Neurenix

Neurenix is ​​an AI framework optimized for embedded devices (Edge AI), with support for multiple GPUs and distributed clusters. The framework specializes in AI agents, with native support for multi-agent, reinforcement learning, and autonomous AI.

Social

Main Features

• Hot-swappable backend functionality:

  • Added DeviceManager class for runtime device switching
  • Created Genesis system for automatic hardware detection and selection
  • Modified Tensor class to support hot-swapping between devices

• ONNX support:

  • Implemented ONNXConverter for model import/export
  • Added convenience functions for easy ONNX integration
  • Support for converting between Neurenix and other ML frameworks

• API support:

  • Added RESTful, WebSocket, and gRPC server implementations
  • Created APIManager for centralized server management
  • Provided convenience functions for serving models

• Dynamic imports from neurenix.binding with NumPy fallbacks for activation functions:

  • relu, sigmoid, tanh, softmax, log_softmax, leaky_relu, elu, selu, gelu

• CPU implementations for BLAS operations:

  • GEMM, dot product, GEMV

• CPU implementations for convolution operations:

  • conv2d, conv_transpose2d

• Conditional compilation for hardware-specific operations:

  • CUDA, ROCm, and WebGPU support for BLAS and convolution operations
  • Proper error handling for unsupported hardware configurations

• Binding functions for tensor operations:

  • backward, no_grad, zero_grad, weight_decay

• WebAssembly SIMD and WASI-NN support for browser-based tensor operations

• Hardware acceleration backends:

  • Vulkan for cross-platform GPU acceleration
  • OpenCL for heterogeneous computing
  • oneAPI for Intel hardware acceleration
  • DirectML for Windows DirectX 12 acceleration
  • oneDNN for optimized deep learning primitives
  • MKL-DNN for Intel CPU optimization
  • TensorRT for NVIDIA GPU optimization

• Automatic quantization support:

  • INT8, FP16, and FP8 precision
  • Model pruning capabilities
  • Quantization-aware training
  • Post-training quantization with calibration

• Graph Neural Networks (GNNs):

  • Implemented various GNN layers (GCN, GAT, GraphSAGE, etc.)
  • Added pooling operations for graph data
  • Provided graph data structures and utilities
  • Implemented common GNN models

• Fuzzy Logic:

  • Added fuzzy sets with various membership functions
  • Implemented fuzzy variables and linguistic variables
  • Created fuzzy rule systems with different operators
  • Implemented Mamdani, Sugeno, and Tsukamoto inference systems
  • Added multiple defuzzification methods

• Federated Learning:

  • Implemented client-server architecture for federated learning
  • Added various aggregation strategies (FedAvg, FedProx, FedNova, etc.)
  • Implemented security mechanisms (secure aggregation, differential privacy)
  • Added utilities for client selection and model compression

• AutoML & Meta-learning:

  • Implemented hyperparameter search strategies (Grid, Random, Bayesian, Evolutionary)
  • Added neural architecture search capabilities
  • Implemented model selection and evaluation utilities
  • Created pipeline optimization tools
  • Added meta-learning algorithms for few-shot learning

• Distributed training technologies:

  • MPI for high-performance computing clusters
  • Horovod for distributed deep learning
  • DeepSpeed for large-scale model training

• Memory management technologies:

  • Unified Memory (UM) for seamless CPU-GPU memory sharing
  • Heterogeneous Memory Management (HMM) for advanced memory optimization

• Specialized hardware acceleration:

  • GraphCore IPU support for intelligence processing
  • FPGA support via multiple frameworks:
    • OpenCL for cross-vendor FPGA programming
    • Xilinx Vitis for Xilinx FPGAs
    • Intel OpenVINO for Intel FPGAs

• DatasetHub: mechanism that allows users to easily load datasets by providing a URL or file path

• CLI

• Continual Learning Module: Allows models to be retrained and updated with new data without forgetting previously learned information. Implements several techniques:

  • Elastic Weight Consolidation (EWC)
  • Experience Replay
  • L2 Regularization
  • Knowledge Distillation
  • Synaptic Intelligence

