Models
Models are the core of Decthings, and the goal of all the features we add to the platform is to enable you to create and run better models, easier.
What is a model?
Artificial intelligence (AI) allows a computer program to makes intelligent decisions based on data. In the end, AI is useful to us humans because we can connect it to automated systems. For example, an AI could monitor a construction site and alert if it sees a construction worker in danger. It could also be used for our daily tasks, such as for automatically switching on the lights when we enter a room.
In Decthings, the task for a model is to take some input data and transform it to some output data. The input data could for example be an image, and the output a text containing the AIs guess as to what the image depicts. In Decthings, this is called an evaluation. The model contains everything that is needed to train and evaluate your model. As you will soon see, this can either be a visual graph which contains the model configuration, or it can be code. A model can also be trained. When training, data is fed to the model. With powerful computers and a bunch of mathematics the model "learns" from this data, and if all goes to plan, it will in the future be better at performing its designated task.
In Decthings, models run in the cloud, meaning they are not executed on your own computer. Instead, the data is sent from your device, over the network, and to Decthings servers. The model is executed inside an isolated environment, and when the response is ready it is sent back to your device. Using the information in the response the device can then make an intelligent decision or perform a task using the information generated by the AI.
Two ways to create a model
The visual editor makes it easy to create both a simple and an advanced model. It is used to build a neural network, which is the most common type of AI. Neural networks are made up of layers, and the visual editor allows you to place different types of layers and connect them.
By selecting different types of layers and being clever about how you connect them you can make many different types of models. The editor makes no assumption about the type of model you are creating, and we have made it as generic as possible in order to allow you to create many different types of models. Here are a few types of models you can create using the visual editor:
- Classifiers, such as image classification, audio classification or classification from tabular data
- More advanced detection models, such as facial features detection, body pose recognition and sentiment analysis from text
- Models that predict the future, such as stock prices and future supply and demand
- Data generation, such as image generation and text-to-speech
You can also create a model by writing code. This is the ultimate flexibility, because you can really create any type of model you would like. The code for your model gets stored on Decthings servers, and when it is time to execute it, the code gets executed in a virtual machine. With code you can create all the above mentioned models and more.
To get started with models, take a look at the Visual editor quick start and Write a model in code.
How models work in Decthings
Decthings models are stored and executed in the cloud (cloud = Decthings servers). A model takes some arbitrary input data, performs some arbitrary computation, and returns some arbitrary output data.
At its core, a model is defined by two pieces of data - code and state.
- The model code determines the actual logic that is performed when the model is either evaluated or trained. This can for example be to multiply input numbers together, or to run a neural network. For models using the visual editor, there is still underlying code, but you don't edit it directly.
- The model state contains some binary information that the code can make use of - for example the weights and biases of a neural network. State is required because on its own, code is static. When we train our model, we want to update parameters within our model, which for a neural network means to update the weights and biases of the network. When the model is trained a new state is created, which in effect means that the updated weights and biases can now be used in the next evaluation.
When a request to evaluate the model reaches Decthings (either via the browser user interface or via the API), the code is executed in a virtual machine. Once the code is ready, the active state (e.g the weights and biases) is loaded into the model, as well as the evaluation input data that was provided by whoever requested the evaluation. The code performs its calculations and returns some evaluation output data which is then sent from Decthings to whoever started the evaluation.
When training a model we usually want to provide a dataset of training data. There are some pre-built datasets available to use and you can also create your own. When a request to train the model reaches Decthings, the code is started and the active state is loaded, in the same way as when evaluating. The code then receives the training data and can perform the training operation. Once complete, the code will output the new state data. As previously mentioned, for a neural network, this new state would contain the updated and trained weights and biases. To use this new and improved state you should mark it as the "active state". The next time an evaluation is performed, the new, trained weights and biases will be used and the model performance will likely be improved.
