Artificial Intelligence (AI) is a lot more than 'prompts' and LLMs (chatGPT, AI search, etc). CyberSym Technologies has been building high end AI (including advanced neural networks) for 3 decades. AI is ideal for problems that require complex pattern recognition, prediction, or optimization using or derived from large amounts of data. Our proprietary AI software libraries include many neural network capabilities:

• highly flexible layers and layer structures for deep learning, including convolution and FHT
• capability for automatic model generation (AI4AI)
• linear, logistic, radial, ReLU, and boolean activation functions
• traditional backpropagation, Levenberg-Marquardt, and approximate matrix diagonalization algorithms for nonlinear regression
• support for temporal difference learning
• training includes automatic model pruning and augmentation, as well as simulated annealing for optimization
• parallel use of multiple models for output (ensemble AI)
• principal component and sensitivity analysis
• tools for generating training pattern sets from SQL databases

A lot goes into successful development of AI solutions.

Multidimensional datasets can often be sparse in certain dimensions, and training datasets are no exception. In some cases it is possible to skirt this problem by training individual layers on subsets of the data, and then clamping those layers while subsequently training other parts of the model.

Projects that involve ongoing state prediction/calculation may require temporal difference learning methods. Not only do our libraries incorporate temporal difference learning, we have extensive experience using these methods (and the regression that is necessary to make them successful).

We believe strongly in the concept of ethical AI. A lot of this begins with the training data set and good data science. In simple terms, the patterns selected for training and testing need to be representative, and there need to be enough of them (distinct) to justify the number of parameters in the model. Data should be obtained legally and should not incorporate privacy violations or cultural or racial appropriation. Trained AI models should be queried behind some kind of firewall or shell application that enforces fair and acceptable use policies. Attempting to train neural networks etc. so that they enforce acceptable use policy is a prescription for disaster. It is important for safety that neural networks undergoing training should never have write access to scratch memory or to the goal or task stack that underlies the development of the neural net. Finally it is important to recognize and avoid training feedback loops that can lead to model collapse.

Computational neural networks are ultimately methods for sophisticated data analysis. As such they need to follow good standards of data analysis (e.g., statistical learning principles and sound data science). At the same time, computational neural networks were designed to mimic biological neuronal networks of the sort found in the brain. It is therefore important that we understand computational neural networks in the context of what biological brains can and cannot do.

AI is actually many different things beyond just computational neural networks. CyberSym Technologies also has experience with knowledge bases, flow graphs, and hybrid AI such as procedural reasoning systems.