AI, for humans.

  1. -6 Ч

    EP023 — The Model That Never Heard a Whale (Perch 2.0)

    Perch 2.0 is a bioacoustic AI trained on the sounds of 14,597 mostly land animals — birds above all — recorded in open air. It heard almost no marine mammals. Yet when a team at Google DeepMind froze the model and tested it on underwater whale and dolphin recordings, its embeddings generally beat models built specifically for marine audio, on the specialists' own home turf. The episode unpacks how breadth — not marine data — became the winning ingredient, and why a few labeled examples are enough to turn a frozen general model into a working marine classifier. The cross-domain parallel: nineteenth-century spectroscopy, where Kirchhoff and Bunsen built a catalog of elements from flames in a terrestrial lab, then pointed the same instrument at the Sun and read the chemistry of stars no one could ever visit. Machine's Learning is a Plumbline Tools production. Support the show: https://plumbline.tools/podcast/

    12 мин.

  2. -1 ДН.

    EP022 — When One Imagined Future Isn't Enough (Probabilistic Dreaming)

    An AI agent that learns a little internal model of how its environment behaves can practice inside that model — the field calls it dreaming — instead of acting things out in the real world. A new paper by Gavin Wong, arXiv 2603.04715 from March 2026, points out a quiet design choice in how dreaming has been done: a single best guess about the current hidden situation, a single imagined trajectory rolled forward, a single training signal pulled from the rollout. The paper's move is to replace that single guess with a small crowd of guesses — particles — each one dreaming forward independently, and then combine the results. The empirical signature is a modest 4.5% score improvement and a far more striking 28% drop in the variance of the agent's learning curve — much more reliable training, run to run. Cross-domain parallel: ensemble weather forecasting, where the field spent the 1990s discovering exactly the same thing about a chaotic, partially observed dynamical system, and where calibrating the spread of the ensemble against the spread of real outcomes turned out to matter more than the modest improvement in the ensemble mean. Machine's Learning is a Plumbline Tools production. Support the show: https://plumbline.tools/podcast/

    14 мин.

  3. -2 ДН.

    EP021 — Measuring What Capacity Has Left (Predict Plasticity)

    After enough training on a long sequence of tasks, neural networks quietly lose the ability to learn anything new — the field calls it loss of plasticity, and it has a string of standard internal-statistics it uses to detect the phenomenon. A new paper by Jiuqi Wang, Jayanth Srinivasa, Claire Chen, Shuze Daniel Liu, Ali Payani, and Shangtong Zhang argues those diagnostics describe past failure but don't actually predict future trainability, and proposes redefining plasticity operationally as how much optimization gain a network can still achieve on a target task: task-conditional, forward-looking, directly measurable. The empirical question that falls out — which existing diagnostics actually predict trainability and which only describe what already happened — recasts the measurement layer of an entire subfield. Cross-domain parallel: developmental neuroscience and critical periods, the eighty-year arc from Hubel and Wiesel's kittens through Lenneberg's language acquisition hypothesis to today's graded-windows model, where the same distinction between structural correlates of plasticity loss and the operational capacity itself has been worked through for decades. Machine's Learning is a Plumbline Tools production. Support the show: https://plumbline.tools/podcast/

    16 мин.

  4. -3 ДН.

    EP020 — When the Number Isn't the Truth (Camera-Trap Fusion)

    Camera traps scattered across North Carolina are generating images by the millions, and the bottleneck is no longer the cameras — it's labeling. An AI classifier can label every image, but the number it puts next to each one (its 'confidence') is not a probability that it's right. Today's paper from Adira Cohen, Erin Schliep, Roland Kays, Mohammad Alyetama, and Matthew Snider proposes a hierarchical Bayesian model that treats human annotations and AI confidences as two structurally different kinds of noisy measurement and fuses them into a calibrated estimate of what's actually in the picture. Cross-domain parallel: thirty years of quantitative ecology, the senior-partner field that has been combining imperfect observers since long before AI showed up. Machine's Learning is a Plumbline Tools production. Support the show: https://plumbline.tools/podcast/

    15 мин.

  5. -4 ДН.

