4.2: MLE.2 ML Model Architecture

Process Definition

Purpose

MLE.2 Purpose: To establish an ML model architecture that satisfies the ML requirements.

Outcomes

Base Practices with AI Integration

Architecture Patterns for Embedded ML

Pattern Selection Guide

The diagram below presents a decision framework for selecting ML architecture patterns based on task type, latency constraints, and available hardware resources.

Model Architecture Specification

Architecture Document Template

# ML Model Architecture Specification
architecture:
  id: MLE-ARCH-001
  name: ADAS Vehicle Detection Model
  version: 2.1
  status: approved

  requirements:
    - MLE-ADAS-001  # Vehicle detection performance
    - MLE-ADAS-002  # Latency requirement
    - MLE-ADAS-003  # Memory constraint

  base_architecture: YOLOv8-nano
  modification: "Quantization-aware training, pruned for automotive"

  input_specification:
    format: RGB image
    resolution: [640, 384]
    dtype: uint8
    normalization: "[0, 255] → [0, 1]"
    preprocessing:
      - resize: "bilinear interpolation"
      - normalize: "divide by 255"

  output_specification:
    format: Detection boxes + confidence
    max_detections: 100
    classes:
      - vehicle_car
      - vehicle_truck
      - vehicle_motorcycle
      - vehicle_bus
    output_format:
      boxes: "[N, 4] (x_center, y_center, width, height)"
      scores: "[N] confidence scores"
      classes: "[N] class indices"

  model_structure:
    backbone:
      type: CSPDarknet-nano
      layers: 24
      channels: [16, 32, 64, 128]
      activation: SiLU

    neck:
      type: PANet
      feature_levels: [P3, P4, P5]

    head:
      type: Decoupled Head
      anchors: anchor-free
      outputs_per_level: 3

  quantization:
    type: INT8
    calibration: "Post-training quantization"
    calibration_dataset: "1000 representative samples"
    accuracy_loss: "< 1% mAP drop"

  resource_estimation:
    parameters: 3.2M
    flops: 8.7 GFLOPs
    model_size:
      fp32: 12.8 MB
      int8: 3.2 MB
    memory_footprint:
      inference: 48 MB
      peak: 64 MB

  hardware_target:
    platform: "NVIDIA Jetson Orin Nano"  # Illustrative; adapt to project-specific target
    accelerator: GPU
    inference_time: 22ms
    power: 7W

  traceability:
    requirements:
      - id: MLE-ADAS-001
        coverage: "Detection performance"
      - id: MLE-ADAS-002
        coverage: "Latency (22ms < 50ms)"
      - id: MLE-ADAS-003
        coverage: "Memory (64MB < 128MB)"

Model Component Specification

Layer-Level Documentation

The following diagram breaks down the ML model into its constituent layers and components, showing input/output shapes, parameter counts, and computational requirements at each stage.

Interface Specification

Input/Output Contract

Note: This C interface follows AUTOSAR conventions for consistency with embedded automotive software.

/**
 * @file ml_detection_interface.h
 * @brief ML Detection Model Interface
 * @trace MLE-ARCH-001
 */

#ifndef ML_DETECTION_INTERFACE_H
#define ML_DETECTION_INTERFACE_H

#include "Std_Types.h"

/*===========================================================================*/
/* TYPE DEFINITIONS                                                          */
/*===========================================================================*/

/** @brief Input image format */
typedef struct {
    uint8* data;           /**< RGB pixel data */
    uint16 width;          /**< Image width (640) */
    uint16 height;         /**< Image height (384) */
    uint8  channels;       /**< Number of channels (3) */
} ML_ImageInput_t;

/** @brief Detection result */
typedef struct {
    float32 x_center;      /**< Box center X (normalized 0-1) */
    float32 y_center;      /**< Box center Y (normalized 0-1) */
    float32 width;         /**< Box width (normalized 0-1) */
    float32 height;        /**< Box height (normalized 0-1) */
    float32 confidence;    /**< Detection confidence (0-1) */
    uint8   class_id;      /**< Class index */
} ML_Detection_t;

/** @brief Inference result */
typedef struct {
    ML_Detection_t detections[100];  /**< Detection array */
    uint8 num_detections;            /**< Number of valid detections */
    uint32 inference_time_us;        /**< Inference time in microseconds */
} ML_InferenceResult_t;

/*===========================================================================*/
/* API FUNCTIONS                                                             */
/*===========================================================================*/

/**
 * @brief Initialize ML model
 * @return E_OK on success
 * @timing Initialization: ~500ms
 * @trace MLE-ARCH-001
 */
Std_ReturnType ML_Detection_Init(void);

/**
 * @brief Run inference on image
 * @param input Input image
 * @param result Output detections
 * @return E_OK on success
 * @timing Max 50ms inference time
 * @trace MLE-ADAS-001, MLE-ADAS-002
 */
Std_ReturnType ML_Detection_Inference(
    const ML_ImageInput_t* input,
    ML_InferenceResult_t* result
);

/**
 * @brief Get model information
 * @param version Model version string
 * @param params Parameter count
 * @return E_OK on success
 */
Std_ReturnType ML_Detection_GetInfo(
    char* version,
    uint32* params
);

#endif /* ML_DETECTION_INTERFACE_H */

AI-Assisted Architecture Design

L2: Architecture Selection

AI Architecture Recommendation:
───────────────────────────────

Requirements Analysis:
• MLE-ADAS-001: Recall ≥ 99.9%, Precision ≥ 98%
• MLE-ADAS-002: Latency ≤ 50ms
• MLE-ADAS-003: Memory ≤ 128MB

Hardware Target:
• NVIDIA Jetson Orin Nano (GPU accelerated)
• Available memory: 128MB for ML
• Power budget: 10W

Recommended Architecture: YOLOv8-nano
─────────────────────────────────────

Rationale:
1. Real-time Performance
   - Expected latency: ~22ms on target
   - Well within 50ms requirement

2. Accuracy Profile
   - mAP@0.5: 0.87 (baseline)
   - With automotive fine-tuning: expected 0.90+

3. Memory Efficiency
   - INT8 model: 3.2MB
   - Inference memory: 48MB peak

4. Deployment Maturity
   - TensorRT support available
   - Automotive deployment references exist

Alternatives Considered:
────────────────────────
1. SSD MobileNet v2
   - Pro: Smaller model
   - Con: Lower accuracy (mAP 0.72)
   - Decision: Accuracy insufficient

2. EfficientDet-D0
   - Pro: Better accuracy (mAP 0.90)
   - Con: Higher latency (~80ms)
   - Decision: Latency exceeded

3. YOLOv8-small
   - Pro: Higher accuracy (mAP 0.92)
   - Con: 2x parameters, 2x memory
   - Decision: Could be future upgrade if needed

Human Review Required:
□ Confirm architecture selection
□ Review quantization strategy
□ Approve resource allocation

Resource Estimation

Computation Analysis

The diagram below estimates the computational resources required for model inference on the target hardware, including memory footprint, FLOPS, and latency budget allocation.

Work Products

Summary

MLE.2 ML Model Architecture:

• AI Level: L1-L2 (AI suggests, human selects)
• Primary AI Value: Architecture search, resource estimation
• Human Essential: Final architecture decision, constraints
• Key Outputs: Architecture specification, interface definition
• Focus: Embedded constraints (latency, memory, power)