No-Code AI Tools: Lobe, Teachable Machine

No-code AI tools allow users to build, train, and deploy machine learning models without writing a single line of code. They provide a graphical user interface (GUI) where users can upload data, label it, and initiate training. The underlying algorithms—typically deep neural networks—are abstracted away. For AI-900, you need to know that these tools are designed for rapid prototyping, learning, and small-scale projects, not for production-grade enterprise solutions. They support image classification, object detection, sound classification, and pose detection.

Lobe is a free, offline desktop application from Microsoft that enables users to train image classification and object detection models. It uses transfer learning with a pre-trained MobileNetV2 architecture optimized for edge devices. Lobe automatically splits your data into training (80%) and validation (20%) sets. It supports up to 10 classes and a few thousand images per class. Training runs locally on your machine using CPU or GPU (if available). The tool provides real-time training metrics like accuracy and loss, and it exports models to TensorFlow Lite, ONNX, or as a local API.

How Lobe Works Internally: - Data Ingestion: Users drag and drop images into labeled folders. Lobe resizes images to 224x224 pixels and normalizes pixel values. - Feature Extraction: The pre-trained MobileNetV2 extracts feature vectors from each image. - Training: Lobe adds a new classification head on top of the pre-trained base and fine-tunes the entire network using stochastic gradient descent. The default learning rate is 0.001, batch size 32, and training runs for 50 epochs (early stopping if validation accuracy plateaus). - Evaluation: After training, Lobe shows a confusion matrix and per-class precision/recall. - Export: Users can export the model as a TensorFlow Lite file (for mobile/edge), ONNX (for cross-platform), or deploy to Azure via a generated Docker container.

Key Limitations for the Exam: - Lobe only supports image data (no audio, video, or text). - Maximum 10 classes. - No automatic hyperparameter tuning; uses fixed defaults. - Models are not production-grade; they are meant for prototyping. - Lobe does not support multi-label classification (each image belongs to exactly one class).

Teachable Machine is a web-based tool from Google that supports image, audio, and pose classification. It uses TensorFlow.js in the browser, meaning all training happens client-side. It supports up to 10 classes for images and audio, and up to 5 classes for poses. For images, it uses a MobileNet pre-trained on ImageNet; for audio, it extracts MFCC features; for poses, it uses PoseNet to extract keypoints.

How Teachable Machine Works: - Image Classification: Same as Lobe but entirely in-browser. Users can use their webcam to capture images live or upload files. The model is trained using transfer learning with MobileNet. - Audio Classification: Users record short audio clips (0.5 to 2 seconds). The tool extracts Mel-frequency cepstral coefficients (MFCCs) and trains a simple neural network. - Pose Classification: Uses PoseNet to detect 17 keypoints (e.g., nose, shoulders, elbows). The relative positions of these keypoints form the input features. Training uses a small neural network. - Export: Models can be exported as TensorFlow.js (for web apps), TensorFlow Lite, or as a link to share.

Key Differences from Lobe: - Teachable Machine is web-based; Lobe is desktop. - Teachable Machine supports audio and pose; Lobe does not. - Teachable Machine's training is less sophisticated (fewer epochs, simpler architecture) and is meant for quick demos. - Lobe provides more detailed metrics (confusion matrix, precision/recall). - Teachable Machine can be used with a live webcam for real-time classification.

Both tools rely on transfer learning: they take a pre-trained model (MobileNet) and retrain only the last few layers on the user's data. This drastically reduces the amount of data needed (hundreds vs. millions of images) and training time (minutes vs. days). The pre-trained model already knows how to detect edges, textures, and shapes; the new layers learn to map those features to the user's specific classes.

AI-900 tests the ability to differentiate between no-code tools and code-based solutions like Azure Machine Learning designer or custom Python scripts. You should know:

A common exam trap is confusing Lobe with Azure Machine Learning designer. The key differentiator: Lobe is specifically for image models with a very simple interface; Azure ML designer is for broader ML tasks (including tabular data) and offers more control but also more complexity. Another trap: assuming Teachable Machine can handle large datasets—it is limited to small samples (a few hundred per class) because it runs in the browser.

A manufacturing company wants to inspect products for defects using a camera. They have no data scientists and need a quick proof-of-concept. They use Lobe to train an image classifier on 500 images of good products and 200 images of defective ones. Lobe's model achieves 95% accuracy. They export the model as TensorFlow Lite and deploy it on a Raspberry Pi with a camera. The system flags defective products in real-time. However, Lobe's limitation of 10 classes means they cannot expand to multiple defect types without retraining. For production, they later migrate to Azure Custom Vision, which supports 100 classes and automated retraining.

A museum wants an exhibit where visitors can control animations using hand gestures. They use Teachable Machine's Pose Classification to recognize 5 poses (e.g., wave, thumbs up, peace sign). They train the model with 50 samples per pose using a webcam. The model is exported as TensorFlow.js and embedded in a web page. Visitors stand in front of a camera, and their gestures trigger different animations. The system works well in controlled lighting but fails in low light because PoseNet relies on clear video. Teachable Machine's limitation of 5 pose classes is sufficient for this demo. For a more robust solution, they would need a custom TensorFlow model.

A teacher wants to create an app that helps students learn animal sounds. They use Teachable Machine's Audio Classification to train on 10 animal sounds (e.g., dog bark, cat meow). They record 20 short clips per animal using a microphone. The model is trained in seconds and exported as a shareable link. Students can play sounds and the app identifies the animal. The accuracy is around 80% because background noise and recording quality vary. For a more accurate educational tool, the teacher would need a larger dataset and noise reduction preprocessing, which Teachable Machine does not offer.