#025 ML Про стажировки в Яндексе и распознавание рукописных цифр

#025 ML Про стажировки в Яндексе и распознавание рукописных цифр

#025 ML About internships at Yandex and handwritten digit recognition

by Machine Learning Podcast · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning
🎙 Machine Learning Podcast 👥 5K 📅 May 19, 2021 ⏱ 50 min 👁 808 📄 tutorial 🧭 2026-10-10
Available in: English (current) Français

Keywords

PythonKerasMNISTgradient descentinternship

Summary

This episode of the Machine Learning Podcast is divided into two parts. In the first part, Alina, a team leader for intern recruitment at Yandex, discusses how paid internships can provide an entry point into machine learning. She explains the application process, including the online form, resume tips, and a test assignment with six tasks covering algorithms, math, and ML. She emphasizes that there are no age or education restrictions and that remote internships are currently being phased out. The second part, hosted by Mikhail, is a tutorial on building a fully-connected feedforward neural network to recognize handwritten digits from the MNIST dataset. He bridges theory and practice by starting with a simple linear regression example of converting Celsius to Fahrenheit, then explaining the concept of a pipeline, data preprocessing, and the training process. He touches on overfitting, model architecture, and optimization. The episode aims to help listeners who have theoretical knowledge but struggle to apply it to real code. Practical resources such as the Yandex internship page and a Google codelab for TensorFlow MNIST are referenced throughout.

178 words

Critical Evaluation

Value of the Information & Strength of the Argument

The information value is moderate to high: the internship segment offers practical, often hard-to-find advice on applying to Yandex, including details about the test assignment and selection criteria, while the MNIST tutorial provides a clear step-by-step demonstration of building and training a simple neural network. The argumentation is solid, using analogies like the Celsius-to-Fahrenheit conversion to illustrate model training and the pipeline concept. However, the discussion is somewhat superficial on technical details, and the author’s personal anecdotes, while relatable, are not substitutes for systematic explanations. The structure is logical, progressing from internship advice to a concrete coding example, which supports the claim of bridging theory and practice.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is acceptable for an educational podcast. The presenter references the MNIST dataset and linear regression concepts correctly, and the Google codelab link is a recognized resource. However, the podcast relies heavily on personal experience and anecdotal evidence, especially in the internship section, without citing external studies or official documents beyond the Yandex page. The sources listed in the description are mostly podcast-related and self-promotional. The title accurately matches the content: it explicitly mentions internships and handwritten digit recognition. The video is well-structured, but the lack of strict factual verification or peer-reviewed references limits its scholarly reliability.

221 words

Title / Content Match

The title accurately reflects the two-part content: first about Yandex internships, then about recognizing handwritten digits with a simple neural network.

Quality & Reliability

7/10

The podcast provides accurate information about ML internships and a practical tutorial on MNIST, based on personal experience and well-known educational resources. However, it lacks rigorous citations and some claims are anecdotal.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

External References

Contribution & Novelties

The podcast’s original contribution lies in combining practical career advice for entering the ML industry via large-company internships with a hands-on tutorial that grounds theoretical concepts in code. The author bridges the gap between knowing neural-network concepts and actually writing a working model, using the MNIST example as a common educational benchmark. This dual focus is relatively unique in educational content.

Pour aller plus loin :

  • MNIST database — The standard dataset for handwritten digit recognition, essential for understanding the tutorial.
  • Gradient descent — The core optimization algorithm used to train neural networks; the podcast briefly mentions it.
  • Overfitting — A key concept discussed in the episode, explaining why test sets are necessary.

113 words

Radar Profile

The radar profile shows a balanced but modest performance across all dimensions. Information quantity is relatively high due to the 50-minute duration, but technical depth and reliability are moderate, reflecting the podcast's accessible nature. The overall score is consistent with a solid educational resource that could benefit from more rigorous sourcing and deeper technical explanations.

Reliability 7/10