Neural networks are becoming the preferred predictive model, especially for dynamic data that moves and changes across space and time. The amount of data and compute they consume is staggering, and engineers now build systems that analyze any kind of data — big or small, image or text, structured or streaming. This pushes the limits of hardware and software, so it has become necessary to optimize neural network computation from many angles: training time, inference time, model size (depth and width), CPU operations, GPU arithmetic, memory, storage and energy.
We especially seek "systems" papers presenting algorithms, data structures, functions, language extensions and optimizations that work well inside popular network libraries — particularly Python and C++ — including papers that build new AI libraries callable from Python or C++.
Call for papers
Topics
A partial, non-exclusive list:
- Reducing the number of iterations
- Early stopping for convergence or numerical issues
- Improving SGD with 2nd-order methods to accelerate convergence
- Sparse matrix multiplication in neural network layers
- Tensor storage, slicing and retrieval
- Compressed storage and reduced floating-point precision
- Stacking or pipelining networks instead of one monolithic net
- Understanding neural network circuits with data subsets
- Low-resource inference on edge devices
- Debugging neuron activations
- Swapping activation functions
- Trading slight accuracy loss for significant speed gains
- Identifying bottlenecks in large networks (e.g. transformers)
- Finding classical ML models that beat neural nets in specific cases
History
Previous editions
OPTINET is a revamped edition of the DEMAI workshop, held successfully in 2023 and 2024 with a healthy set of submissions, accepted presentations and papers invited for journal submission. Given new trends — more complex activation functions, LLMs, larger-scale networks, new-generation GPUs and faster storage — we have refocused on neural networks as the main challenge, with database and big data aspects as a supporting theme.