Matching Items (104)
Description
Nearly 60% of the world population uses a mobile phone, which is typically powered by a system-on-chip (SoC). While the mobile platform capabilities range widely, responsiveness, long battery life and reliability are common design concerns that are crucial to remain competitive. Consequently, state-of-the-art mobile platforms have become highly heterogeneous by combining a powerful SoC with numerous other resources, including display, memory, power management IC, battery and wireless modems. Furthermore, the SoC itself is a heterogeneous resource that integrates many processing elements, such as CPU cores, GPU, video, image, and audio processors. Therefore, CPU cores do not dominate the platform power consumption under many application scenarios.
Competitive performance requires higher operating frequency, and leads to larger power consumption. In turn, power consumption increases the junction and skin temperatures, which have adverse effects on the device reliability and user experience. As a result, allocating the power budget among the major platform resources and temperature control have become fundamental consideration for mobile platforms. Dynamic thermal and power management algorithms address this problem by putting a subset of the processing elements or shared resources to sleep states, or throttling their frequencies. However, an adhoc approach could easily cripple the performance, if it slows down the performance-critical processing element. Furthermore, mobile platforms run a wide range of applications with time varying workload characteristics, unlike early generations, which supported only limited functionality. As a result, there is a need for adaptive power and performance management approaches that consider the platform as a whole, rather than focusing on a subset. Towards this need, our specific contributions include (a) a framework to dynamically select the Pareto-optimal frequency and active cores for the heterogeneous CPUs, such as ARM big.Little architecture, (b) a dynamic power budgeting approach for allocating optimal power consumption to the CPU and GPU using performance sensitivity models for each PE, (c) an adaptive GPU frame time sensitivity prediction model to aid power management algorithms, and (d) an online learning algorithm that constructs adaptive run-time models for non-stationary workloads.
Competitive performance requires higher operating frequency, and leads to larger power consumption. In turn, power consumption increases the junction and skin temperatures, which have adverse effects on the device reliability and user experience. As a result, allocating the power budget among the major platform resources and temperature control have become fundamental consideration for mobile platforms. Dynamic thermal and power management algorithms address this problem by putting a subset of the processing elements or shared resources to sleep states, or throttling their frequencies. However, an adhoc approach could easily cripple the performance, if it slows down the performance-critical processing element. Furthermore, mobile platforms run a wide range of applications with time varying workload characteristics, unlike early generations, which supported only limited functionality. As a result, there is a need for adaptive power and performance management approaches that consider the platform as a whole, rather than focusing on a subset. Towards this need, our specific contributions include (a) a framework to dynamically select the Pareto-optimal frequency and active cores for the heterogeneous CPUs, such as ARM big.Little architecture, (b) a dynamic power budgeting approach for allocating optimal power consumption to the CPU and GPU using performance sensitivity models for each PE, (c) an adaptive GPU frame time sensitivity prediction model to aid power management algorithms, and (d) an online learning algorithm that constructs adaptive run-time models for non-stationary workloads.
ContributorsGupta, Ujjwala (Author) / Ogras, Umit Y. (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Kishinevsky, Michael (Committee member) / Dutt, Nikil (Committee member) / Arizona State University (Publisher)
Created2018
Description
Deep neural networks (DNN) have shown tremendous success in various cognitive tasks, such as image classification, speech recognition, etc. However, their usage on resource-constrained edge devices has been limited due to high computation and large memory requirement.
To overcome these challenges, recent works have extensively investigated model compression techniques such as element-wise sparsity, structured sparsity and quantization. While most of these works have applied these compression techniques in isolation, there have been very few studies on application of quantization and structured sparsity together on a DNN model.
