Developing AI methods to encode non-lattice-structured data is one main challenge in current AI systems. Such data structures are common in the scientific applications to be explored in A3D3. AI researchers have to carefully consider the specifics of the data irregularity, the chosen model architecture, and the system constraints.
A3D3 proposes to investigate optimal encoders for these non-lattice-structured scientific data. Within the A3D3 science drivers, fast identification of sporadic events from large-scale data streams provides a unique opportunity for discovery.
Traditional selection strategies use either subjective heuristics (e.g., a high energy collision or a bright transient) or model-based classifiers that are trained with specific labels. Such strategies exclude unconventional signatures with unexpected properties where manually defined heuristics and labeled examples are insufficient to explain them.
Decoders based on encoded data can be designed to overcome these challenges. A3D3 will explore semi-supervised/unsupervised learning methods for effective anomaly, or out-of-class, detection. Recent work on variational autoencoders (VAEs) and generative models learn the distribution of the input data and detect anomalies based on their likelihood scores. Hardware-AI co-design is the tight coupling of the design of AI algorithms with the domain science, hardware, and system constraints. A hallmark of co-design is either an encoding of the design constraints directly in the cost function of the training algorithm or a fast-to-evaluate approximation of the acceptability of an on-device solution to allow for fast exploration and evaluation.
Our primary focus is on achieving low latency, real-time processing of scientific data. A3D3 aims to design end-to-end workflows integrated with user-friendly tools for hardware-aware co-design of AI algorithms.