简体中文
English
Website review
This review may include AI-generated content and is not individually verified. Domain, certificate and other facts come from the observed data on this page.
What is Jingles writes about Machine Learning & Neuroscience?
It is the personal site of Hong Jing (who publishes as "Jingles"), a data scientist and Alibaba PhD student who writes about machine learning and neuroscience. The site collects blog articles, project showcases, and published work, so it works less like a course and more like a working notebook: explanations of concepts, experiments, and reflections on how AI and brain science inform each other. You can find it at Jingles writes about Machine Learning & Neuroscience.
What the articles cover
Based on the post titles on the page, the material spans several practical areas:
- Deep learning fundamentals — how convolutional layers work, an illustrated guide to Transformers, and why learned optimizers can outperform Adam.
- Applied and emerging topics — coordinating multiple reinforcement-learning agents, 3D object understanding with Objectron, GANs beyond art, and realistic deepfakes on Colab.
- Neuroscience crossover — decoding brain signals with machine learning, the relationship between AI and neuroscience, and evaluating "intelligent" AI.
- Getting started and context — a ranked path into ML from easiest to hardest, sequence-prediction problem types, and language models as of 2020.
Who it suits
The mix points to readers who want conceptual walkthroughs with a research flavor rather than API tutorials: students deciding where to start, engineers curious about neuroscience-inspired ideas, or anyone who wants a readable explanation of a model before reading the paper. The dated pieces (such as the 2020 language-model overview) are best treated as historical context, not current guidance.
A practical next step: if you are new to the field, start with the ranked "7 ways to get started into ML" post, then follow the convolutional-layers and Transformer guides; if you already work in ML, go straight to the reinforcement-learning and brain-signal decoding articles, which are closer to active research.
How can I get started in machine learning based on the ranked difficulty guide?
The ranked guide is a practical ladder: start with the easiest entry point that gets you running code, then add theory as you hit limits. Hong Jing's site lists "7 ways to get started into ML, ranked from easiest to hardest," alongside beginner-oriented explainers such as "How Convolutional Layers Work in Deep Learning Neural Networks?" and "3 Types of Sequence Prediction Problems" on Jingles writes about Machine Learning & Neuroscience. The ranking itself is the author's framing; treat the order as a suggested sequence rather than a rule.
A workable order
- Run something end to end. Pick a small tabular dataset and train a simple model. The goal is familiarity with the loop: load data, split it, fit, evaluate, iterate.
- Learn the core concepts behind what you just ran. Loss functions, overfitting, train/test splits, and why a baseline matters. Jingles' convolutional-layer explainer is useful once you reach images, not before.
- Move to a domain that interests you. Sequence problems, computer vision, or reinforcement learning each have their own vocabulary; the site's posts on sequence prediction and coordinating RL agents show how quickly the topics branch.
- Read the "why," not just the "how." Posts like the one on learned optimizers versus Adam are more valuable after you have trained a few models and felt the tuning pain they address.
Choosing your entry point
| Your situation | Sensible start | Trade-off |
|---|---|---|
| No coding background yet | A guided course or tutorial with runnable notebooks | Slower to reach real projects, but fewer dead ends |
| Comfortable with Python | Implement a model from scratch on a small dataset | More effort, but concepts stick |
| Strong maths, weak tooling | Theory first, then a framework tutorial | Risks knowing the maths without debugging skills |
| Domain expert (biology, finance, medicine) | Apply ML to your own data early | Narrower foundation, faster relevance |
A concrete first month
Spend week one on a single tutorial project you fully understand. Weeks two and three, rebuild it from a blank file without copying. Week four, change one thing — a different dataset, a different model — and write down what broke. That written record is more useful than a certificate when you decide what to study next.
For broader structured paths, compare with Google Developers machine learning resources or scikit-learn's user guide, which pairs each algorithm with practical examples.
How do convolutional layers work in deep learning neural networks?
Convolutional layers work by sliding a small set of learnable filters (kernels) across the input and computing a dot product at each position, producing a feature map that records where a given pattern appears. Because the same filter is reused at every location, the layer detects a pattern regardless of where it sits in the image, and it needs far fewer parameters than a fully connected layer over the same pixels. Stacking layers builds a hierarchy: early filters respond to edges and colour transitions, deeper ones to textures and object parts.
What happens inside one layer
- Filter and receptive field: each kernel covers a small window of the input, so it only sees local structure at a time.
- Stride and padding: stride controls how far the window jumps between positions; padding controls whether the output shrinks or keeps its spatial size.
- Channels: a filter spans all input channels at once, so one kernel produces one output channel; several kernels produce several.
- Nonlinearity and pooling: an activation (typically ReLU) follows the convolution, and pooling or a strided convolution reduces spatial resolution before deeper layers.
Where the trade-offs show up
A small kernel with modest depth keeps computation cheap and generalises well on limited data, which suits a reader training a first image classifier on a few thousand labelled photos. Large kernels or very deep stacks capture broader context but demand more data, memory and tuning. If your model memorises the training set quickly, adding pooling or reducing filter count is usually a better first move than adding layers.
For a hands-on walkthrough of the mechanics, Hong Jing's article on how convolutional layers work in deep learning neural networks at Jingles writes about Machine Learning & Neuroscience is a practical starting point; pair it with the visual explanations at Setosa if you want to see the sliding-window arithmetic step by step.