This thesis co-optimizes structured sparsity and quantization constraints on DNN models during training. Specifically, it obtains optimal setting of 2-bit weight and 2-bit activation coupled with 4X structured compression by performing combined exploration of quantization and structured compression settings. The optimal DNN model achieves 50X weight memory reduction compared to floating-point uncompressed DNN. This memory saving is significant since applying only structured sparsity constraints achieves 2X memory savings and only quantization constraints achieves 16X memory savings. The algorithm has been validated on both high and low capacity DNNs and on wide-sparse and deep-sparse DNN models. Experiments demonstrated that deep-sparse DNN outperforms shallow-dense DNN with varying level of memory savings depending on DNN precision and sparsity levels. This work further proposed a Pareto-optimal approach to systematically extract optimal DNN models from a huge set of sparse and dense DNN models. The resulting 11 optimal designs were further evaluated by considering overall DNN memory which includes activation memory and weight memory. It was found that there is only a small change in the memory footprint of the optimal designs corresponding to the low sparsity DNNs. However, activation memory cannot be ignored for high sparsity DNNs.
To overcome these challenges, recent works have extensively investigated model compression techniques such as element-wise sparsity, structured sparsity and quantization. While most of these works have applied these compression techniques in isolation, there have been very few studies on application of quantization and structured sparsity together on a DNN model.
This thesis co-optimizes structured sparsity and quantization constraints on DNN models during training. Specifically, it obtains optimal setting of 2-bit weight and 2-bit activation coupled with 4X structured compression by performing combined exploration of quantization and structured compression settings. The optimal DNN model achieves 50X weight memory reduction compared to floating-point uncompressed DNN. This memory saving is significant since applying only structured sparsity constraints achieves 2X memory savings and only quantization constraints achieves 16X memory savings. The algorithm has been validated on both high and low capacity DNNs and on wide-sparse and deep-sparse DNN models. Experiments demonstrated that deep-sparse DNN outperforms shallow-dense DNN with varying level of memory savings depending on DNN precision and sparsity levels. This work further proposed a Pareto-optimal approach to systematically extract optimal DNN models from a huge set of sparse and dense DNN models. The resulting 11 optimal designs were further evaluated by considering overall DNN memory which includes activation memory and weight memory. It was found that there is only a small change in the memory footprint of the optimal designs corresponding to the low sparsity DNNs. However, activation memory cannot be ignored for high sparsity DNNs.
ContributorsSrivastava, Gaurav (Author) / Seo, Jae-Sun (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Berisha, Visar (Committee member) / Arizona State University (Publisher)
Created2018
Description
Vision processing on traditional architectures is inefficient due to energy-expensive off-chip data movements. Many researchers advocate pushing processing close to the sensor to substantially reduce data movements. However, continuous near-sensor processing raises the sensor temperature, impairing the fidelity of imaging/vision tasks.
The work characterizes the thermal implications of using 3D stacked image sensors with near-sensor vision processing units. The characterization reveals that near-sensor processing reduces system power but degrades image quality. For reasonable image fidelity, the sensor temperature needs to stay below a threshold, situationally determined by application needs. Fortunately, the characterization also identifies opportunities -- unique to the needs of near-sensor processing -- to regulate temperature based on dynamic visual task requirements and rapidly increase capture quality on demand.
Based on the characterization, the work proposes and investigate two thermal management strategies -- stop-capture-go and seasonal migration -- for imaging-aware thermal management. The work present parameters that govern the policy decisions and explore the trade-offs between system power and policy overhead. The work's evaluation shows that the novel dynamic thermal management strategies can unlock the energy-efficiency potential of near-sensor processing with minimal performance impact, without compromising image fidelity.
The work characterizes the thermal implications of using 3D stacked image sensors with near-sensor vision processing units. The characterization reveals that near-sensor processing reduces system power but degrades image quality. For reasonable image fidelity, the sensor temperature needs to stay below a threshold, situationally determined by application needs. Fortunately, the characterization also identifies opportunities -- unique to the needs of near-sensor processing -- to regulate temperature based on dynamic visual task requirements and rapidly increase capture quality on demand.
Based on the characterization, the work proposes and investigate two thermal management strategies -- stop-capture-go and seasonal migration -- for imaging-aware thermal management. The work present parameters that govern the policy decisions and explore the trade-offs between system power and policy overhead. The work's evaluation shows that the novel dynamic thermal management strategies can unlock the energy-efficiency potential of near-sensor processing with minimal performance impact, without compromising image fidelity.
