nips nips2008 nips2008-3 knowledge-graph by maker-knowledge-mining

3 nips-2008-A Massively Parallel Digital Learning Processor

Source: pdf

Author: Hans P. Graf, Srihari Cadambi, Venkata Jakkula, Murugan Sankaradass, Eric Cosatto, Srimat Chakradhar, Igor Dourdanovic

Abstract: We present a new, massively parallel architecture for accelerating machine learning algorithms, based on arrays of vector processing elements (VPEs) with variable-resolution arithmetic. Groups of VPEs operate in SIMD (single instruction multiple data) mode, and each group is connected to an independent memory bank. The memory bandwidth thus scales with the number of VPEs, while the main data flows are local, keeping power dissipation low. With 256 VPEs, implemented on two FPGAs (field programmable gate array) chips, we obtain a sustained speed of 19 GMACS (billion multiplyaccumulate per sec.) for SVM training, and 86 GMACS for SVM classification. This performance is more than an order of magnitude higher than that of any FPGA implementation reported so far. The speed on one FPGA is similar to the fastest speeds published on a Graphics Processor for the MNIST problem, despite a clock rate that is an order of magnitude lower. Tests with Convolutional Neural Networks show similar compute performances. This massively parallel architecture is particularly attractive for embedded applications, where low power dissipation is critical. 1 I n trod u cti on Machine learning demands higher and higher compute-performance, but serial processors are not improving that much anymore - at least not as quickly as they used to. Mainstream processor development is moving to multi-core systems, using shared memory technology to hide the parallel nature of the processors. But shared memory technology does not scale to hundreds or thousands of cores. In order to reach such levels of parallelization alternative approaches have to be developed. Massively parallel general-purpose computers had limited success so far, because of difficulties programming these machines, and they remain a niche market, mostly in highperformance computing. Yet processors specialized for certain application domains, such as graphics processors or routing processors 1, have been parallelized to several hundred cores and are successful mass products. They improve performance over general-purpose processors by focusing on a few key algorithmic elements, yet still maintain enough flexibility that they can be programmed for a variety of applications. We explore in this paper if a similar approach can lead to efficient machine learning processors. 1 e.g. Nvidia, Quadro FX 5600 graphics processor; Cisco, CRS-1 routing processor Several processors optimized for machine learning, in particular for neural networks, were developed during the 1980’s and 90’s. Examples are the Synapse-1 architecture [1], or the Connectionist Network Supercomputer, CNS1 [2]. Recently there has been less activity in this field, but some accelerators are sold today for specific applications, such as the Axeon [3] processor for power train control of cars. Beside digital processors a large number of analog circuits were built, emulating neural network structures. Extremely high performance with low power dissipation is achievable, see e.g. [4][5], but these networks have little flexibility. SVM implementations on FPGA have been demonstrated in recent years [6-8], yet reached only low compute-performances. All machine learning processors had only limited success so far, indicating how difficult it is to find a good combination of performance, flexibility, price and ease of use. An important consideration is that many applications of machine learning, such as video analysis, data mining, or personalization of services, show the most promise in embedded systems. Embedded learning requires high compute performance while dissipating little power, a combination that is difficult to achieve, and so far required application specific IC (ASIC). Our aim is to develop architectures that meet the requirements for embedded learning, but are programmable and therefore can be used in a wide range of applications. With the goal of analyzing different architectures we designed a development and testing environment where the parallel computation is mapped onto FPGA’s. Initially this system was intended only for experimentation, but its performance is so high that this platform is useful in its own right as accelerator for high-performance systems. While the experiments shown here emphasize high performance, the architecture has been designed from the start for low power dissipation. The main features for achieving this goal are: low-resolution arithmetic, keeping the main data flow local, low operating frequencies, and a modular design, so that unused parts can be powered down dynamically. All results shown here are from the test platform; migration to lowpower FPGA or chip designs are done in a later stage. 2 Al gori th ms - A ri th meti c - A rch i te ctu re For a substantial improvement over a general purpose processor, the algorithms, the arithmetic units, as well as the architecture have to be optimized simultaneously. This is not just an exercise in hardware design, but algorithms and their software implementations have to be developed concurrently. Most machine learning algorithms have not been developed with parallelization in mind. Therefore, we first need to find good parallel versions, identify their performance bottlenecks, and then extract common computational patterns that can be mapped into accelerator hardware. 2.1 Algorithms Characteristic for machine learning is that large amounts of data need to be processed, often with predictable data access patterns and no dependency between operations over large segments of the computation. This is why data-parallelization can often provide good accelerations on multi-core chips, clusters of machines, or even on loosely coupled networks of machines. Using MapReduce, speedups linear with the number of processors have been reported in [9] for several machine learning algorithms. Up to 16 cores were tested, and simulations indicate good scaling to more processors in some cases. Many algorithms, such as KNN, K-means clustering, LVQ, and Neural Networks can be reduced to forms where the computation is dominated by vector-matrix multiplications, which are easily parallelizable. For Convolutional Neural Networks (CNN) the data flow can be complex, yet the core of the computation is a convolution, an operation which has been studied extensively for parallel implementations. For Support Vector Machines (SVM), several parallel algorithms were described, but most saturate quickly for more than 16 processors. Scaling to larger numbers of processors has been demonstrated, applying MapReduce on a graphics processor with 128 cores [10]. Another implementation on a cluster of 48 dual-core machines (with 384 MMX units) [11] scales even super-linearly, and, according to simulations, scales to thousands of cores. Based on this analysis it is clear that vector-matrix and matrix-matrix multiplications for large vector dimensionalities and large numbers of vectors must be handled efficiently. Yet this alone is not sufficient since data access patterns vary greatly between algorithms. We analyze this here in more detail for SVM and CNN. These algorithms were chosen, because they are widely used for industrial applications and cover a broad range of computation, I/O, and memory requirements. The characteristics of the SVM training are summarized in Table 1. We use an approach similar to the one described in [11] to split different parts of the computation between a host CPU and the FPGA accelerator. For large dimensions d of the vectors the calculation of the columns of the kernel matrix dominates by far. This is needed to update the gradients, and in the present implementation, only this part is mapped onto the FPGA. If the dimensionality d is smaller than around 100, operations 2 and 5 can become bottlenecks and should also be mapped onto the accelerator. Challenging is that for each kernel computation a new data vector has to be loaded 4 into the processor, leading to very high I/O requirements. We consider here dimensions of 10 - 10 5 7 and numbers of training data of 10 - 10 , resulting easily in Gigabytes that need to be transferred to the processors at each iteration. 1 2 3 4 5 6 Operation Initialize all αx, Gx Do Find working set αi, αj Update αi, αj Get 2 columns of kernel matrix Update gradients Gx While not converged Computation 2n IO 2n Unit CPU I I I I I * 2n I*2 I * (2d+2dn) I*n CPU CPU FPGA CPU * * * * 2n 10 2nd n Table 1: Compute- and IO-requirements of each step for SVM training (SMO algorithm). n: number of training data; d: dimension of the vectors; G: gradients; α: support vector factors; I: number of iterations. The last column indicates whether the execution happens on the host CPU or the accelerator FPGA. It is assumed that the kernel computation requires a dot product between vectors (e.g. rbf, polynomial, tanh kernels). Neural network algorithms are essentially sequences of vector-matrix multiplications, but networks with special connectivity patterns, such as convolutional networks have very different IO characteristics than fully connected networks. Table 2 shows the computation and IO requirements for scanning several convolution kernels over one input plane. A full network requires multiple of these operations for one layer, with nonlinearities between layers. We map all operations onto the FPGA accelerator, since intermediate results are re-used right away. The most significant 2 difference to between the SVM and CNN is the Compute/IO ratio: SVM: ~ 1; CNN: ~ L*k > 100. Therefore the requirements for these two algorithms are very different, and handling both cases efficiently is quite a challenge for an architecture design. Operation Load L kernels For all input pixels Shift in new pixel Multiply kernels Shift out result 1 2 3 4 Computation IO 2 L* k n* m 2 n*m*L*k n*m Unit FPGA FPGA FPGA FPGA FPGA Table 2: Compute- and IO-requirements for CNN computation (forward pass), where l kernels of size k*k are scanned simultaneously over an input plane of size n*m. This is representative for implementations with kernel unrolling (kernel pixels processed in parallel). Internal shifts, computation of the non-linearity, and border effects not shown. 