With convolutional neural networks (CNN) becoming more of a commodity in the computer vision field, many have attempted to improve CNN in a bid to achieve better accuracy to a point that CNN accuracies have surpassed that of human's capabilities. However, with deeper networks, the number of computations and consequently the energy needed per classification has grown considerably. In this chapter, an iterative approach is introduced, which transforms the CNN from a single feed-forward network that processes a large image into a sequence of smaller networks, each processing a subsample of each image. Each smaller network combines the features extracted from all the earlier networks, to produce classification results. Such a multistage approach allows the CNN function to be dynamically approximated by creating the possibility of early termination and performing the classification with far fewer operations compared to a conventional CNN.

Chapter Contents:

• 9.1 Motivation
• 9.2 Background on CNN
• 9.3 Optimization of CNN
• 9.4 Iterative learning
• 9.4.1 Case study: iterative AlexNet
• 9.5 ICNN training schemes
• 9.5.1 Sequential training
• 9.5.2 Parallel training
• 9.6 Complexity analysis
• 9.7 Visualization
• 9.7.1 Background on CNN visualization
• 9.7.2 Visualizing features learned by ICNN
• 9.8 Contextual awareness in ICNN
• 9.8.1 Prediction rank
• 9.8.2 Pruning neurons in FC layers
• 9.8.3 Pruning filters in CONV layers
• 9.9 Policies for exploiting energy-accuracy trade-off in ICNN
• 9.9.1 Dynamic deadline (DD) policy for real-time applications
• 9.9.2 Thresholding policy (TP) for dynamic complexity reduction
• 9.9.2.1 Fixed thresholding policy
• 9.9.2.2 Variable thresholding policy
• 9.9.3 Context-aware pruning policy
• 9.9.3.1 Context-aware pruning policy for FC layer
• 9.9.3.2 Context-aware pruning policy for CONV layer
• 9.9.4 Pruning and thresholding hybrid policy
• 9.9.4.1 Fixed percentage PTHP
• 9.9.4.2 Confidence-tracking PTHP
• 9.9.5 Variable and dynamic bit-length selection
• 9.10 ICNN implementation results
• 9.10.1 Implementation framework
• 9.10.2 Dynamic deadline policy for real-time applications
• 9.10.3 Thresholding policy for dynamic complexity reduction
• 9.10.3.1 Fixed thresholding policy
• 9.10.3.2 Variable thresholding policy
• 9.10.4 Context-aware pruning policy for parameter reduction
• 9.10.4.1 Context-aware pruning policy for FC layer
• 9.10.4.2 Context-aware pruning policy for CONV layer
• 9.11 Pruning and thresholding hybrid policy
• 9.11.1 Fixed percentage PTHP
• 9.11.2 Confidence-tracking PTHP
• 9.11.3 Run-time and overall accuracy
• 9.11.3.1 Pruning and/or thresholding
• 9.11.3.2 Deadline-driven
• 9.12 Conclusions
• References

Inspec keywords: feature extraction; iterative methods; image classification; feedforward neural nets; real-time systems; convolutional neural nets; computer vision

Other keywords: feature extraction; feed-forward network; low power system; real-time system; classification; iterative convolutional neural network; image processing; computer vision; iterative CNN

Subjects: Computer vision and image processing techniques; Image recognition; Interpolation and function approximation (numerical analysis); Neural computing techniques; Interpolation and function approximation (numerical analysis)