Key Terminology to Know

A glossary of key terminology to know when working with the AI Studio platform.

Projects

Projects in AI Studio are self-contained, foundational units that contain models, labels, datasets and fashion data. Projects are used for managing, processing, modeling and analyzing fashion data. They contain fashion datasets that are used to build and deploy AI models for solving specific, real-life fashion problems.

Projects are easy to create and maintain. Choose from among the available project types, define your specific problem, follow the simple instructions and use the easy-to-use interconnected suite of tools to create your project and build your AI models. You can create as many projects as you like to solve your compelling fashion problems.

You can create multiple models, add datasets and create labels specific to each project. Multiple models can be trained, evaluated, retrained with different datasets, compared with other models, deployed and archived within the project itself.

(Insert flowchart or infographic)

Labels

Labels in AI Studio are like keywords used to annotate, categorize and tag inputs for specific fashion attributes. In AI Studio, you add labels when creating projects. However, if you want to retrain the model or add labels, you can do so with ease. Labels are critical for training machines to make effective predictions. Generally speaking, AI models/ machines cannot predict labels that are not defined.

For instance, you are building an image tagging project to categorize inputs based on color of garments. Here, the labels will be the list of colors you want the model to identify and tag. If you train the model to identify 5 colors, then it will only identify those colors unless you add more labels and train the model to tag those.

Categories

Categories annotate the type of fashion objects the project focuses on. For instance, top wear, bottom wear, hats, outerwear, bags, footwear, one-piece, etc. In pre-processing, the categories are identified, and bounding boxes are added to focus on these areas of interest. This helps prepare the data for building models and enables the building of robust fashion AI models.

Dataset

Datasets form the core of the AI models that are built with the AI Studio platform. The fashion data contained in datasets are the inputs you provide to the AI Studio platform to build fashion AI models to solve real-life challenges. The dataset differs based on the problems you are solving and could be images, text, spreadsheets, etc. Uploading, managing and maintaining datasets is simple and intuitive.

Datasets used in AI Studio can be categorized as:

• Training Dataset

• Evaluation Dataset

Training Dataset

Training datasets provide the foundation for building fashion AI models. The machine detects and learns patterns in the training datasets by applying different parameters, classifiers and algorithms. Given the criticality of training datasets in building models, they need to be kept as close to the datasets on which predictions will be made. While the training dataset must contain at least 100 images per label, 300+ images per label are recommended where greater nuances are involved. This enables you in improving the accuracy of the model.

The training dataset must contain images with diverse variations. For instance, if the training dataset is of human models wearing fashion items, then the images should contain variations such as multiple angles, poses and backgrounds. If the training dataset is of print swatches, then the variations could include various colors, patterns, blocks, concepts, etc.

Evaluation Dataset

Evaluation datasets are created at a project level to set a benchmark to measure the performance of multiple models. Evaluation datasets are used to determine the error rate, accuracy and performance of the fashion AI model trained. It tells you if the model is applying the right parameters, classifiers and detectors to the problem at hand. Armed with insights, statistics and suggestions for improvement, you can retrain a model, train another model, compare multiple trained models and choose a high-performance AI model to deploy.

Model

Models are mathematical algorithms that are trained using fashion data to replicate actions that a human expert would take in fulfilling the tasks at hand using the same information provided. For instance, identifying and tagging colors of garments, categorizing footwear, generating collages, etc.

AI Studio makes it hassle-free for domain experts to train, evaluate, compare, improve, deploy and archive AI models. Domain experts can custom build AI models to solve fashion problems specific to their organization/ role/ field of work. The process is intuitive and iterative to ensure that the models are robust, accurate and effective.

Training

Training is the process through which you help the model learn from the fashion data in the training datasets. It is during training that the machine detects patterns in the dataset by applying different parameters, classifiers and algorithms. With AI Studio, domain experts can train models in just a few clicks, without writing a single line of code. Retraining with updated or new datasets is as hassle-free as training models. With the help of the insights and statistics, you can keep iterating the model to achieve higher accuracy and performance before deployment.

Evaluation

Evaluation is the process of testing and analyzing the trained model for accuracy, errors and performance. AI Studio provides an aggregate set of evaluation metrics such as confusion matrices, proponents and opponents, accuracy scores, violin plots, TSNE plots and PR scores, among others. Using these evaluation metrics, you can easily understand how well the model will perform for each of the labels. You can iterate accordingly to strengthen the model.

Deployment

Deployment refers to the application of your production-ready AI model to your workflow where it expedites manual, data-intensive tasks and solves your challenges. Deployment in AI Studio is as simple as copying the URL and applying it to the target environment.

we plan to offer more ways of deployment - urls, apis, ability for user to download the model, etc. Maybe we can mention we are working on many useful ways to deploy.

Deployment Status

The deployment status tells you whether or not models within a project have been deployed.

Insights

AI Studio takes a data-centric approach to Machine Learning. Actionable insights are evaluation metrics provided by the AI Studio platform. These enable fashion users to understand how well the trained model is performing and thereon iterate to make the model more accurate and effective. Insights include accuracy scores, accuracy by image and labels, statistical metrics along with suggestions for improvement.

Accuracy

The accuracy of the model informs the user how precise and correct the AI model is at performing the trained task. Higher the accuracy rate, the better the performance of the model.

Model Robustness

An AI model is considered robust if it consistently produces accurate results even when one or more of the variables are changed.

Test

Testing enables you to evaluate the performance of the trained and iterated model on a fresh dataset. This helps you to check the model for biases, effectiveness or bugs before deploying the model into your production workflow. With the help of testing, you can prevent such issues before the model goes into production.