CSX-AI Solutions

Solutions through examples of use and demos

1. Perception

Deep learning models such as convolutional neural networks and vision transformers have become the status quo for perception modules of many systems, such as cars, drones, and trains, for detecting objects and understanding the environment. Multiple sensory modalities can be used together to collect inputs to a perception model.

First, our CSX platform can support the development and training of a perception model with high accuracy. Second, in cases where the sensor modalities may exhibit either benign (e.g., white noise, color shifting, or motion blur) or malicious (e.g., adversarial perturbations or patches) attacks, the robust accuracy of the perception model will deteriorate significantly. Our CSX platform offers functionalities to detect and analyze the risks and improve the model accordingly.

2. Guidance, Navigation, and Control

Modern vehicles are increasingly driven either partially or entirely by deep learning models, which determine the location of the vehicles, conduct motion planning, and control the movement of the cars. The deep learning models can be, e.g., the classifiers implemented by convolutional neural networks and others, the reinforcement learning models, and so on.

As we discussed in perception models, these models also suffer from generalization and robustness risks, and our CSX platform facilitates their safe and robust development. In addition, uncertainties of these models may also lead to a lack of confidence in driving the vehicles. Our CSX platform offers the functionalities to measure the uncertainties, reduce the uncertainties, and improve the performance of the models.

3. Security and Privacy in Cyberspace

AI has been applied to various cyber applications, acting as either attackers or defenders or both in cybersecurity scenarios and providing the infrastructure for constructing useful applications that protect sensitive information.

A notable example of the latter case is federated/distributed learning, where local agents collectively learn a model without transferring the data collected locally. CSX platform integrates functionalities to facilitate the development of such AI systems with provable guarantees on their security and privacy performance.

4. LLMs for Autonomous Driving

Large language models (LLMs) have been widely studied as intelligent agents in autonomous driving scenarios, where some information, such as the bird-view of a vehicle, is translated into a prompt on which an LLM will act by outputting a motion plan or a control behavior of the vehicle.

While promising, it is known that LLMs may suffer from various safety and trustworthiness issues similar to other simpler deep learning models. Our CSX platform will focus on the guardrails, a safeguarding mechanism that can effectively detect and correct the unexpected behaviors of the LLMs