Introducing NIMCP: A Biologically-Inspired AI Model

After a year of development, the NIMCP (Neural Integrated Multimodal Cognitive Processor) project has finally come to fruition. NIMCP is an artificial brain built in C that trains six different types of neural networks simultaneously, with a unique approach to safety and ethics. In this blog post, we'll take a closer look at the architecture and features of NIMCP.

Training Multiple Neural Network Types Simultaneously

NIMCP is capable of training six different neural network types simultaneously, including:

1. Spiking Neural Networks (SNNs)

2. Liquid State Machines (LSMs)

3. Convolutional Neural Networks (CNNs)

4. Fourier Neural Networks (FNNs)

5. Hamiltonian Neural Networks (HNNs)

6. Adaptive Neural Networks (ANNs)

These networks are not only trained in parallel, but also interact with each other through learnable bridges, allowing for the exchange of information and gradient flow between them.

Safety and Ethics

One of the most interesting aspects of NIMCP is its safety and ethics module. Unlike traditional AI models, NIMCP's ethics module is not a learned weight that can be fine-tuned away or jailbroken. Instead, it's a function call in the inference code path, ensuring that the model's behavior is always aligned with the predetermined ethics rules. These rules can only be made stricter, not looser, providing a high degree of safety and reliability.

Emergent Properties and Developmental Stages

NIMCP's training methodology is based on a 4-stage developmental curriculum, inspired by human cognitive development:

1. Sensory stage: The model learns to process sensory data

2. Naming stage: The model learns to associate names with sensory data

3. Feedback stage: The model learns to provide feedback based on its associations

4. Reasoning stage: The model learns to reason and make decisions based on its feedback

Currently, the training is in the naming stage, with metrics updating every 60 seconds on the NIMCP website. The model has already achieved some impressive results, including 26 Hz firing rates with 67% sparsity in its SNN, within the range of mammalian cortical activity.

Code and Documentation

The NIMCP codebase is available on GitHub, with over 2,600 source files, 240 Python API methods, and 8 language bindings. The system runs on a single RTX 4000 GPU with 20 GB of VRAM. Eight technical papers on the NIMCP website provide a comprehensive overview of the math, training methodology, safety architecture, and emergent dynamics of the model.

In conclusion, NIMCP is a groundbreaking biologically-inspired AI model that offers a unique approach to safety, ethics, and development. Its ability to train multiple neural network types simultaneously and its emergent properties make it an exciting area of research. We look forward to seeing the continued development and progress of the NIMCP project.