AI Modeling/軽量化
(On-Device AI)

Running DL Models with the same
performance in small device

Saving 17% of compression rate with the same
accuracy(90%) on image recognition task

  • Saving 17% of compression rate & Same Accuracy(90%)
  • Related patents : 3 PCT & 6 KR

競合他社

Nota

  • 同一正確度

    正確度

    既存モデル 97%
    SAME

    97%

    Raspberry Pi 3+
    基盤の顔面認識
    モデル基準

  • 高効率

    演算量

    既存モデル 14.4B
    -85.3%

    2.3B

    Raspberry Pi 3+
    基盤の顔面認識
    モデル基準

  • 速い速度

    推論速度

    既存モデル 0.38秒
    -77%

    0.08秒

    Raspberry Pi 3+
    基盤の顔面認識
    モデル基準

  • 省電力

    電力消費量

    既存モデル 100%
    -40%

    60%

    顧客での実使用の
    データ基準

  • 高いコスパ

    AIシステム費用

    既存モデル 100%
    -85%

    15%

    顧客での実使用の
    データ基準

Compare to Intel OpenVINO &
Nvidia TensorRT
Intel UP2
  • Intel ATOM x7-E3950
  • Intel HD Graphic 505
  • 4 GB
  • 20 ~ 30 W

* Object Detection on
Intel UP2 board

  • CPU
  • GPU
  • Memory
  • Power
NVIDIA Jetson TX2
  • ARM Cortex-A5(4core)
  • 256 Cuda core(Pascal)
  • 8 GB
  • 7.5 W

* Object Detection on
Nvidia Jetson TX2 board

Partnership

Nvidia Embedded Partner

Conventional AI model compresion

  • Pretrained Model
  • Compression
    Technique
    Pruning Quantization
    Knowledge Distillation NAS
  • Compressed Model

Problems of current
network compression

  • DL engineers manualy compress the model
  • Compression methods are developed in different places and forms
  • Hard to know which compression method or combination to use
  • Compression metric does not fit to practical metric

Nota’s NetsPresso(Automatic Model Compression Platform)

  • Problem Soving
    • - Automatic compression without manpower
    • - Combination of multiple compression methods
    • - Fitted metric for practical usage

    Nota’s Automatic AI Model Compression Platform : NetsPresso

  • Optimum compression platform for :
    • - Target task
    • - Target dataset
    • - Target device
    • - Target accuracy / latency / model size