2.2 Arithmetic Hardware can be built much more compactly and runs with lower power dissipation, if it uses fixed-point instead of floating-point operations. Fortunately, many learning algorithms tolerate a low resolution in most of the computations. This has been investigated extensively for neural networks [12][13], but less so for other learning algorithms. Learning from data is inherently a noisy process, because we see only a sparse sampling of the true probability distributions. A different type of noise is introduced in gradient descent algorithms, when only a few training data are used at a time to move the optimization forward iteratively. This noise is particularly pronounced for stochastic gradient descent. There is no point in representing noisy variables with high resolution, and it is therefore a property inherent to many algorithms that low-resolution computation can be used. It is important, not to confuse this tolerance to low resolution with the resolution required to avoid numeric instabilities. Some of the computations have to be performed with a high resolution, in particular for variables that are updated incrementally. They maintain the state of the optimization and may change in very small steps. But usually by far the largest part of the computation can be executed at a low resolution. Key is that the hardware is flexible enough and can take advantage of reduced resolution while handling high resolution where necessary. Problem Adult Forest MNIST NORB Kernel: Float Obj. f. # SV 31,930.77 11,486 653,170.7 49,333 4,960.13 6,172 1,243.71 3,077 F-score 77.58 98.29 99.12 93.34 Kernel: 16 bit fixed point Obj. f. # SV F-score 31,930.1 11,490 77.63 652,758 49,299 98.28 4,959.64 6,166 99.11 1,244.76 3,154 93.26 F-sc. (4b in) NA NA 99.11 92.78 Table 3: Comparison of the results of SVM training when the kernels are represented with floating point numbers (32 or 64 bits) (left half) and with 16 bit fixed point (right half). The last column shows the results when the resolution of the training data is reduced from 8 bit to 4 bit. For NORB this reduces the accuracy; all other differences in accuracy are not significant. All are two class problems: Adult: n=32,562, d=122; Forest: n=522,000, d=54 (2 against the rest); MNIST: n=60,000, d=784 (odd–even); NORB: n=48,560, d=5,184. We developed a simulator that allows running the training algorithms with various resolutions in each of the variables. A few examples for SVM training are shown in Table 3. Reducing the resolution of the kernel values from double or float to 16 bit fixed point representations does not affect the accuracy for any of the problems. Therefore all the multiplications in the dot products for the kernel computation can be done in low resolutions (4–16 bit in the factors), but the accumulator needs sufficient resolution to avoid over/under flow (48 bit). Once the calculation of the kernel value is completed, it can be reduced to 16 bit. A low resolution of 16 bit is also tolerable for the α values, but a high resolution is required for the gradients (double). For Neural Networks, including CNN, several studies have confirmed that states and gradients can be kept at low resolutions (<16 bit), but the weights must be maintained at a high resolution (float) (see e.g. [12]). In our own evaluations 24 bits in the weights tend to be sufficient. Once the network is trained, for the classification low resolutions can be used for the weights as well (<16 bit). 2.3 A rc h i t e c t u re Figure 1: Left: Schematic of the architecture with the main data flows; on one FPGA 128 VPE are configured into four SIMD groups; L-S: Load-store units. Right: Picture of an FPGA board; in our experiments one or two of them are used, connected via PCI bus to a host CPU. Based on the analysis above, it is clear that the architecture must be optimized for processing massive amounts of data with relatively low precision. Most of the time, data access patterns are predictable and data are processed in blocks that can be stored contiguously. This type of computation is well suited for vector processing, and simple vector processing elements (VPE) with fixed-point arithmetic can handle the operations. Since typically large blocks of data are processed with the same operation, groups of VPE can work in SIMD (single instruction multiple data) mode. Algorithms must then be segmented to map the highvolume, low precision parts onto the vector accelerators and parts requiring high precision arithmetic onto the CPU. The most important design decision is the organization of the memory. Most memory accesses are done in large blocks, so that the data can be streamed, making complex caching unnecessary. This is fortunate, since the amounts of data to be loaded onto the processor are so large that conventional caching strategies would be overwhelmed anyway. Because the blocks tend to be large, a high data bandwidth is crucial, but latency for starting a block transfer is less critical. Therefore we can use regular DDR memories and still get high IO rates. This led to the design shown schematically in Figure 1, where independent memory banks are connected via separate IO ports for each group of 32 VPE. By connecting multiple of the units shown in Figure 1 to a CPU, this architecture scales to larger numbers of VPE. Parallel data IO and parallel memory access scale simultaneously with the number of parallel cores, and we therefore refer to this as the P3 (P-cube) architecture. Notice also that the main data flow is only local between a group of VPE and its own memory block. Avoiding movements of data over long distances is crucial for low power dissipation. How far this architecture can reasonably scale with one CPU depends on the algorithms, the amount of data and the vector dimensionality (see below). A few hundred VPE per CPU have provided good accelerations in all our tests, and much higher numbers are possible with multi-core CPUs and faster CPU-FPGA connections. 3 I mp l e men tati on of th e P 3 A rch i t ectu re This architecture fits surprisingly well onto some of the recent FPGA chips that are available with several hundred Digital Signal Processors (DSP) units and over 1,000 IO pins for data transfers. The boards used here contain each one Xilinx Virtex 5 LX330T-2 FPGA coupled to 4 independent DDR2 SDRAM with a total of 1GB, and 2 independent 4MB SSRAM memory banks (commercial board from AlphaData). One FPGA chip contains 192 DSP with a maximum speed of 550MHz, which corresponds to a theoretical compute-performance of 105.6 GMACS (18 bit and 25 bit operands). There is a total of 14 Mbit of on-chip memory, and the chip incorporates 960 pins for data IO. Due to routing overhead, not all DSP units can be used and the actual clock frequencies tend to be considerably lower than what is advertised for such chips (typically 230MHz or less for our designs). Nevertheless, we obtain high performances because we can use a large number of DSP units for executing the main computation. The main architecture features are: • Parallel processing (on one chip): 128 VPE (hardware DSP) are divided into 4 blocks of 32, each group controlled by one sequencer with a vector instruction set. • Custom Precision: Data are represented with 1 to 16 bit resolution. Higher resolutions are possible by operating multiple DSP as one processor. • Overlapping Computation and Communication: CPU-FPGA communication is overlapped with the FPGA computation. • Overlap Memory Operations with Computation: All loads and stores from the FPGA to off-chip memory are performed concurrently with computations. • High Off-chip Memory Bandwidth: 6 independent data ports, each 32 bits wide, access banked memories concurrently (12GB/s per chip). • • Streaming Data Flow, Simple Access Patterns: Load/store units are tailored for streaming input and output data, and for simple, bursty access patterns. Caching is done under application control with dual-port memory on chip. Load/store with (de)compression: For an increase of effective IO bandwidth the load/store units provide compression and decompression in hardware. Figure 2 shows the configuration of the VPEs for vector dot product computation used for SVM training and classification. For training, the main computation is the calculation of one column of the kernel matrix. One vector is pre-fetched and stored in on-chip memory. All other vectors are streamed in from off-chip memory banks 1-4. Since this is a regular and predictable access pattern, we can utilize burst-mode, achieving a throughput of close to one memory word per cycle. But the speed is nevertheless IO bound. When several vectors can be stored on-chip, as is the case for classification, then the speed becomes compute-bound. Figure 2: Architecture for vector dot-product computation. The left side shows a high-level schematic with the main data flow. The data are streamed from memory banks 1-4 to the VPE arrays, while memory banks 5 and 6, alternatively receive results or stream them back to the host. The right side shows how a group of VPE is pipelined to improve clock speed. The operation for SVM training on the FPGA corresponds to a vector-matrix multiplication and the one for classification to a matrix-matrix multiplication. Therefore the configuration of Figure 2 is useful for many other algorithms as well, where operations with large vectors and matrices are needed, such as Neural Networks. We implemented a specialized configuration for Convolutional Neural Networks, for more efficiency and lower power dissipation. The VPE are daisy-chained and operate as systolic array. In this way we can take advantage of the high computation to IO ratio (Table 2) to reduce the data transfers from memory. 