Largescale Object Recognition with CUDAaccelerated Hierarchical Neural Networks

1
Large-scale Object Recognitionwith
CUDA-accelerated Hierarchical Neural Networks
Rafael Uetz and Sven Behnke
2
Motivation

• Ultimate goal
• An object recognition system with human-like
  recognition performance
• Our approach
• Basic model Biologically-inspired artificial
  neural network
• Very large neural network (i.e., large number of
  neurons)
• Realistic training and testing images from
  natural scenes
• Very large number of training images
• High-performance, parallel implementation for
  faster processing

2
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
3
Outline
The Locally-connected Neural Pyramid (LCNP) model
The LabelMe-12-50k dataset
Parallel implementation using the Nvidia CUDA
framework
Recognition results
3
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
4
The LCNP model
Output Layer
Layer 2
Neurons

Forward Propagation
Maps
Backpropagation of Error
Layer 1
Layer 0
4
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
5
Local connections

• Each neuron has a local receptive field
• No weight sharing is employed

Map of layer n1
Map of layer n
5
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
6
Training techniques

• Input maps in every layer (except the output
  layer)
• Edge-filtered channels in addition to the color
  channels
• Random translations during the training phase

6
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
7
The LabelMe-12-50k dataset

• Goal Creating a large, realistic dataset for
  object recognition
• Great variance in appearance, lighting
  conditions, and angle of view
• Extracted from LabelMe (Russell et al. 2008)
• 50,000 color JPEG images (40,000 for training
  and 10,000 for testing)
• 256x256 pixels
• 12 object classes
• 50 objects, 50 clutter

7
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
8
Parallel implementation

• Parallel implementation using the Nvidia CUDA
  framework
• GPU GeForce GTX 285, CPU Intel Core i7 940
• One epoch of LabelMe-12-50k training 27 sec
  (GPU), 37 min (CPU)

8
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
9
Recognition results

• Current state-of-the-art recognition rates
  without translations and additional input
  channels (Jarrett et al. 2009) MNIST 0.53
  NORB 5.6
• No significant overtraining with random
  translations

9
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
10
Future work

• Unsupervised pretraining with de-noising
  autoencoders
• Much larger training datasets
• Recurrent network structures for an iterative
  refinement of the output

10
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks
11
Conclusions

• The LCNP is a hierarchical, locally-connected
  neural network model with several input maps
  on every layer
• The parallel CUDA implementation allows for
  extremely large datasets as well as large
  network structures
• The LCNP achieves state-of-the-art performance
  on the NORB dataset
• LabelMe-12-50k is a new, challenging dataset of
  natural images (download http//www.ais.uni-bo
  nn.de/download/datasets.html)

Thank you!
11
R. Uetz and S. Behnke Large-scale Object
Recognition with CUDA-accelerated Hierarchical
Neural Networks