• Asynchronous and Interruptible Training Module: Provides functionality for asynchronous training with continuous checkpointing and automatic resume, even in unstable environments. Features include:

  • Continuous checkpointing with atomic writes
  • Automatic resume after interruptions
  • Resource monitoring and proactive checkpointing
  • Signal handling for graceful interruptions
  • Distributed checkpointing for multi-node training

• Docker Support:

  • Container management
  • Image building and management
  • Volume management
  • Network configuration
  • Registry integration

• Kubernetes Support:

  • Deployment management
  • Service configuration
  • Pod management
  • ConfigMap handling
  • Secret management
  • Job scheduling
  • Cluster management

• Apache Arrow Support:

  • Efficient in-memory columnar data format
  • Seamless conversion between Arrow tables and Neurenix tensors
  • Support for various data types and operations

• Parquet Support:

  • High-performance columnar storage format
  • Reading and writing Parquet files
  • Dataset management with partitioning support
  • Integration with Arrow for efficient data processing

• SHAP (SHapley Additive exPlanations):

  • KernelSHAP for model-agnostic explanations
  • TreeSHAP for tree-based models
  • DeepSHAP for deep learning models

• LIME (Local Interpretable Model-agnostic Explanations):

  • Support for tabular, text, and image data
  • Customizable sampling and kernel functions

• Additional Explanation Techniques

  • Feature importance analysis
  • Partial dependence plots
  • Counterfactual explanations
  • Activation visualization

• Multi-Scale Model Architectures:

  • MultiScaleModel: Base class for multi-scale architectures
  • PyramidNetwork: Feature pyramid network implementation
  • UNet: U-Net architecture with skip connections

• Multi-Scale Pooling Operations:

  • MultiScalePooling: Base class for multi-scale pooling
  • PyramidPooling: Pyramid pooling module (PPM)
  • SpatialPyramidPooling: Spatial pyramid pooling (SPP)

• Feature Fusion Mechanisms:

  • FeatureFusion: Base class for feature fusion
  • ScaleFusion: Fusion of features from different scales
  • AttentionFusion: Attention-based fusion of multi-scale features

• Multi-Scale Transformations:

  • MultiScaleTransform: Base class for multi-scale transforms
  • Rescale: Rescaling to multiple scales
  • PyramidDownsample: Pyramid downsampling for multi-scale representations
  • GaussianPyramid: Gaussian pyramid implementation
  • LaplacianPyramid: Laplacian pyramid implementation

• Zero-shot Learning

• NVIDIA Tensor Cores support

• WebAssembly multithreaded support

• gRPC-Streaming support

• Neuroevolution + Evolutionary Algorithms Support:

  • Genetic Algorithms: Implementation of population-based optimization with selection, crossover, and mutation operators
  • NEAT (NeuroEvolution of Augmenting Topologies): Algorithm for evolving both neural network topologies and weights
  • HyperNEAT: Extension of NEAT that uses CPPNs to generate large-scale neural networks with geometric regularities
  • CMA-ES (Covariance Matrix Adaptation Evolution Strategy): State-of-the-art evolutionary algorithm for continuous optimization
  • Evolution Strategies: Implementation of various ES variants with adaptive learning rates and population-based optimization

• Hybrid Neuro-Symbolic Models Support:

  • Symbolic reasoning components with rule-based inference
  • Neural-symbolic integration with multiple interaction modes
  • Differentiable logic with support for fuzzy and probabilistic logic
  • Knowledge distillation between symbolic and neural systems
  • Advanced reasoning paradigms (constraint satisfaction, logical inference, abductive/deductive/inductive reasoning)

• Multi-Agent Systems (MAS) Support:

  • Agent framework with reactive, deliberative, and hybrid agent types
  • Communication protocols and message passing infrastructure
  • Coordination mechanisms (task allocation, auctions, contract nets, voting)
  • Multi-agent learning algorithms (independent learners, joint action learners, team learning)
  • Environment abstractions for agent interaction

Documentation

Neurenix Documentation

License

This project is licensed under the Apache License 2.0.