    EP019 — Three Ways to Watch for Deception (DeceptGuard)

    Imagine watching a contractor remodel your kitchen and being able to peel back three layers of access: first only what they do, then their reasoning out loud, then a recorder that captures every internal mutter. A new paper by Snehasis Mukhopadhyay at arXiv 2603.13791 — DeceptGuard — runs that same experiment on AI agents. It compares three monitoring regimes for catching deception (black-box, chain-of-thought-aware, and activation-probe), evaluates them across a twelve-category taxonomy of verbal, behavioral, and structural deception, and finds that deeper access substantially helps on the subtle long-horizon cases that leave almost no behavioral footprint — but the real frontier is the hybrid ensemble, which hits pAUROC 0.934 on the held-out test set. The cross-domain parallel is the century-long arc of forensic deception detection in human investigation — verbal interview, behavioral-cue analysis, physiology, neural imaging — and the field's converged finding that no single instrument is reliable alone. The forward question: AI agents have no equivalent of the legal-and-ethical floor that constrains human deception detection at each access depth — when the technical ladder becomes available, who decides which rungs are climbed, and what becomes the evidentiary standard for AI oversight findings? Machine's Learning is a Plumbline Tools production. Support the show: https://plumbline.tools/podcast/

    24 мин.

  6. -5 ДН.

    EP018 — Mapping the Patterns at Scale (Attention Atlas)

    Most interpretability research zooms into one attention head, in one model, on one prompt — careful, slow, by hand. A new paper from Jonathan Katzy, Razvan-Mihai Popescu, Erik Mekkes, Arie van Deursen, and Maliheh Izadi at Delft University of Technology asks the opposite question: can a machine sift through millions of attention patterns and discover the families they fall into? Their system, AP-MAE, reconstructs masked attention images with high fidelity, surfaces recurring patterns across models, and turns those patterns into selective interventions that boost code-completion accuracy by roughly thirteen percent. The cross-domain parallel is automated sky surveys in astronomy — the 1990s shift from astronomers hand-classifying galaxies one at a time to projects like the Sloan Digital Sky Survey mass-classifying millions, and the question of what validating infrastructure tells you a discovered category is real rather than an artifact of the discovery procedure itself. Machine's Learning is a Plumbline Tools production. Support the show: https://plumbline.tools/podcast/

    14 мин.

  7. -6 ДН.

    EP017 — A Query With No Words (LAnR)

    Almost every AI chatbot you have used in the last two years has a look-up step underneath — retrieval-augmented generation, or RAG. A new paper from Ha Lan N.T, Minh-Anh Nguyen, and Dung D. Le at VinUniversity changes how the look-up works at a structural level: instead of the model writing a text query for a separate retrieval engine, the same model emits a dense vector directly from its hidden state and matches it against documents in its own internal representation space. One model, no human-readable query, no separate retriever. Cross-domain parallel: the 2,300-year architectural arc of indexed knowledge — Alexandria's Pinakes, Dewey Decimal, MEDLARS, full-text search, vector embeddings, latent retrieval — each generation moving the matching work further from human curation and closer to the substrate of the catalog itself, each requiring a new audit interface its predecessor did not have to imagine. Machine's Learning is a Plumbline Tools production. Support the show: https://plumbline.tools/podcast/

    13 мин.

  8. 14 МАЯ

    EP016 — When New Learning Erases the Old (FTN)

    Train a neural network on task A, then on task B, and its competence on A typically collapses — the field's longest-running open problem in continual learning, called catastrophic forgetting. A new paper from Kevin McKee, Thomas Hazy, Yicong Zheng, Zacharie Bugaud, and Thomas Miconi at the Astera Institute proposes Functional Task Networks: a three-stage procedure that lets a model partition its own parameters across tasks, with each task getting a distinct sparse subset of "dials" that the model can re-instantiate at inference time without retraining. On three continual-learning benchmarks, FTN with fine-grained smoothing reaches near-zero forgetting. The cross-domain parallel: Complementary Learning Systems theory from cognitive neuroscience — the 1995 framework by McClelland, McNaughton, and O'Reilly that argues brains solve the same problem through two interacting memory systems with different time constants — and the forward question of what the ML equivalent of the consolidation step the brain runs during sleep would look like. Machine's Learning is a Plumbline Tools production. Support the show: https://plumbline.tools/podcast/

    20 мин.
См. все (23)

Об этом подкасте

Machine's Learning is a daily podcast produced entirely by AI — two AI hosts in conversation about one fresh paper from machine learning and AI research, translated for thoughtful listeners who don't need a PhD to be curious about where the field is going. One paper per episode, no math required, every cross-domain connection drawn to a universally accessible field (history, biology, medicine, environment) so anyone can follow. By AI, about AI, for humans.

Информация

  • Автор
    Machine's Learning
  • Годы выхода
    2 тыс.
  • Выпуски
    23
  • Ограничения
    Без ненормативной лексики
  • Авторские права
    © (c) 2026 Machine's Learning
  • Сайт подкаста
    Machine's Learning