Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce "deep compression", a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35× to 49× without affecting their accuracy. Our method first prunes the network by learning only the important connections. Next, we quantize the weights to enforce weight sharing, finally, we apply Huffman coding. After the first two steps we retrain the network to fine tune the remaining connections and the quantized centroids. Pruning, reduces the number of connections by 9× to 13×; Quantization then reduces the number of bits that represent each connection from 32 to 5. On the ImageNet dataset, our method reduced the storage required by AlexNet by 35×, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49× from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3× to 4× layerwise speedup and 3× to 7× better energy efficiency.

Oxidation of graphite produces graphite oxide, which is dispersible in water as individual platelets. After deposition onto Si/SiO2 substrates, chemical reduction produces graphene sheets. Electrical conductivity measurements indicate a 10000-fold increase in conductivity after chemical reduction to graphene. Tapping mode atomic force microscopy measurements show one to two layer graphene steps. Electrodes patterned onto a reduced graphite oxide film demonstrate a field effect response when the gate voltage is varied from +15 to -15 V. Temperature-dependent conductivity indicates that the graphene-like sheets exhibit semiconducting behavior.

We demonstrate detection of NO2 down to ppb levels using transistors based on both single and multiple In2O3 nanowires operating at room temperature. This represents orders-of-magnitude improvement over previously reported metal oxide film or nanowire/nanobelt sensors. A comparison between the single and multiple nanowire sensors reveals that the latter have numerous advantages in terms of great reliability, high sensitivity, and simplicity in fabrication. Furthermore, selective detection of NO2 can be readily achieved with multiple-nanowire sensors even with other common chemicals such as NH3, O2, CO, and H2 around.

We present an approach to use individual In2O3 nanowire transistors as chemical sensors working at room temperature. Upon exposure to a small amount of NO2 or NH3, the nanowire transistors showed a decrease in conductance up to six or five orders of magnitude and also substantial shifts in the threshold gate voltage. These devices exhibited significantly improved chemical sensing performance compared to existing solid-state sensors in many aspects, such as the sensitivity, the selectivity, the response time, and the lowest detectable concentrations. Furthermore, the recovery time of our devices can be shortened to just 30 s by illuminating the devices with UV light in vacuum.