4 E val u ati on s We evaluated SVM training and classification with the NORB and MNIST problems, the latter with up to 2 million training samples (data from [11]). Both are benchmarks with vectors of high dimensionality, representative for applications in image and video analysis. The computation is split between CPU and FPGA as indicated by Table 1. The DDR2 memory banks are clocked at 230MHz, providing double that rate for data transfers. The data may be compressed to save IO bandwidth. On the FPGA they are decompressed first and distributed to the VPE. In our case, a 32 bit word contains eight 4-bit vector components. Four 32 bit words are needed to feed all 32 VPEs of a group; therefore clocking the VPE faster than 115MHz does not improve performance. A VPE executes a multiplication plus add operation in one clock cycle, resulting in a theoretical maximum of 14.7 GMACS per chip. The sustained compute-rate is lower, about 9.4 GMACS, due to overhead (see Table 4). The computation on the host CPU overlaps with that on the FPGA, and has no effect on the speed in the experiments shown here. For the classification the VPE can be clocked higher, at 230 MHz. By using 4-bit operands we can execute 2 multiply-accumulates simultaneously on one DSP, resulting in speed that is more than four times higher and a sustained 43.0 GMACS limited by the number and speed of the VPE. Adding a second FPGA card doubles the speed, showing little saturation effects yet, but for more FPGA per CPU there will be saturation (see Fig. 3). The compute speed in GMACS obtained for NORB is almost identical. # 60k 2M Iterations 8,000 266,900 CPU time 754s -- speed 0.5 -- CPU+MMX time speed 240 s 1.57 531,534 s 1.58 CPU+FPGA time speed 40 s 9.42 88,589 s 9.48 CPU+2 FPGA time speed 21 s 17.9 48,723 s 17.2 Table 4: Training times and average compute speed for SVM training. Systems tested: CPU, Opteron, 2.2GHz; CPU using MMX; CPU with one FPGA; CPU with two FPGA boards. Results are shown for training sizes of 60k and 2M samples. Compute speed is in GMACS (just kernel computations). Training algorithm: SMO with second order working set selection. Parallelizations of SVM training have been reported recently for a GPU [10] and for a cluster [11], both using the MNIST data. In [10] different bounds for stopping were used than here and in [11]. Nevertheless, a comparison of the compute performance is possible, because based on the number of iterations we can compute the average GMACS for the kernel computations. As can be seen in Table 5 a single FPGA is similar in speed to a GPU with 128 stream processors, despite a clock rate that is about 5.5 times lower for I/O and 11 times lower for the VPE. The cluster with 384 MMX units is about 6 times faster than one FPGA with 128 VPE, but dissipates about two orders of magnitude more electric power. For the FPGA this calculation includes only the computation of the kernel values while the part on the CPU is neglected. This is justified for this study, because the rest of the calculations can be mapped on the FPGA as well and will increase the power dissipation only minimally. Number Clock Operand Power Average of cores speed type dissipation compute speed CPU (Opteron) 1 2.2 GHz float 40 W 0.5 GMACS GPU (from [10]) 128 1.35 GHz float 80 W 7.4 GMACS Cluster (from [11]) 384 1.6 GHz byte > 1 kW 54 GMACS FPGA 128 0.12 GHz 4 bit nibble 9W 9.4 GMACS Table 5: Comparison of performances for SVM training (MNIST data). GPU: Nvidia 8800 GTX. Cluster: 48 dual core CPU (Athlon), 384 MMX units. The GPU was training with 60k samples ([10], table 2, second order), the cluster trained with 2 million samples. Processor Figure 3: Acceleration of SVM training as a function of the number of VPE. MNIST n: 2,000,000, d=784; NORB: n=48,560, d=5,184. The points for 128 and 256 VPE are experimental, the higher ones are simulations. Curves MNIST, NORB: Multiple FPGA are attached to one CPU. Curve MNIST C: Each FPGA is attached to a separate host CPU. Scaling of the acceleration with the number of VPEs is shown in Figure 3. The reference speed is that of one FPGA attached to a CPU. The evaluation has been done experimentally for 128 and 256 VPEs, and beyond that with a simulator. The onset of saturation depends on the dimensionality of the vectors, but to a much lesser extent on the number of training vectors (up to the limit of the memory on the FPGA card). MNIST saturates for more than two FPGAs because then the CPU and FPGA computation times become comparable. For the larger vectors of NORB (d=5,184) this saturation starts to be noticeable for more than 4 FPGA. Alternatively, a system can be scaled by grouping multiple CPU, each with one attached FPGA accelerator. Then the scaling follows a linear or even super-linear acceleration (MNIST C) to several thousand VPE. If the CPUs are working in a cluster arrangement, the scaling is similar to the one described in [11]. For convolutional neural networks, the architecture of Figure 2 is modified to allow a block of VPE to operate as systolic array. In this way convolutions can be implemented with minimal data movements. In addition to the convolution, also sub-sampling and non-linear functions plus the logistics to handle multiple layers with arbitrary numbers of kernels in each layer are done on the FPGA. Four separate blocks of such convolvers are packed onto one FPGA, using 100 VPE. Clocked at 115MHz, this architecture provides a maximum of 11.5 GMACS. Including all the overhead the sustained speed is about 10 GMACS. 5 Con cl u s i on s By systematically exploiting characteristic properties of machine learning algorithms, we developed a new massively parallel processor architecture that is very efficient and can be scaled to thousands of processing elements. The implementation demonstrated here is more than an order of magnitude higher in performance than previous FPGA implementations of SVM or CNN. For the MNIST problem it is comparable to the fastest GPU implementations reported so far. These results underline the importance of flexibility over raw compute-speed for massively parallel systems. The flexibility of the FPGA allows more efficient routing and packing of the data and the use of computations with the lowest resolution an algorithm permits. The results of Table 5 indicate the potential of this architecture for low-power operation in embedded applications. R e f e re n c e s [1] Ramacher, et al. (1995) Synapse-1: A high-speed general purpose parallel neurocomputer system. In Proc. 9th Intl. Symposium on Parallel Processing (IPPS'95), pp. 774-781. [2] Asanovic, K., Beck, Feldman, J., Morgan, N. & Wawrzynek, J. (1994) A Supercomputer for Neural Computation, Proc. IEEE Intl. Joint Conference on Neural Networks, pp. 5-9, Orlando, Florida. [3] Neil, P., (2005) Combining hardware with a powerful automotive MCU for powertrain applications. In Industrial Embedded Resource Guide, p. 88. [4] Korekado, et al. (2003) A Convolutional Neural Network VLSI for Image Recognition Using Merged/Mixed Analog-Digital Architecture, in Proc. 7th KES 2003, Oxford, pp 169-176. [5] Murasaki, M., Arima, Y. & Shinohara, H. (1993) A 20 Tera-CPS Analog Neural Network Board. In Proc. Int. Joint Conf. Neural Networks, pp. 3027 – 3030. [6] Pedersen, R., Schoeberl, M. (2006), An Embedded Support Vector Machine, WISE 2006. [7] Dey, S., Kedia, M. Agarwal, N., Basu, A., Embedded Support Vector Machine: Architectural Enhancements and Evaluation, in Proc 20th Int. Conf. VLSI Design. [8] Anguita, D., Boni, A., Ridella, S., (2003) A Digital Architecture for Support Vector Machines: Theory, Algorithm, and FPGA Implementation, IEEE Trans. Neural Networks, 14/5, pp.993-1009. [9] Chu, C., Kim, S., Lin, Y., Yu, Y., Bradski, G., Ng, A. & Olukotun, K. (2007) Map-Reduce for Machine Learning on Multicore, Advances in Neural Information Processing Systems 19, MIT Press. [10] Catanzaro, B., Sundaram, N., & Keutzer, K. (2008) Fast Support Vector Machine Training and Classification on Graphics Processors, Proc. 25th Int. Conf. Machine Learning, pp 104-111. [11] Durdanovic, I., Cosatto, E. & Graf, H. (2007) Large Scale Parallel SVM Implementation. In L. Bottou, O. Chapelle, D. DeCoste, J. Weston (eds.), Large Scale Kernel Machines, pp. 105-138, MIT Press. [12] Simard, P & Graf, H. (1994) Backpropagation without Multiplication. In J. Cowan, G. Tesauro, J. Alspector, (eds.), Neural Information Processing Systems 6, pp. 232 – 239, Morgan Kaufmann. [13] Savich, A., Moussa, M., Areibi, S., (2007) The Impact of Arithmetic Representation on Implementing MLP-BP on FPGAs: A Study, IEEE Trans. Neural Networks, 18/1, pp. 240-252.