ContributorsKodukula, Venkatesh (Author) / LiKamWa, Robert (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Brunhaver, John (Committee member) / Arizona State University (Publisher)
Created2019
Description
Articial Neural Network(ANN) has become a for-bearer in the field of Articial Intel-
ligence. The innovations in ANN has led to ground breaking technological advances
like self-driving vehicles,medical diagnosis,speech Processing,personal assistants and
many more. These were inspired by evolution and working of our brains. Similar
to how our brain evolved using a combination of epigenetics and live stimulus,ANN
require training to learn patterns.The training usually requires a lot of computation
and memory accesses. To realize these systems in real embedded hardware many
Energy/Power/Performance issues needs to be solved. The purpose of this research
is to focus on methods to study data movement requirement for generic Neural Net-
work along with the energy associated with it and suggest some ways to improve the
design.Many methods have suggested ways to optimize using mix of computation and
data movement solutions without affecting task accuracy. But these methods lack a
computation model to calculate the energy and depend on mere back of the envelope calculation. We realized that there is a need for a generic quantitative analysis
for memory access energy which helps in better architectural exploration. We show
that the present architectural tools are either incompatible or too slow and we need
a better analytical method to estimate data movement energy. We also propose a
simplistic yet effective approach that is robust and expandable by users to support
various systems.
ligence. The innovations in ANN has led to ground breaking technological advances
like self-driving vehicles,medical diagnosis,speech Processing,personal assistants and
many more. These were inspired by evolution and working of our brains. Similar
to how our brain evolved using a combination of epigenetics and live stimulus,ANN
require training to learn patterns.The training usually requires a lot of computation
and memory accesses. To realize these systems in real embedded hardware many
Energy/Power/Performance issues needs to be solved. The purpose of this research
is to focus on methods to study data movement requirement for generic Neural Net-
work along with the energy associated with it and suggest some ways to improve the
design.Many methods have suggested ways to optimize using mix of computation and
data movement solutions without affecting task accuracy. But these methods lack a
computation model to calculate the energy and depend on mere back of the envelope calculation. We realized that there is a need for a generic quantitative analysis
for memory access energy which helps in better architectural exploration. We show
that the present architectural tools are either incompatible or too slow and we need
a better analytical method to estimate data movement energy. We also propose a
simplistic yet effective approach that is robust and expandable by users to support
various systems.
ContributorsChowdary, Hidayatullah (Author) / Cao, Yu (Thesis advisor) / Seo, JaeSun (Committee member) / Chakrabarti, Chaitali (Committee member) / Arizona State University (Publisher)
Created2018
Description
Hardware implementation of deep neural networks is earning significant importance nowadays. Deep neural networks are mathematical models that use learning algorithms inspired by the brain. Numerous deep learning algorithms such as multi-layer perceptrons (MLP) have demonstrated human-level recognition accuracy in image and speech classification tasks. Multiple layers of processing elements called neurons with several connections between them called synapses are used to build these networks. Hence, it involves operations that exhibit a high level of parallelism making it computationally and memory intensive. Constrained by computing resources and memory, most of the applications require a neural network which utilizes less energy. Energy efficient implementation of these computationally intense algorithms on neuromorphic hardware demands a lot of architectural optimizations. One of these optimizations would be the reduction in the network size using compression and several studies investigated compression by introducing element-wise or row-/column-/block-wise sparsity via pruning and regularization. Additionally, numerous recent works have concentrated on reducing the precision of activations and weights with some reducing to a single bit. However, combining various sparsity structures with binarized or very-low-precision (2-3 bit) neural networks have not been comprehensively explored. Output activations in these deep neural network algorithms are habitually non-binary making it difficult to exploit sparsity. On the other hand, biologically realistic models like spiking neural networks (SNN) closely mimic the operations in biological nervous systems and explore new avenues for brain-like cognitive computing. These networks deal with binary spikes, and they can exploit the input-dependent sparsity or redundancy to dynamically scale the amount of computation in turn leading to energy-efficient hardware implementation. This work discusses configurable spiking neuromorphic architecture that supports multiple hidden layers exploiting hardware reuse. It also presents design techniques for minimum-area/-energy DNN hardware with minimal degradation in accuracy. Area, performance and energy results of these DNN and SNN hardware is reported for the MNIST dataset. The Neuromorphic hardware designed for SNN algorithm in 28nm CMOS demonstrates high classification accuracy (>98% on MNIST) and low energy (51.4 - 773 (nJ) per classification). The optimized DNN hardware designed in 40nm CMOS that combines 8X structured compression and 3-bit weight precision showed 98.4% accuracy at 33 (nJ) per classification.
