Collect and Prepare Data for Teaching a Machine

Think about the last time you tried a new recipe. Remember flipping through that cookbook or scrolling through that food blog, and selecting that mouth-watering dish you wanted to try? Now, the recipe calls for specific ingredients, doesn’t it? Fresh tomatoes, aged parmesan, hand-torn basil… Each ingredient, carefully chosen, plays a pivotal role in turning that dish into a culinary masterpiece.

But what if you used ketchup instead of fresh tomatoes? Or swapped out that parmesan for any random cheese you found in the fridge? The end result might not be the gourmet experience you anticipated.

In many ways, venturing into the world of machine learning, especially with tools like Google’s Teachable Machine, is quite similar to cooking. The ‘recipe’ for training a machine requires specific, high-quality data. And just like in cooking, the outcome largely depends on the ingredients you use. Let’s explore the art and science of collecting and preparing the finest ‘ingredients’ or data, to ensure our machine learning ‘dish’ is not just palatable but exemplary!


Understanding Data Needs for Teachable Machine Projects

Imagine you’re setting up a vast library, and you want to sort books into different genres. But here’s the catch – you have 20 copies of the same book. How useful is that library going to be? Just like this library, a machine learning model is only as good as the variety and relevance of the ‘books’ or data you provide it.

Diversity of data
For a machine learning model to be robust and effective, it needs diverse data. It’s akin to understanding a topic by reading multiple books, articles, and viewpoints.

  • Variety of Angles: When teaching a machine to recognize objects, imagine it as someone who’s never seen that object before. By providing images from multiple perspectives – top, front, side – you’re giving it a holistic view.
  • Variety of Tones: For sound-based tasks, nuances matter. Different brands of guitars or even the same guitar played differently can produce varied tones. By capturing these variations, the machine can better understand the essence of a ‘guitar sound.’
  • Variety of Movements: Humans never replicate movements exactly. A seasoned dancer might perform a move slightly differently than a beginner. By feeding these differences into the machine, you make it more versatile and understanding of the core movement.
  • Different Environments: A plant might look different under the morning sun compared to under a cloudy sky. By training the machine with these nuances, you’re preparing it for real-world unpredictability.
  • Variations: Whether it’s the calm of an elderly person’s smile or the exuberance of a child’s grin, the subtle differences in facial expressions across ages and ethnicities can greatly enrich your dataset. Similarly, the richness of voice variations or exercise forms makes the model more comprehensive and adaptable.

Data relevance and quality
While diversity is crucial, relevance is equally paramount. Imagine trying to learn French, but every fifth word in your lesson is in German. Confusing, right? If you’re teaching the machine about cats and dogs, throwing in a parrot might just confuse it. It’s like adding a sci-fi novel to the romance section of our library. If the focus is on a juicy apple, a cluttered kitchen backdrop might pull focus. Similarly, in audio tasks, you’d want the primary sound to stand out, not be muddled by background music.


Collecting and Preparing Data for Teachable Machine

Imagine you’re preparing a gourmet meal. The ingredients you select, their quality, and how you prepare them can make all the difference between a memorable dish and a forgettable one. Similarly, the quality, diversity, and preparation of your data dictate the success of your machine learning model.

  1. Determining the type of data needed
    Before diving into the details, it’s crucial to pinpoint the type of data you’ll be working with. Whether it’s images of flowers, audio snippets of musical instruments, or poses of various yoga postures, the nature of your project will guide this choice.
    Imagine a jar of unlabeled spices. How do you know which is paprika and which is cayenne pepper? Labels serve a similar purpose, providing clarity and direction to our machine model.
  2. Collection of necessary data
    Teachable Machine is like your digital fishing net, allowing you to gather data through its intuitive interface, be it through snaps or voice recordings. A rich recipe is made of varied ingredients. Similarly, the more diverse your data, the more nuanced and adaptive your machine model will be.
  3. Importing data into Google’s Teachable Machine
    Whether it’s a cherished family recipe (your own dataset) or fresh produce (live captures), Teachable Machine offers flexibility in ingredient sourcing. Ensure that your ingredients (data files) are fresh and suitable (supported formats). And just as ingredients are separated in a recipe, segregate your data into distinct classes during upload.
  4. Preprocessing of data
    An overripe tomato can ruin a salad. Similarly, ensure that your data—be it a blurry image or a noisy audio clip—is of the best quality. While Teachable Machine eases the cooking process, sometimes a little manual intervention, such as enhancing image contrast or filtering out noise, can elevate your dish (model).
  5. Data labelling
    As with labeled spices in a rack, clear and descriptive names help in navigating and understanding your data landscape. During the upload, name each category descriptively, ensuring the model has clear indications of each class.
  6. Data splitting
    Imagine taste-testing a sauce before serving it. The training data is where your model gets its primary flavor, while the validation set ensures it’s seasoned just right. If you only taste the sauce while adding salt, you might over-season it. Similarly, using only the training data can lead to overfitting. Validation keeps this in check.
  7. Saving your project
    Just as you’d store the leftover sauce for another day, saving your project ensures that you can revisit, refine, or reuse your model as needed. Saving allows for iterative refinement. Like tweaking a recipe based on feedback, models can be refined based on results and changing requirements.