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 com NEC Laboratories, America 4 Independence Way, Suite 200; Princeton, NJ 07738, USA Abstract We present a new, massively parallel architecture for accelerating machine learning algorithms, based on arrays of vector processing elements (VPEs) with variable-resolution arithmetic. [sent-8, score-0.323]

2 Groups of VPEs operate in SIMD (single instruction multiple data) mode, and each group is connected to an independent memory bank. [sent-9, score-0.133]

3 The memory bandwidth thus scales with the number of VPEs, while the main data flows are local, keeping power dissipation low. [sent-10, score-0.286]

4 With 256 VPEs, implemented on two FPGAs (field programmable gate array) chips, we obtain a sustained speed of 19 GMACS (billion multiplyaccumulate per sec. [sent-11, score-0.159]

5 The speed on one FPGA is similar to the fastest speeds published on a Graphics Processor for the MNIST problem, despite a clock rate that is an order of magnitude lower. [sent-14, score-0.15]

6 This massively parallel architecture is particularly attractive for embedded applications, where low power dissipation is critical. [sent-16, score-0.536]

7 1 I n trod u cti on Machine learning demands higher and higher compute-performance, but serial processors are not improving that much anymore - at least not as quickly as they used to. [sent-17, score-0.167]

8 Mainstream processor development is moving to multi-core systems, using shared memory technology to hide the parallel nature of the processors. [sent-18, score-0.323]

9 But shared memory technology does not scale to hundreds or thousands of cores. [sent-19, score-0.097]

10 Massively parallel general-purpose computers had limited success so far, because of difficulties programming these machines, and they remain a niche market, mostly in highperformance computing. [sent-21, score-0.105]

11 Yet processors specialized for certain application domains, such as graphics processors or routing processors 1, have been parallelized to several hundred cores and are successful mass products. [sent-22, score-0.671]

12 They improve performance over general-purpose processors by focusing on a few key algorithmic elements, yet still maintain enough flexibility that they can be programmed for a variety of applications. [sent-23, score-0.219]

13 Nvidia, Quadro FX 5600 graphics processor; Cisco, CRS-1 routing processor Several processors optimized for machine learning, in particular for neural networks, were developed during the 1980’s and 90’s. [sent-27, score-0.384]

14 Examples are the Synapse-1 architecture [1], or the Connectionist Network Supercomputer, CNS1 [2]. [sent-28, score-0.129]

15 Recently there has been less activity in this field, but some accelerators are sold today for specific applications, such as the Axeon [3] processor for power train control of cars. [sent-29, score-0.19]

16 Beside digital processors a large number of analog circuits were built, emulating neural network structures. [sent-30, score-0.202]

17 Extremely high performance with low power dissipation is achievable, see e. [sent-31, score-0.162]

18 All machine learning processors had only limited success so far, indicating how difficult it is to find a good combination of performance, flexibility, price and ease of use. [sent-35, score-0.167]

19 With the goal of analyzing different architectures we designed a development and testing environment where the parallel computation is mapped onto FPGA’s. [sent-39, score-0.222]

20 While the experiments shown here emphasize high performance, the architecture has been designed from the start for low power dissipation. [sent-41, score-0.202]

21 The main features for achieving this goal are: low-resolution arithmetic, keeping the main data flow local, low operating frequencies, and a modular design, so that unused parts can be powered down dynamically. [sent-42, score-0.094]

22 2 Al gori th ms - A ri th meti c - A rch i te ctu re For a substantial improvement over a general purpose processor, the algorithms, the arithmetic units, as well as the architecture have to be optimized simultaneously. [sent-44, score-0.226]

23 Therefore, we first need to find good parallel versions, identify their performance bottlenecks, and then extract common computational patterns that can be mapped into accelerator hardware. [sent-47, score-0.214]

24 1 Algorithms Characteristic for machine learning is that large amounts of data need to be processed, often with predictable data access patterns and no dependency between operations over large segments of the computation. [sent-49, score-0.132]

25 Using MapReduce, speedups linear with the number of processors have been reported in [9] for several machine learning algorithms. [sent-51, score-0.167]

26 Up to 16 cores were tested, and simulations indicate good scaling to more processors in some cases. [sent-52, score-0.241]

27 For Convolutional Neural Networks (CNN) the data flow can be complex, yet the core of the computation is a convolution, an operation which has been studied extensively for parallel implementations. [sent-54, score-0.246]

28 For Support Vector Machines (SVM), several parallel algorithms were described, but most saturate quickly for more than 16 processors. [sent-55, score-0.105]

29 Scaling to larger numbers of processors has been demonstrated, applying MapReduce on a graphics processor with 128 cores [10]. [sent-56, score-0.41]

30 These algorithms were chosen, because they are widely used for industrial applications and cover a broad range of computation, I/O, and memory requirements. [sent-61, score-0.097]

31 We use an approach similar to the one described in [11] to split different parts of the computation between a host CPU and the FPGA accelerator. [sent-63, score-0.101]

32 If the dimensionality d is smaller than around 100, operations 2 and 5 can become bottlenecks and should also be mapped onto the accelerator. [sent-66, score-0.126]

33 Challenging is that for each kernel computation a new data vector has to be loaded 4 into the processor, leading to very high I/O requirements. [sent-67, score-0.11]

34 We consider here dimensions of 10 - 10 5 7 and numbers of training data of 10 - 10 , resulting easily in Gigabytes that need to be transferred to the processors at each iteration. [sent-68, score-0.199]

35 The last column indicates whether the execution happens on the host CPU or the accelerator FPGA. [sent-71, score-0.108]

36 It is assumed that the kernel computation requires a dot product between vectors (e. [sent-72, score-0.111]

37 Neural network algorithms are essentially sequences of vector-matrix multiplications, but networks with special connectivity patterns, such as convolutional networks have very different IO characteristics than fully connected networks. [sent-75, score-0.147]

38 Therefore the requirements for these two algorithms are very different, and handling both cases efficiently is quite a challenge for an architecture design. [sent-80, score-0.129]

39 This is representative for implementations with kernel unrolling (kernel pixels processed in parallel). [sent-82, score-0.107]

40 Fortunately, many learning algorithms tolerate a low resolution in most of the computations. [sent-86, score-0.118]

41 It is important, not to confuse this tolerance to low resolution with the resolution required to avoid numeric instabilities. [sent-92, score-0.202]

42 Key is that the hardware is flexible enough and can take advantage of reduced resolution while handling high resolution where necessary. [sent-96, score-0.209]

43 78 Table 3: Comparison of the results of SVM training when the kernels are represented with floating point numbers (32 or 64 bits) (left half) and with 16 bit fixed point (right half). [sent-119, score-0.169]

44 The last column shows the results when the resolution of the training data is reduced from 8 bit to 4 bit. [sent-120, score-0.218]

45 Reducing the resolution of the kernel values from double or float to 16 bit fixed point representations does not affect the accuracy for any of the problems. [sent-125, score-0.299]

46 Therefore all the multiplications in the dot products for the kernel computation can be done in low resolutions (4–16 bit in the factors), but the accumulator needs sufficient resolution to avoid over/under flow (48 bit). [sent-126, score-0.472]

47 A low resolution of 16 bit is also tolerable for the α values, but a high resolution is required for the gradients (double). [sent-128, score-0.337]

48 For Neural Networks, including CNN, several studies have confirmed that states and gradients can be kept at low resolutions (<16 bit), but the weights must be maintained at a high resolution (float) (see e. [sent-129, score-0.211]

49 Once the network is trained, for the classification low resolutions can be used for the weights as well (<16 bit). [sent-133, score-0.183]

50 3 A rc h i t e c t u re Figure 1: Left: Schematic of the architecture with the main data flows; on one FPGA 128 VPE are configured into four SIMD groups; L-S: Load-store units. [sent-135, score-0.129]

51 Based on the analysis above, it is clear that the architecture must be optimized for processing massive amounts of data with relatively low precision. [sent-137, score-0.163]

52 Most of the time, data access patterns are predictable and data are processed in blocks that can be stored contiguously. [sent-138, score-0.171]

53 This type of computation is well suited for vector processing, and simple vector processing elements (VPE) with fixed-point arithmetic can handle the operations. [sent-139, score-0.112]

54 Since typically large blocks of data are processed with the same operation, groups of VPE can work in SIMD (single instruction multiple data) mode. [sent-140, score-0.103]

55 Algorithms must then be segmented to map the highvolume, low precision parts onto the vector accelerators and parts requiring high precision arithmetic onto the CPU. [sent-141, score-0.215]

56 Most memory accesses are done in large blocks, so that the data can be streamed, making complex caching unnecessary. [sent-143, score-0.13]

57 This is fortunate, since the amounts of data to be loaded onto the processor are so large that conventional caching strategies would be overwhelmed anyway. [sent-144, score-0.222]

58 This led to the design shown schematically in Figure 1, where independent memory banks are connected via separate IO ports for each group of 32 VPE. [sent-147, score-0.187]

59 By connecting multiple of the units shown in Figure 1 to a CPU, this architecture scales to larger numbers of VPE. [sent-148, score-0.174]

60 Parallel data IO and parallel memory access scale simultaneously with the number of parallel cores, and we therefore refer to this as the P3 (P-cube) architecture. [sent-149, score-0.351]