ContributorsKolala Venkataramanaiah, Shreyas (Author) / Seo, Jae-Sun (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Cao, Yu (Committee member) / Arizona State University (Publisher)
Created2018
Description
Medical ultrasound imaging is widely used today because of it being non-invasive and cost-effective. Flow estimation helps in accurate diagnosis of vascular diseases and adds an important dimension to medical ultrasound imaging. Traditionally flow estimation is done using Doppler-based methods which only estimate velocity in the beam direction. Thus when blood vessels are close to being orthogonal to the beam direction, there are large errors in the estimation results. In this dissertation, a low cost blood flow estimation method that does not have the angle dependency of Doppler-based methods, is presented.
First, a velocity estimator based on speckle tracking and synthetic lateral phase is proposed for clutter-free blood flow.
Speckle tracking is based on kernel matching and does not have any angle dependency. While velocity estimation in axial dimension is accurate, lateral velocity estimation is challenging due to reduced resolution and lack of phase information. This work presents a two tiered method which estimates the pixel level movement using sum-of-absolute difference, and then estimates the sub-pixel level using synthetic phase information in the lateral dimension. Such a method achieves highly accurate velocity estimation with reduced complexity compared to a cross correlation based method. The average bias of the proposed estimation method is less than 2% for plug flow and less than 7% for parabolic flow.
Blood is always accompanied by clutter which originates from vessel wall and surrounding tissues. As magnitude of the blood signal is usually 40-60 dB lower than magnitude of the clutter signal, clutter filtering is necessary before blood flow estimation. Clutter filters utilize the high magnitude and low frequency features of clutter signal to effectively remove them from the compound (blood + clutter) signal. Instead of low complexity FIR filter or high complexity SVD-based filters, here a power/subspace iteration based method is proposed for clutter filtering. Excellent clutter filtering performance is achieved for both slow and fast moving clutters with lower complexity compared to SVD-based filters. For instance, use of the proposed method results in the bias being less than 8% and standard deviation being less than 12% for fast moving clutter when the beam-to-flow-angle is $90^o$.
Third, a flow rate estimation method based on kernel power weighting is proposed. As the velocity estimator is a kernel-based method, the estimation accuracy degrades near the vessel boundary. In order to account for kernels that are not fully inside the vessel, fractional weights are given to these kernels based on their signal power. The proposed method achieves excellent flow rate estimation results with less than 8% bias for both slow and fast moving clutters.
The performance of the velocity estimator is also evaluated for challenging models. A 2D version of our two-tiered method is able to accurately estimate velocity vectors in a spinning disk as well as in a carotid bifurcation model, both of which are part of the synthetic aperture vector flow imaging (SA-VFI) challenge of 2018. In fact, the proposed method ranked 3rd in the challenge for testing dataset with carotid bifurcation. The flow estimation method is also evaluated for blood flow in vessels with stenosis. Simulation results show that the proposed method is able to estimate the flow rate with less than 9% bias.
First, a velocity estimator based on speckle tracking and synthetic lateral phase is proposed for clutter-free blood flow.
Speckle tracking is based on kernel matching and does not have any angle dependency. While velocity estimation in axial dimension is accurate, lateral velocity estimation is challenging due to reduced resolution and lack of phase information. This work presents a two tiered method which estimates the pixel level movement using sum-of-absolute difference, and then estimates the sub-pixel level using synthetic phase information in the lateral dimension. Such a method achieves highly accurate velocity estimation with reduced complexity compared to a cross correlation based method. The average bias of the proposed estimation method is less than 2% for plug flow and less than 7% for parabolic flow.