The Art of Data Preparation for Teachable Machines

Imagine you’re about to paint a masterpiece. Your success isn’t determined solely by your painting skills but also by the quality of your brushes, paints, and canvas. Similarly, in the realm of machine learning, the outcome isn’t determined only by the algorithm or tool used, but crucially by the quality and preparation of your data. Here are some best practices:

  • Balance the data
    Think of this as the symmetrical balance in a painting. If your canvas is overwhelmingly filled with one subject (say, cats) over the other (dogs), the viewer’s attention will naturally drift toward the dominant subject. Likewise, an imbalanced dataset can cause the model to lean heavily towards one category.  A balanced dataset ensures that the learning process is fair, providing the algorithm with an impartial view, thereby enhancing accuracy and reducing overfitting risks.
  • Diversity of data
    Just as an artist uses a wide palette to capture the vibrancy of life, ensure your dataset encompasses varied instances. From the pale hues of dawn to the deep tones of dusk, the more diverse your data, the richer the model’s understanding. This ensures that your model isn’t just theoretically sound but performs effectively in real, unpredictable situations, enhancing its reliability and adaptability.
  • Clean the data
    Before an artist starts, they ensure the canvas is clean and primed. Analogously, irrelevant ‘noise’ in your data can obscure the model’s understanding. This might mean omitting punctuation in text data or ensuring clear backgrounds in image data. A clean dataset translates to faster training times, increased model accuracy, and a more efficient machine learning process.
  • Understand and handle missing data
    In art, blank spaces can either add meaning or indicate incompleteness. In data science, missing or null values must be consciously addressed, whether that’s by imputing them or making strategic deletions. Handling these gaps ensures that the final model is both robust and reflective of the true nature of the data, reinforcing the reliability of your outcomes.
  • Splitting the data
    Partition your dataset into training, validation, and testing sets. Each serves its distinct purpose, guiding the model’s journey from learning to validation and eventual real-world testing. This trifecta not only bolsters the model’s ability to learn but also provides checkpoints to refine and evaluate its performance, ensuring its robustness against overfitting or underfitting.


Navigating the Pitfalls of Data Preparation in Teachable Machines

The art of teaching machines isn’t merely about feeding data and expecting accurate outcomes. It’s akin to cultivating a garden; if you don’t tend to the soil, prune the plants, or water them appropriately, you cannot expect a lush harvest.

  • The trap of biased and insufficient data
    Imagine launching facial recognition software and witnessing it falter, not recognizing a significant proportion of the world’s population. This was the fate of a company that relied heavily on a dataset dominated by Caucasian male faces. Their model’s understanding became a mirror of this limited perspective.
    The narrow horizon of their dataset didn’t capture the richness and diversity of the human populace. The bedrock of machine learning is data, and if that foundation is skewed, the superstructure, no matter how advanced, will wobble.
    A model’s inability to perform outside its training constraints can diminish user trust, affect business decisions, and may even hold ethical implications, especially in sensitive applications.
    Fix: Treat data collection as a mission to represent the real world in all its diversity. A deliberate and strategic approach, ensuring that all facets of the problem space are covered, is vital.
  • Overlooking the cleanup
    An ambitious e-commerce platform sought to predict its future sales trajectory. They enthusiastically fed their model all sales data, without discerning between genuine sales, canceled orders, and returns. The result? A model that misread past sales and mispredicted future trends.
    Amidst the excitement of leveraging machine learning, the crucial step of refining the raw data was overlooked. Feeding a model without vetting the data is akin to building a house on unexamined land; surprises, often unpleasant, are bound to spring up.
    Without a clean dataset, the model learned from noise and inconsistencies, leading to predictions that, instead of guiding the business, could mislead it into flawed strategies.
    Fix: Dedicate time and resources to data preprocessing. This isn’t just a preliminary step; it’s foundational. Address anomalies, rectify missing values, and ensure data is consistent and aligned with the problem statement.

Related Tags: