Brain Builder » Recognition » Brain Training Beginner's Guide » Walk-Through 2: Getting Smarter

In This Section:

In This Walk-Through:

  • Training a more complex Brain
  • Data quality
  • Brain Scoring
  • Time: 10-15 minutes

Robot Room Recognition

Imagine you are developing a household floor-cleaning robot. To distinguish your product from others in the market, you want your robot to have a Brain that it can use to understand what room it is in so that it can adapt its cleaning protocol accordingly (for example, sweeping the kitchen twice).

You have amassed a dataset with a number of images of common household rooms and you're ready to train your Brain.


The quality of your training data is the single most important factor in determining how successful your Brain will be. Though we have provided the data for you in this Walk-Through, be sure to read the tips in Uploading Data for Recognition Training to ensure that your data is as effective as possible.

Train your Room Recognition Brain

  1. Navigate to the Dataset in your project called 2. Room Recognition, and click Resume Tagging. This Dataset is preloaded with 64 images of three types of rooms: Bedroom, Living Room, Kitchen.

  2. Start tagging images to train your Brain. As you tag, you'll notice a few important differences between these images and the Sheet Metal images in the first walk-through. There is much greater variety in the room recognition images.

    Information Simplicity & Complexity

    The Sheet Metal Brain started making accurate predictions after learning the first four or five images. This was possible because of the simplicity and consistency of the data. Those images were all taken by the same camera from the same perspective showing the same part of a piece of sheet metal. Any differences between the images were relevant to determine which type of metal defect was present.

    In the Room Recognition Dataset, there is a lot more variety. Each living room, kitchen, and bedroom is different—not only from the other room types but even from other images of the same room type. The Brain must learn the range of variety that exists for each class so that it can generalize the knowledge to accurately recognize bedrooms that it hasn't previously seen.

    Training a Brain with a more complex dataset requires more images than training a Brain with a simple dataset.

  3. Check your Brain Score located at the top of the Workspace to know when it's smart enough.

    For the first 10-20 images, your Brain Score will be Still Learning. You've probably tagged images of mostly one room type, so your Brain simply hasn't learned enough to have a score yet. The Brain will probably make predictions on some of the images, but it won't be able to make many quite yet.

    After about 20 images, you'll start adding images for another room type. At this point your Brain score may change to Low. Other possible scores are Good or Great.

    Click the arrow next to Brain Score to reveal the full Brain Score details.

    Warning The scores in this walk-through might not exactly match your experience. If you skip images or tag them in a different order, you may get a different Brain score or accuracy value. For more details on the Brain Score methodology, see Understanding Brain Performance.

    Here you can see the Brain's accuracy, the accuracy of each class, and recommendations for how to make your Brain smarter.

    The Brain measures accuracy by testing its predictions on some of the images you've already tagged. Click the Images bubble to see exactly which of those images the Brain gets right and wrong.

  4. Keep tagging.

    The Brain Score suggests tagging more Living Room and Kitchen images, so let's keep going.

    After you have tagged 42 images, you will have trained the Brain on all of the Bedroom and Kitchen images. Before starting the Living Room images, let's check on the Score.

    Well, that's not good. The Brain accuracy has gone down!

    Brain Builder uses patented Lifelong-AITM technology to reduce the impact of Catastrophic Forgetting—a term that refers to AI systems forgetting previous knowledge when they learn new information. However, when similar types of images are learned in batches like this, behavior similar to catastrophic forgetting can still occur. Fortunately, there's a way to fix it.

    But first, let's tag the Bedroom images and finish training the Brain. After tagging all the images, the accuracy is back up to 80 percent but the Brain Score is still Low.

    The Brain has learned the most recent Living Room images quite well, but it's not doing as well on the images it learned previously.

To Prevent Unbalanced Training

  • Randomize your data so it's not uploaded in order by class. This helps the Brain learn in a more balanced way, which should improve performance after all of the images have been learned.

  • Use Brain Builder's Optimize Brain feature to achieve the highest possible Brain accuracy score given the current data. This feature shuffles the images in different orders on the fly inside the platform to optimize Brain performance. See Retraining your Brain for more information.

Retraining the Brain

Navigate back to the Dataset page, and click the Deploy Brain button.

This takes you to the Publishing Workflow (which we'll discuss more in Walk-Through 3).

At the top-right of the Publishing page, click the Brain Actions drop-down menu.

Click the Optimize Brain button.

Click Continue.

After the optimization process completes, the screen updates with the new Brain accuracy score.

Ta-da! Retraining increased the accuracy of the Brain to 100%!

This is an extreme example, but it shows how Brain Builder can help you quickly and easily address some of the common pitfalls that people encounter when creating AI solutions.


After this walk-through, you should feel confident about the following subjects:

  • The importance of data and what makes a good dataset
  • How to evaluate your Brain's performance with a Brain Score
  • The impact of the order of training data, and how Retraining can improve Brain performance

Next: Walk-Through 3: Batch Training & Deployment >>