Blood is always accompanied by clutter which originates from vessel wall and surrounding tissues. As magnitude of the blood signal is usually 40-60 dB lower than magnitude of the clutter signal, clutter filtering is necessary before blood flow estimation. Clutter filters utilize the high magnitude and low frequency features of clutter signal to effectively remove them from the compound (blood + clutter) signal. Instead of low complexity FIR filter or high complexity SVD-based filters, here a power/subspace iteration based method is proposed for clutter filtering. Excellent clutter filtering performance is achieved for both slow and fast moving clutters with lower complexity compared to SVD-based filters. For instance, use of the proposed method results in the bias being less than 8% and standard deviation being less than 12% for fast moving clutter when the beam-to-flow-angle is $90^o$.
Third, a flow rate estimation method based on kernel power weighting is proposed. As the velocity estimator is a kernel-based method, the estimation accuracy degrades near the vessel boundary. In order to account for kernels that are not fully inside the vessel, fractional weights are given to these kernels based on their signal power. The proposed method achieves excellent flow rate estimation results with less than 8% bias for both slow and fast moving clutters.
The performance of the velocity estimator is also evaluated for challenging models. A 2D version of our two-tiered method is able to accurately estimate velocity vectors in a spinning disk as well as in a carotid bifurcation model, both of which are part of the synthetic aperture vector flow imaging (SA-VFI) challenge of 2018. In fact, the proposed method ranked 3rd in the challenge for testing dataset with carotid bifurcation. The flow estimation method is also evaluated for blood flow in vessels with stenosis. Simulation results show that the proposed method is able to estimate the flow rate with less than 9% bias.
ContributorsWei, Siyuan (Author) / Chakrabarti, Chaitali (Thesis advisor) / Papandreou-Suppappola, Antonia (Committee member) / Ogras, Umit Y. (Committee member) / Wenisch, Thomas F. (Committee member) / Arizona State University (Publisher)
Created2018
Description
With the end of Dennard scaling and Moore's law, architects have moved towards
heterogeneous designs consisting of specialized cores to achieve higher performance
and energy efficiency for a target application domain. Applications of linear algebra
are ubiquitous in the field of scientific computing, machine learning, statistics,
etc. with matrix computations being fundamental to these linear algebra based solutions.
Design of multiple dense (or sparse) matrix computation routines on the
same platform is quite challenging. Added to the complexity is the fact that dense
and sparse matrix computations have large differences in their storage and access
patterns and are difficult to optimize on the same architecture. This thesis addresses
this challenge and introduces a reconfigurable accelerator that supports both dense
and sparse matrix computations efficiently.
The reconfigurable architecture has been optimized to execute the following linear
algebra routines: GEMV (Dense General Matrix Vector Multiplication), GEMM
(Dense General Matrix Matrix Multiplication), TRSM (Triangular Matrix Solver),
LU Decomposition, Matrix Inverse, SpMV (Sparse Matrix Vector Multiplication),
SpMM (Sparse Matrix Matrix Multiplication). It is a multicore architecture where
each core consists of a 2D array of processing elements (PE).
The 2D array of PEs is of size 4x4 and is scheduled to perform 4x4 sized matrix
updates efficiently. A sequence of such updates is used to solve a larger problem inside
a core. A novel partitioned block compressed sparse data structure (PBCSC/PBCSR)
is used to perform sparse kernel updates. Scalable partitioning and mapping schemes
are presented that map input matrices of any given size to the multicore architecture.
Design trade-offs related to the PE array dimension, size of local memory inside a core
and the bandwidth between on-chip memories and the cores have been presented. An
optimal core configuration is developed from this analysis. Synthesis results using a 7nm PDK show that the proposed accelerator can achieve a performance of upto
32 GOPS using a single core.
heterogeneous designs consisting of specialized cores to achieve higher performance
and energy efficiency for a target application domain. Applications of linear algebra
are ubiquitous in the field of scientific computing, machine learning, statistics,
etc. with matrix computations being fundamental to these linear algebra based solutions.
Design of multiple dense (or sparse) matrix computation routines on the
same platform is quite challenging. Added to the complexity is the fact that dense
and sparse matrix computations have large differences in their storage and access
patterns and are difficult to optimize on the same architecture. This thesis addresses
this challenge and introduces a reconfigurable accelerator that supports both dense
and sparse matrix computations efficiently.
The reconfigurable architecture has been optimized to execute the following linear
algebra routines: GEMV (Dense General Matrix Vector Multiplication), GEMM
(Dense General Matrix Matrix Multiplication), TRSM (Triangular Matrix Solver),
LU Decomposition, Matrix Inverse, SpMV (Sparse Matrix Vector Multiplication),
SpMM (Sparse Matrix Matrix Multiplication). It is a multicore architecture where
each core consists of a 2D array of processing elements (PE).
The 2D array of PEs is of size 4x4 and is scheduled to perform 4x4 sized matrix
updates efficiently. A sequence of such updates is used to solve a larger problem inside
a core. A novel partitioned block compressed sparse data structure (PBCSC/PBCSR)
is used to perform sparse kernel updates. Scalable partitioning and mapping schemes
are presented that map input matrices of any given size to the multicore architecture.
Design trade-offs related to the PE array dimension, size of local memory inside a core
and the bandwidth between on-chip memories and the cores have been presented. An
optimal core configuration is developed from this analysis. Synthesis results using a 7nm PDK show that the proposed accelerator can achieve a performance of upto
32 GOPS using a single core.
ContributorsAnimesh, Saurabh (Author) / Chakrabarti, Chaitali (Thesis advisor) / Brunhaver, John (Committee member) / Ren, Fengbo (Committee member) / Arizona State University (Publisher)
Created2018
Description
Deep learning (DL) has proved itself be one of the most important developements till date with far reaching impacts in numerous fields like robotics, computer vision, surveillance, speech processing, machine translation, finance, etc. They are now widely used for countless applications because of their ability to generalize real world data, robustness to noise in previously unseen data and high inference accuracy. With the ability to learn useful features from raw sensor data, deep learning algorithms have out-performed tradinal AI algorithms and pushed the boundaries of what can be achieved with AI. In this work, we demonstrate the power of deep learning by developing a neural network to automatically detect cough instances from audio recorded in un-constrained environments. For this, 24 hours long recordings from 9 dierent patients is collected and carefully labeled by medical personel. A pre-processing algorithm is proposed to convert event based cough dataset to a more informative dataset with start and end of coughs and also introduce data augmentation for regularizing the training procedure. The proposed neural network achieves 92.3% leave-one-out accuracy on data captured in real world.
Deep neural networks are composed of multiple layers that are compute/memory intensive. This makes it difficult to execute these algorithms real-time with low power consumption using existing general purpose computers. In this work, we propose hardware accelerators for a traditional AI algorithm based on random forest trees and two representative deep convolutional neural networks (AlexNet and VGG). With the proposed acceleration techniques, ~ 30x performance improvement was achieved compared to CPU for random forest trees. For deep CNNS, we demonstrate that much higher performance can be achieved with architecture space exploration using any optimization algorithms with system level performance and area models for hardware primitives as inputs and goal of minimizing latency with given resource constraints. With this method, ~30GOPs performance was achieved for Stratix V FPGA boards.
Hardware acceleration of DL algorithms alone is not always the most ecient way and sucient to achieve desired performance. There is a huge headroom available for performance improvement provided the algorithms are designed keeping in mind the hardware limitations and bottlenecks. This work achieves hardware-software co-optimization for Non-Maximal Suppression (NMS) algorithm. Using the proposed algorithmic changes and hardware architecture
With CMOS scaling coming to an end and increasing memory bandwidth bottlenecks, CMOS based system might not scale enough to accommodate requirements of more complicated and deeper neural networks in future. In this work, we explore RRAM crossbars and arrays as compact, high performing and energy efficient alternative to CMOS accelerators for deep learning training and inference. We propose and implement RRAM periphery read and write circuits and achieved ~3000x performance improvement in online dictionary learning compared to CPU.
