I6 · Neural Networks and Deep Learning Essentials

Training a Network Without Code

No-code deep-learning tools let you build, train and test neural networks through a visual interface. Google's Teachable Machine trains image, sound and pose classifiers from examples you record, using transfer learning on a pre-trained model. TensorFlow Playground shows a small network learning to separate coloured dots, letting you change layers, neurons, activation, learning rate and regularisation and see the effect immediately.

12 min Intermediate · Sublevel Intermediate Jump to the activity

What you need: The earlier topics of this module.

Definition

You can train a real neural network in your browser in five minutes, using your webcam and no programming at all, and then watch it make the same mistakes that professional models make.

No-code deep-learning tools let you build, train and test neural networks through a visual interface. Google's Teachable Machine trains image, sound and pose classifiers from examples you record, using transfer learning on a pre-trained model. TensorFlow Playground shows a small network learning to separate coloured dots, letting you change layers, neurons, activation, learning rate and regularisation and see the effect immediately.

In simple words

It is like a cooking show where the ingredients are already chopped. You still choose the recipe, taste the result and fix mistakes, but you skip the knife work, so you can focus on what makes a dish good.

Overview

  • No-code tools let you train neural networks by uploading or recording examples.
  • Most use transfer learning: a pre-trained network is fine-tuned on your examples.
  • You still decide the classes, collect balanced examples and test on new data.
  • Tools show accuracy and confusion so you can spot weak classes.
  • Models can be exported for websites, apps or code.
  • Good results still depend on varied, representative examples.

Why it matters

No-code tools build intuition fast: you see directly how data quality, class balance, network size and learning rate change the result. They are also used by teachers, artists and small businesses for real prototypes.

Key terms

Teachable Machine
A free browser tool that trains image, sound and pose classifiers from your own examples.
TensorFlow Playground
An interactive visual of a small network learning to classify points.
Class
A category the model learns to recognise, with its own set of examples.
Decision boundary
The line or curve where the model switches from one class to another.
Export
Saving the trained model to use in a web page, an app or a microcontroller.

Types

No-code neural-network tools come in several types, each suited to different data and goals.

1. Teach-by-example web tools

Browser tools where you show examples from a webcam, microphone or files.

  • How it works: They fine-tune a small pre-trained network in the browser within seconds and let you test it live.
  • Example: Teachable Machine learning to recognise thumbs-up and thumbs-down gestures.
  • When to use it: Use them for quick prototypes, demonstrations and learning how classifiers behave.

2. Interactive playgrounds

Visual simulations of small networks training on toy data.

  • How it works: You change layers, neurons, activations and learning rates and watch the decision boundary and loss change in real time.
  • Example: TensorFlow Playground showing that a spiral data set needs more hidden layers.
  • When to use it: Use them to build intuition about how architecture and settings affect learning.

3. Visual workflow tools

Desktop tools such as Orange and KNIME with neural-network widgets.

  • How it works: You connect data, preprocessing, a neural-network learner and evaluation widgets, and compare it with other models.
  • Example: Comparing a neural network with a random forest on a student-performance data set.
  • When to use it: Use them to compare models on tabular data without writing code.

4. Cloud AutoML and no-code platforms

Services that train image, text or tabular models from uploaded data.

  • How it works: They search architectures and settings automatically and return a deployable model with evaluation reports.
  • Example: Training a product-defect detector from labelled factory photos in a cloud console.
  • When to use it: Use them for business prototypes; watch costs and keep control of private data.

Explained in detail

Teachable Machine, step by step

Open Teachable Machine and choose an image project. Create classes, for example "thumbs up", "thumbs down" and "nothing". Hold each gesture in front of the webcam and record a few hundred frames per class. Click Train: behind the scenes the tool takes a pre-trained MobileNet, extracts features from your images and trains a small new head, which is why it finishes in seconds. The preview shows live confidence bars. Open the advanced settings to see epochs, batch size and learning rate, the same settings covered in this module. Finally you can export the model for a web page, for an Android app or, for some project types, for a small microcontroller.

Experiments that teach real lessons

ExperimentWhat you will seeLesson
Record one class in a bright room and another in a dark roomThe model learns lighting, not gesturesShortcut learning and spurious correlations
Give the first category only 50 images and the second 500The model favours the larger classClass imbalance
Train only on your own face, then test a friendConfidence drops or predictions flipGeneralisation needs diverse data
Leave out a "nothing" classSomething is always predicted, even for an empty frameModels must choose among known classes

TensorFlow Playground

In TensorFlow Playground, choose the spiral data set. With no hidden layers the network can only draw a straight boundary, and it fails. Add two hidden layers of six neurons with ReLU, and press play: the boundary bends around the spiral over a few hundred epochs. Raise the learning rate to 3 and the test loss jumps about; lower it to 0.00001 and nothing happens. Add noise and train a big network for a long time, and the boundary wraps around individual noisy points, which is overfitting; add L2 regularisation and it smooths out. Each knob in the interface corresponds to a line of code you will write in the next topic.

  • Never upload photos of other people without their permission, and avoid recording private spaces.
  • Teachable Machine trains in your browser by default, so recorded examples stay on your device unless you choose to save them to Google Drive.

How it works

  1. Decide on two to four clearly different classes.
  2. Record varied examples for each: different angles, backgrounds, lighting and people.
  3. Train, then test with new examples the model has not seen.
  4. Find failures and fix them by adding the missing kinds of examples.
  5. Export the model, or note what you would change before real use.

Try it yourself

Train a classifier

Add examples to each class, train the model, then test it on new examples. Try making one class much smaller and see what happens.

The gesture model that learned the wall

Two friends build a rock-paper-scissors gesture model in Teachable Machine.

Common mistakes

  • Testing only on the same examples used for training.
  • Recording each class in a different setting.
  • Uploading other people's images without consent.

Check yourself

Question 1

1. Why does Teachable Machine train in seconds?

Question 2

2. A model works in one room but fails in another. The most likely cause is…

Question 3

3. In TensorFlow Playground, a network with no hidden layers on spiral data…

Summary

  • Teachable Machine and TensorFlow Playground let you train real networks without code.
  • Experiments reveal shortcut learning, class imbalance and the need for diverse data.
  • Every knob in these tools matches a setting you will later control in code.

Read it. Now try it yourself.

Open the interactive activity

Finished reading?

Sources: Google Teachable Machine · TensorFlow Playground