Instead of empirical design and trial and errors you can use formal tools to ensure your neural network performance success.
Choosing a neural network architecture has a major impact on its performance. In order to have good performance it is necessary to optimize any chosen architecture to the task at hand. This is a complex and time-consuming task with no guarantee of success once you move into production. Being able to control step by step the influence of any change of the architecture on your system performance will help optimizing it quicker and with greater chance of success.
Understanding the inner dynamics of a neural network will help your engineers save up a lot of time during the design phase.
Confronting a neural network to its performance is not only about statistical measurements. It is also about explaining why some tests are failing and some others do not. Tracing the behavior step by step may be easy but making sense of it is very hard. Only observing value flowing through the network of neurons is not appropriate. To have a better insight of why a neural network is not performant, you need to know at which point it starts failing. For example you need to understand at which point the system is starting to hesitate between two different answers. Pinpointing the layers where the flaw occurs helps you improve the design.
Using Saimple during the process of optimizing an architecture allows you to discover quicker a robust solution that matches your requirements.
When choosing architecture for your network, performance is the first concern to be addressed. Using classical metric, the engineers verify the robustness of the network and make an opinion on what architecture seems better suited. Improving on an existing architecture by optimizing its hyper-parameters is however a challenge. By doing it by hand or using stochastic techniques, the outcome can be costly while being both unpredictable and disappointing in terms of robustness.
Saimple is a tool capable of evaluating the robustness of a network at any stage it is in. Using it during your optimization process saves you time since you can take the robustness into account all along the way of any proposed solution.
By making the network explainable step by step you can guide the design, justify every choice you made and document accordingly the outcome of your design phase.
Each layer of a neural network plays a role in the larger model that has been trained. But it is often difficult to perceive how each one has contributed (either positively or negatively). When the neural network has yet to achieve the targeted performance it can be a challenge to discover why the overall network does not work. Sometimes it is because of the data, sometimes because of the architecture.
With Saimple you can see directly in the architecture how a result is made and why. At each layer you have the information to understand how each one is impacting the process.