This work also examines the realistic RRAM devices and their non-idealities. We do an in-depth study of the effects of RRAM non-idealities on inference accuracy when a pretrained model is mapped to RRAM based accelerators. To mitigate this issue, we propose Random Sparse Adaptation (RSA), a novel scheme aimed at tuning the model to take care of the faults of the RRAM array on which it is mapped. Our proposed method can achieve inference accuracy much higher than what traditional Read-Verify-Write (R-V-W) method could achieve. RSA can also recover lost inference accuracy 100x ~ 1000x faster compared to R-V-W. Using 32-bit high precision RSA cells, we achieved ~10% higher accuracy using fautly RRAM arrays compared to what can be achieved by mapping a deep network to an 32 level RRAM array with no variations.
Deep neural networks are composed of multiple layers that are compute/memory intensive. This makes it difficult to execute these algorithms real-time with low power consumption using existing general purpose computers. In this work, we propose hardware accelerators for a traditional AI algorithm based on random forest trees and two representative deep convolutional neural networks (AlexNet and VGG). With the proposed acceleration techniques, ~ 30x performance improvement was achieved compared to CPU for random forest trees. For deep CNNS, we demonstrate that much higher performance can be achieved with architecture space exploration using any optimization algorithms with system level performance and area models for hardware primitives as inputs and goal of minimizing latency with given resource constraints. With this method, ~30GOPs performance was achieved for Stratix V FPGA boards.
Hardware acceleration of DL algorithms alone is not always the most ecient way and sucient to achieve desired performance. There is a huge headroom available for performance improvement provided the algorithms are designed keeping in mind the hardware limitations and bottlenecks. This work achieves hardware-software co-optimization for Non-Maximal Suppression (NMS) algorithm. Using the proposed algorithmic changes and hardware architecture
With CMOS scaling coming to an end and increasing memory bandwidth bottlenecks, CMOS based system might not scale enough to accommodate requirements of more complicated and deeper neural networks in future. In this work, we explore RRAM crossbars and arrays as compact, high performing and energy efficient alternative to CMOS accelerators for deep learning training and inference. We propose and implement RRAM periphery read and write circuits and achieved ~3000x performance improvement in online dictionary learning compared to CPU.
This work also examines the realistic RRAM devices and their non-idealities. We do an in-depth study of the effects of RRAM non-idealities on inference accuracy when a pretrained model is mapped to RRAM based accelerators. To mitigate this issue, we propose Random Sparse Adaptation (RSA), a novel scheme aimed at tuning the model to take care of the faults of the RRAM array on which it is mapped. Our proposed method can achieve inference accuracy much higher than what traditional Read-Verify-Write (R-V-W) method could achieve. RSA can also recover lost inference accuracy 100x ~ 1000x faster compared to R-V-W. Using 32-bit high precision RSA cells, we achieved ~10% higher accuracy using fautly RRAM arrays compared to what can be achieved by mapping a deep network to an 32 level RRAM array with no variations.
ContributorsMohanty, Abinash (Author) / Cao, Yu (Thesis advisor) / Seo, Jae-Sun (Committee member) / Vrudhula, Sarma (Committee member) / Chakrabarti, Chaitali (Committee member) / Arizona State University (Publisher)
Created2018
Description
In recent years, conventional convolutional neural network (CNN) has achieved outstanding performance in image and speech processing applications. Unfortunately, the pooling operation in CNN ignores important spatial information which is an important attribute in many applications. The recently proposed capsule network retains spatial information and improves the capabilities of traditional CNN. It uses capsules to describe features in multiple dimensions and dynamic routing to increase the statistical stability of the network.
In this work, we first use capsule network for overlapping digit recognition problem. We evaluate the performance of the network with respect to recognition accuracy, convergence and training time per epoch. We show that capsule network achieves higher accuracy when training set size is small. When training set size is larger, capsule network and conventional CNN have comparable recognition accuracy. The training time per epoch for capsule network is longer than conventional CNN because of the dynamic routing algorithm. An analysis of the GPU timing shows that adjusting the capsule structure can help decrease the time complexity of the dynamic routing algorithm significantly.
Next, we design a capsule network for speech recognition, specifically, overlapping word recognition. We use both capsule network and conventional CNN to recognize 2 overlapping words in speech files created from 5 word classes. We show that capsule network achieves a considerably higher recognition accuracy (96.92%) compared to conventional CNN (85.19%). Our results show that capsule network recognizes overlapping word by recognizing each individual word in the speech. We also verify the scalability of capsule network by increasing the number of word classes from 5 to 10. Capsule network still shows a high recognition accuracy of 95.42% in case of 10 words while the accuracy of conventional CNN decreases sharply to 73.18%.
In this work, we first use capsule network for overlapping digit recognition problem. We evaluate the performance of the network with respect to recognition accuracy, convergence and training time per epoch. We show that capsule network achieves higher accuracy when training set size is small. When training set size is larger, capsule network and conventional CNN have comparable recognition accuracy. The training time per epoch for capsule network is longer than conventional CNN because of the dynamic routing algorithm. An analysis of the GPU timing shows that adjusting the capsule structure can help decrease the time complexity of the dynamic routing algorithm significantly.
Next, we design a capsule network for speech recognition, specifically, overlapping word recognition. We use both capsule network and conventional CNN to recognize 2 overlapping words in speech files created from 5 word classes. We show that capsule network achieves a considerably higher recognition accuracy (96.92%) compared to conventional CNN (85.19%). Our results show that capsule network recognizes overlapping word by recognizing each individual word in the speech. We also verify the scalability of capsule network by increasing the number of word classes from 5 to 10. Capsule network still shows a high recognition accuracy of 95.42% in case of 10 words while the accuracy of conventional CNN decreases sharply to 73.18%.
ContributorsXiong, Yan (Author) / Chakrabarti, Chaitali (Thesis advisor) / Berisha, Visar (Thesis advisor) / Weng, Yang (Committee member) / Arizona State University (Publisher)
Created2018
Description
Speech recognition and keyword detection are becoming increasingly popular applications for mobile systems. While deep neural network (DNN) implementation of these systems have very good performance,
they have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this thesis, techniques to reduce the memory and computation cost
of keyword detection and speech recognition networks (or DNNs) are presented.
The first technique is based on representing all weights and biases by a small number of bits and mapping all nodal computations into fixed-point ones with minimal degradation in the
accuracy. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Additional reduction in memory is achieved by a technique called weight pruning,
where the weights are classified as sensitive and insensitive and the sensitive weights are represented with higher precision. A combination of these two techniques helps reduce the memory
footprint by 81 - 84% for speech recognition and keyword detection networks respectively.
Further reduction in memory size is achieved by judiciously dropping connections for large blocks of weights. The corresponding technique, termed coarse-grain sparsification, introduces
hardware-aware sparsity during DNN training, which leads to efficient weight memory compression and significant reduction in the number of computations during classification without
loss of accuracy. Keyword detection and speech recognition DNNs trained with 75% of the weights dropped and classified with 5-6 bit weight precision effectively reduced the weight memory
requirement by ~95% compared to a fully-connected network with double precision, while showing similar performance in keyword detection accuracy and word error rate.
they have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this thesis, techniques to reduce the memory and computation cost
of keyword detection and speech recognition networks (or DNNs) are presented.
The first technique is based on representing all weights and biases by a small number of bits and mapping all nodal computations into fixed-point ones with minimal degradation in the
accuracy. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Additional reduction in memory is achieved by a technique called weight pruning,
where the weights are classified as sensitive and insensitive and the sensitive weights are represented with higher precision. A combination of these two techniques helps reduce the memory
footprint by 81 - 84% for speech recognition and keyword detection networks respectively.
Further reduction in memory size is achieved by judiciously dropping connections for large blocks of weights. The corresponding technique, termed coarse-grain sparsification, introduces
hardware-aware sparsity during DNN training, which leads to efficient weight memory compression and significant reduction in the number of computations during classification without
loss of accuracy. Keyword detection and speech recognition DNNs trained with 75% of the weights dropped and classified with 5-6 bit weight precision effectively reduced the weight memory
requirement by ~95% compared to a fully-connected network with double precision, while showing similar performance in keyword detection accuracy and word error rate.
ContributorsArunachalam, Sairam (Author) / Chakrabarti, Chaitali (Thesis advisor) / Seo, Jae-Sun (Thesis advisor) / Cao, Yu (Committee member) / Arizona State University (Publisher)
Created2016