61 Notice also that the main data flow is only local between a group of VPE and its own memory block. [sent-150, score-0.157]

62 How far this architecture can reasonably scale with one CPU depends on the algorithms, the amount of data and the vector dimensionality (see below). [sent-152, score-0.129]

63 3 I mp l e men tati on of th e P 3 A rch i t ectu re This architecture fits surprisingly well onto some of the recent FPGA chips that are available with several hundred Digital Signal Processors (DSP) units and over 1,000 IO pins for data transfers. [sent-154, score-0.298]

64 The boards used here contain each one Xilinx Virtex 5 LX330T-2 FPGA coupled to 4 independent DDR2 SDRAM with a total of 1GB, and 2 independent 4MB SSRAM memory banks (commercial board from AlphaData). [sent-155, score-0.157]

65 One FPGA chip contains 192 DSP with a maximum speed of 550MHz, which corresponds to a theoretical compute-performance of 105. [sent-156, score-0.14]

66 Due to routing overhead, not all DSP units can be used and the actual clock frequencies tend to be considerably lower than what is advertised for such chips (typically 230MHz or less for our designs). [sent-159, score-0.208]

67 The main architecture features are: • Parallel processing (on one chip): 128 VPE (hardware DSP) are divided into 4 blocks of 32, each group controlled by one sequencer with a vector instruction set. [sent-161, score-0.199]

68 • Custom Precision: Data are represented with 1 to 16 bit resolution. [sent-162, score-0.102]

69 • Overlap Memory Operations with Computation: All loads and stores from the FPGA to off-chip memory are performed concurrently with computations. [sent-165, score-0.097]

70 Caching is done under application control with dual-port memory on chip. [sent-168, score-0.097]

71 Figure 2 shows the configuration of the VPEs for vector dot product computation used for SVM training and classification. [sent-170, score-0.122]

72 For training, the main computation is the calculation of one column of the kernel matrix. [sent-171, score-0.111]

73 All other vectors are streamed in from off-chip memory banks 1-4. [sent-173, score-0.229]

74 Since this is a regular and predictable access pattern, we can utilize burst-mode, achieving a throughput of close to one memory word per cycle. [sent-174, score-0.174]

75 When several vectors can be stored on-chip, as is the case for classification, then the speed becomes compute-bound. [sent-176, score-0.114]

76 The data are streamed from memory banks 1-4 to the VPE arrays, while memory banks 5 and 6, alternatively receive results or stream them back to the host. [sent-179, score-0.359]

77 The operation for SVM training on the FPGA corresponds to a vector-matrix multiplication and the one for classification to a matrix-matrix multiplication. [sent-181, score-0.157]

78 Therefore the configuration of Figure 2 is useful for many other algorithms as well, where operations with large vectors and matrices are needed, such as Neural Networks. [sent-182, score-0.1]

79 4 E val u ati on s We evaluated SVM training and classification with the NORB and MNIST problems, the latter with up to 2 million training samples (data from [11]). [sent-186, score-0.153]

80 The DDR2 memory banks are clocked at 230MHz, providing double that rate for data transfers. [sent-189, score-0.196]

81 In our case, a 32 bit word contains eight 4-bit vector components. [sent-192, score-0.102]

82 Four 32 bit words are needed to feed all 32 VPEs of a group; therefore clocking the VPE faster than 115MHz does not improve performance. [sent-193, score-0.102]

83 A VPE executes a multiplication plus add operation in one clock cycle, resulting in a theoretical maximum of 14. [sent-194, score-0.099]

84 The computation on the host CPU overlaps with that on the FPGA, and has no effect on the speed in the experiments shown here. [sent-198, score-0.188]

85 For the classification the VPE can be clocked higher, at 230 MHz. [sent-199, score-0.128]

86 By using 4-bit operands we can execute 2 multiply-accumulates simultaneously on one DSP, resulting in speed that is more than four times higher and a sustained 43. [sent-200, score-0.159]

87 As can be seen in Table 5 a single FPGA is similar in speed to a GPU with 128 stream processors, despite a clock rate that is about 5. [sent-221, score-0.15]

88 For the FPGA this calculation includes only the computation of the kernel values while the part on the CPU is neglected. [sent-224, score-0.111]

89 This is justified for this study, because the rest of the calculations can be mapped on the FPGA as well and will increase the power dissipation only minimally. [sent-225, score-0.158]

90 Number Clock Operand Power Average of cores speed type dissipation compute speed CPU (Opteron) 1 2. [sent-226, score-0.337]

91 The reference speed is that of one FPGA attached to a CPU. [sent-243, score-0.121]

92 The onset of saturation depends on the dimensionality of the vectors, but to a much lesser extent on the number of training vectors (up to the limit of the memory on the FPGA card). [sent-245, score-0.198]

93 For convolutional neural networks, the architecture of Figure 2 is modified to allow a block of VPE to operate as systolic array. [sent-251, score-0.192]

94 Clocked at 115MHz, this architecture provides a maximum of 11. [sent-255, score-0.129]

95 Including all the overhead the sustained speed is about 10 GMACS. [sent-257, score-0.157]

96 5 Con cl u s i on s By systematically exploiting characteristic properties of machine learning algorithms, we developed a new massively parallel processor architecture that is very efficient and can be scaled to thousands of processing elements. [sent-258, score-0.483]

97 These results underline the importance of flexibility over raw compute-speed for massively parallel systems. [sent-261, score-0.246]

98 The flexibility of the FPGA allows more efficient routing and packing of the data and the use of computations with the lowest resolution an algorithm permits. [sent-262, score-0.223]

99 The results of Table 5 indicate the potential of this architecture for low-power operation in embedded applications. [sent-263, score-0.216]