How is machine learning used to decode brain signals in neuroscience research?
Machine learning helps neuroscience researchers decode brain signals by learning patterns that map recorded neural activity to behaviour, stimuli or mental states, rather than relying only on hand-crafted analysis rules. On Jingles writes about Machine Learning & Neuroscience, Hong Jing writes from a data-science and neuroscience perspective, including an article on decoding brain signals with machine learning and neuroscience, so the site is a useful starting point for seeing how the two fields are combined in practice.
What decoding typically involves
- Recording: brain activity is captured with methods such as EEG, fMRI or implanted electrodes, producing noisy, high-dimensional time series.
- Feature learning: models learn which patterns in that activity carry information, instead of researchers specifying every feature by hand.
- Prediction and validation: the model predicts a label, such as a movement intention or viewed image, and is tested on held-out data to check it generalises.
- Interpretation: researchers inspect which signals or time windows drive predictions, connecting model behaviour back to brain function.
Practical uses and audiences
A concrete scenario: a lab records EEG while participants imagine moving a hand, then trains a classifier to distinguish imagined left from right movements. This supports brain–computer interface work and gives a measurable test of how much movement information the signal contains. Neuroscientists use decoding to test theories about what information a brain region carries; engineers use it to build interfaces; clinicians are interested in biomarkers and assistive technology.
The trade-off is between flexible models, which can capture subtle patterns but need more data and are harder to interpret, and simpler models, which are easier to trust but may miss complex structure. Careful cross-validation and awareness of noise and artefact contamination matter more than model sophistication.
Next step: read the site's decoding article alongside a general primer such as scikit-learn documentation to see which classifiers fit small neural datasets.
How can I set up LaTeX in the Atom editor for writing technical documents?
The short answer: Atom itself is discontinued, so treat any guide for it as a way to understand the setup pattern rather than a long-term workflow. If you still use Atom, the usual approach is to install a LaTeX package (such as latex or latex-plus) from Atom's package manager, configure it to call your installed TeX distribution, and use a per-project build setting so it compiles the right file. If you are choosing a tool now, a maintained editor with a LaTeX extension or an online LaTeX editor will save you maintenance work.
What the setup actually involves
- A TeX distribution first. Install TeX Live (Linux/macOS) or MiKTeX (Windows) before touching the editor. The editor only launches the compiler; it does not provide LaTeX.
- A build command. Most Atom LaTeX packages compile with
pdflatex,xelatexorlatexmk.latexmkis usually the least fiddly because it reruns passes automatically for references and bibliographies. - A viewer. Either let the package open the PDF after each build, or use a PDF viewer that reloads on file change.
- Project-level settings. Store the main file and build recipe in the project rather than in global settings, so multi-file documents compile consistently.
A concrete reader scenario
Suppose you are writing a paper with a main.tex file, a refs.bib bibliography and several included section files. You would point the package at main.tex, choose latexmk as the builder, and keep the PDF open beside the editor. When you add a citation, the automatic rerun resolves it without manual passes. If you later switch editors, the .tex files and build recipe carry over unchanged, which is the main reason to keep document structure independent of the editor.
Choosing between options
| Option | Good for | Trade-off |
|---|---|---|
| Atom plus a LaTeX package | Reusing an existing Atom setup, quick local editing | Atom is no longer maintained, so packages and integrations can break |
| A maintained code editor with a LaTeX extension | Long-term technical writing with Git and multi-file projects | Requires configuring a build recipe and a PDF viewer |
| An online LaTeX editor | Collaboration, zero local install | Less control over packages and offline work |
For a practical next step, decide based on how long the document will live: for a one-off assignment, an online editor is fastest; for a thesis or a document you will revise for months, set up TeX Live with latexmk in a maintained editor. Jingles writes about Machine Learning & Neuroscience includes a post on setting up LaTeX in the Atom editor, useful as a reference for the package-and-build-command pattern even if you apply it elsewhere.
What is the relationship between AI and neuroscience according to the site?
The site treats AI and neuroscience as a two-way exchange rather than a one-way borrowing. Jingles (Hong Jing), a data scientist and PhD student writing at Jingles writes about Machine Learning & Neuroscience, frames the connection as "The Fascinating Relationship between AI and Neuroscience," and pairs that overview with a hands-on piece on "Decoding Brain Signals with Machine Learning and Neuroscience." The recurring theme across the blog is that neural networks were inspired by biological brains, and that modern ML tools now return the favour by helping researchers make sense of brain data.
Two concrete directions follow from that framing:
- Brain as inspiration for AI. Posts like "How Convolutional Layers Work in Deep Learning Neural Networks?" and "An Introduction to Signals" start from concepts that have biological analogues, which is useful if you are learning ML and want intuition before the maths.
- AI as a tool for neuroscience. The brain-signal decoding article sits alongside applied work such as computer vision for medical diagnosis, suggesting the author's interest in ML as an instrument for interpreting messy biological measurements.
For a reader deciding whether this angle matters to them: if you want model architecture explained through biological analogy, start with the CNN and signals pieces. If you care about applying ML to neural recordings, go to the brain-signal decoding post first and treat the relationship essay as the conceptual frame. The trade-off is depth versus breadth—these are blog-level introductions, so expect orientation and worked examples rather than a survey of the research literature.
Amazon Route 53
Fastly
User reviews (0)