100 (1995) Synapse-1: A high-speed general purpose parallel neurocomputer system. [sent-265, score-0.105]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('fpga', 0.67), ('vpe', 0.223), ('cpu', 0.21), ('gmacs', 0.193), ('io', 0.179), ('processors', 0.167), ('architecture', 0.129), ('processor', 0.121), ('vpes', 0.119), ('parallel', 0.105), ('norb', 0.104), ('bit', 0.102), ('mnist', 0.098), ('memory', 0.097), ('dsp', 0.091), ('classification', 0.089), ('dissipation', 0.089), ('massively', 0.089), ('svm', 0.087), ('speed', 0.087), ('resolution', 0.084), ('gpu', 0.078), ('cores', 0.074), ('float', 0.074), ('mmx', 0.074), ('arithmetic', 0.067), ('cnn', 0.065), ('clock', 0.063), ('convolutional', 0.063), ('banks', 0.06), ('flow', 0.06), ('resolutions', 0.06), ('host', 0.056), ('chip', 0.053), ('accelerator', 0.052), ('chips', 0.052), ('flexibility', 0.052), ('embedded', 0.051), ('graphics', 0.048), ('multiplications', 0.048), ('routing', 0.048), ('computation', 0.045), ('units', 0.045), ('configuration', 0.045), ('cosatto', 0.045), ('fpgas', 0.045), ('simd', 0.045), ('streamed', 0.045), ('access', 0.044), ('networks', 0.042), ('saturation', 0.042), ('sustained', 0.042), ('onto', 0.042), ('hardware', 0.041), ('ghz', 0.04), ('clocked', 0.039), ('efficient', 0.039), ('graf', 0.039), ('kernel', 0.039), ('power', 0.039), ('operation', 0.036), ('instruction', 0.036), ('kernels', 0.035), ('implementations', 0.035), ('digital', 0.035), ('blocks', 0.034), ('attached', 0.034), ('low', 0.034), ('caching', 0.033), ('predictable', 0.033), ('gradients', 0.033), ('processed', 0.033), ('training', 0.032), ('cluster', 0.032), ('bandwidth', 0.031), ('mapped', 0.03), ('accelerators', 0.03), ('cadambi', 0.03), ('cpus', 0.03), ('durdanovic', 0.03), ('flows', 0.03), ('jakkula', 0.03), ('mapreduce', 0.03), ('nvidia', 0.03), ('operands', 0.03), ('opteron', 0.03), ('ports', 0.03), ('programmable', 0.03), ('rch', 0.03), ('supercomputer', 0.03), ('acceleration', 0.028), ('overhead', 0.028), ('operations', 0.028), ('machines', 0.027), ('table', 0.027), ('patterns', 0.027), ('vectors', 0.027), ('calculation', 0.027), ('bottlenecks', 0.026), ('loaded', 0.026)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 1.0000001 3 nips-2008-A Massively Parallel Digital Learning Processor

Author: Hans P. Graf, Srihari Cadambi, Venkata Jakkula, Murugan Sankaradass, Eric Cosatto, Srimat Chakradhar, Igor Dourdanovic

2 0.12106465 28 nips-2008-Asynchronous Distributed Learning of Topic Models

Author: Padhraic Smyth, Max Welling, Arthur U. Asuncion

Abstract: Distributed learning is a problem of fundamental interest in machine learning and cognitive science. In this paper, we present asynchronous distributed learning algorithms for two well-known unsupervised learning frameworks: Latent Dirichlet Allocation (LDA) and Hierarchical Dirichlet Processes (HDP). In the proposed approach, the data are distributed across P processors, and processors independently perform Gibbs sampling on their local data and communicate their information in a local asynchronous manner with other processors. We demonstrate that our asynchronous algorithms are able to learn global topic models that are statistically as accurate as those learned by the standard LDA and HDP samplers, but with signiﬁcant improvements in computation time and memory. We show speedup results on a 730-million-word text corpus using 32 processors, and we provide perplexity results for up to 1500 virtual processors. As a stepping stone in the development of asynchronous HDP, a parallel HDP sampler is also introduced. 1

3 0.086309962 148 nips-2008-Natural Image Denoising with Convolutional Networks

Author: Viren Jain, Sebastian Seung

Abstract: We present an approach to low-level vision that combines two main ideas: the use of convolutional networks as an image processing architecture and an unsupervised learning procedure that synthesizes training samples from speciﬁc noise models. We demonstrate this approach on the challenging problem of natural image denoising. Using a test set with a hundred natural images, we ﬁnd that convolutional networks provide comparable and in some cases superior performance to state of the art wavelet and Markov random ﬁeld (MRF) methods. Moreover, we ﬁnd that a convolutional network offers similar performance in the blind denoising setting as compared to other techniques in the non-blind setting. We also show how convolutional networks are mathematically related to MRF approaches by presenting a mean ﬁeld theory for an MRF specially designed for image denoising. Although these approaches are related, convolutional networks avoid computational difﬁculties in MRF approaches that arise from probabilistic learning and inference. This makes it possible to learn image processing architectures that have a high degree of representational power (we train models with over 15,000 parameters), but whose computational expense is signiﬁcantly less than that associated with inference in MRF approaches with even hundreds of parameters. 1 Background Low-level image processing tasks include edge detection, interpolation, and deconvolution. These tasks are useful both in themselves, and as a front-end for high-level visual tasks like object recognition. This paper focuses on the task of denoising, deﬁned as the recovery of an underlying image from an observation that has been subjected to Gaussian noise. One approach to image denoising is to transform an image from pixel intensities into another representation where statistical regularities are more easily captured. For example, the Gaussian scale mixture (GSM) model introduced by Portilla and colleagues is based on a multiscale wavelet decomposition that provides an effective description of local image statistics [1, 2]. Another approach is to try and capture statistical regularities of pixel intensities directly using Markov random ﬁelds (MRFs) to deﬁne a prior over the image space. Initial work used handdesigned settings of the parameters, but recently there has been increasing success in learning the parameters of such models from databases of natural images [3, 4, 5, 6, 7, 8]. Prior models can be used for tasks such as image denoising by augmenting the prior with a noise model. Alternatively, an MRF can be used to model the probability distribution of the clean image conditioned on the noisy image. This conditional random ﬁeld (CRF) approach is said to be discriminative, in contrast to the generative MRF approach. Several researchers have shown that the CRF approach can outperform generative learning on various image restoration and labeling tasks [9, 10]. CRFs have recently been applied to the problem of image denoising as well [5]. 1 The present work is most closely related to the CRF approach. Indeed, certain special cases of convolutional networks can be seen as performing maximum likelihood inference on a CRF [11]. The advantage of the convolutional network approach is that it avoids a general difﬁculty with applying MRF-based methods to image analysis: the computational expense associated with both parameter estimation and inference in probabilistic models. For example, naive methods of learning MRFbased models involve calculation of the partition function, a normalization factor that is generally intractable for realistic models and image dimensions. As a result, a great deal of research has been devoted to approximate MRF learning and inference techniques that meliorate computational difﬁculties, generally at the cost of either representational power or theoretical guarantees [12, 13]. Convolutional networks largely avoid these difﬁculties by posing the computational task within the statistical framework of regression rather than density estimation. Regression is a more tractable computation and therefore permits models with greater representational power than methods based on density estimation. This claim will be argued for with empirical results on the denoising problem, as well as mathematical connections between MRF and convolutional network approaches. 2 Convolutional Networks Convolutional networks have been extensively applied to visual object recognition using architectures that accept an image as input and, through alternating layers of convolution and subsampling, produce one or more output values that are thresholded to yield binary predictions regarding object identity [14, 15]. In contrast, we study networks that accept an image as input and produce an entire image as output. Previous work has used such architectures to produce images with binary targets in image restoration problems for specialized microscopy data [11, 16]. Here we show that similar architectures can also be used to produce images with the analog ﬂuctuations found in the intensity distributions of natural images. Network Dynamics and Architecture A convolutional network is an alternating sequence of linear ﬁltering and nonlinear transformation operations. The input and output layers include one or more images, while intermediate layers contain “hidden

4 0.082164444 56 nips-2008-Deep Learning with Kernel Regularization for Visual Recognition

Author: Kai Yu, Wei Xu, Yihong Gong

Abstract: In this paper we aim to train deep neural networks for rapid visual recognition. The task is highly challenging, largely due to the lack of a meaningful regularizer on the functions realized by the networks. We propose a novel regularization method that takes advantage of kernel methods, where an oracle kernel function represents prior knowledge about the recognition task of interest. We derive an efﬁcient algorithm using stochastic gradient descent, and demonstrate encouraging results on a wide range of recognition tasks, in terms of both accuracy and speed. 1

5 0.073438734 121 nips-2008-Learning to Use Working Memory in Partially Observable Environments through Dopaminergic Reinforcement

Author: Michael T. Todd, Yael Niv, Jonathan D. Cohen

Abstract: Working memory is a central topic of cognitive neuroscience because it is critical for solving real-world problems in which information from multiple temporally distant sources must be combined to generate appropriate behavior. However, an often neglected fact is that learning to use working memory effectively is itself a difficult problem. The Gating framework [14] is a collection of psychological models that show how dopamine can train the basal ganglia and prefrontal cortex to form useful working memory representations in certain types of problems. We unite Gating with machine learning theory concerning the general problem of memory-based optimal control [5-6]. We present a normative model that learns, by online temporal difference methods, to use working memory to maximize discounted future reward in partially observable settings. The model successfully solves a benchmark working memory problem, and exhibits limitations similar to those observed in humans. Our purpose is to introduce a concise, normative definition of high level cognitive concepts such as working memory and cognitive control in terms of maximizing discounted future rewards. 1 I n t ro d u c t i o n Working memory is loosely defined in cognitive neuroscience as information that is (1) internally maintained on a temporary or short term basis, and (2) required for tasks in which immediate observations cannot be mapped to correct actions. It is widely assumed that prefrontal cortex (PFC) plays a role in maintaining and updating working memory. However, relatively little is known about how PFC develops useful working memory representations for a new task. Furthermore, current work focuses on describing the structure and limitations of working memory, but does not ask why, or in what general class of tasks, is it necessary. Borrowing from the theory of optimal control in partially observable Markov decision problems (POMDPs), we frame the psychological concept of working memory as an internal state representation, developed and employed to maximize future reward in partially observable environments. We combine computational insights from POMDPs and neurobiologically plausible models from cognitive neuroscience to suggest a simple reinforcement learning (RL) model of working memory function that can be implemented through dopaminergic training of the basal ganglia and PFC. The Gating framework is a series of cognitive neuroscience models developed to explain how dopaminergic RL signals can shape useful working memory representations [1-4]. Computationally this framework models working memory as a collection of past observations, each of which can occasionally be replaced with the current observation, and addresses the problem of learning when to update each memory element versus maintaining it. In the original Gating model [1-2] the PFC contained a unitary working memory representation that was updated whenever a phasic dopamine (DA) burst occurred (e.g., due to unexpected reward or novelty). That model was the first to connect working memory and RL via the temporal difference (TD) model of DA firing [7-8], and thus to suggest how working memory might serve a normative purpose. However, that model had limited computational flexibility due to the unitary nature of the working memory (i.e., a singleobservation memory controlled by a scalar DA signal). More recent work [3-4] has partially repositioned the Gating framework within the Actor/Critic model of mesostriatal RL [9-10], positing memory updating as but another cortical action controlled by the dorsal striatal

6 0.056942269 196 nips-2008-Relative Margin Machines

7 0.053299811 29 nips-2008-Automatic online tuning for fast Gaussian summation

8 0.051479246 103 nips-2008-Implicit Mixtures of Restricted Boltzmann Machines

9 0.049328335 160 nips-2008-On Computational Power and the Order-Chaos Phase Transition in Reservoir Computing

10 0.046114802 62 nips-2008-Differentiable Sparse Coding

11 0.043017793 231 nips-2008-Temporal Dynamics of Cognitive Control

12 0.041526172 79 nips-2008-Exploring Large Feature Spaces with Hierarchical Multiple Kernel Learning

13 0.038280744 158 nips-2008-Offline Handwriting Recognition with Multidimensional Recurrent Neural Networks

14 0.031050423 97 nips-2008-Hierarchical Fisher Kernels for Longitudinal Data

15 0.030283924 228 nips-2008-Support Vector Machines with a Reject Option

16 0.029904 219 nips-2008-Spectral Hashing

17 0.028152585 248 nips-2008-Using matrices to model symbolic relationship

18 0.028007852 209 nips-2008-Short-Term Depression in VLSI Stochastic Synapse

19 0.027600601 30 nips-2008-Bayesian Experimental Design of Magnetic Resonance Imaging Sequences

20 0.027367527 63 nips-2008-Dimensionality Reduction for Data in Multiple Feature Representations

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, -0.103), (1, -0.006), (2, 0.033), (3, 0.006), (4, 0.03), (5, -0.007), (6, -0.011), (7, -0.009), (8, -0.017), (9, 0.003), (10, 0.082), (11, 0.047), (12, -0.062), (13, 0.081), (14, -0.039), (15, -0.045), (16, -0.002), (17, -0.086), (18, -0.003), (19, -0.058), (20, -0.059), (21, 0.006), (22, 0.075), (23, 0.017), (24, 0.012), (25, 0.061), (26, -0.001), (27, 0.024), (28, -0.009), (29, -0.002), (30, -0.078), (31, 0.041), (32, -0.033), (33, -0.021), (34, -0.05), (35, 0.095), (36, 0.097), (37, 0.022), (38, 0.24), (39, 0.043), (40, 0.058), (41, 0.018), (42, 0.132), (43, 0.05), (44, -0.037), (45, -0.027), (46, 0.097), (47, -0.099), (48, 0.028), (49, -0.147)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.93476254 3 nips-2008-A Massively Parallel Digital Learning Processor

Author: Hans P. Graf, Srihari Cadambi, Venkata Jakkula, Murugan Sankaradass, Eric Cosatto, Srimat Chakradhar, Igor Dourdanovic

2 0.54949361 158 nips-2008-Offline Handwriting Recognition with Multidimensional Recurrent Neural Networks

Author: Alex Graves, Jürgen Schmidhuber

Abstract: Ofﬂine handwriting recognition—the automatic transcription of images of handwritten text—is a challenging task that combines computer vision with sequence learning. In most systems the two elements are handled separately, with sophisticated preprocessing techniques used to extract the image features and sequential models such as HMMs used to provide the transcriptions. By combining two recent innovations in neural networks—multidimensional recurrent neural networks and connectionist temporal classiﬁcation—this paper introduces a globally trained ofﬂine handwriting recogniser that takes raw pixel data as input. Unlike competing systems, it does not require any alphabet speciﬁc preprocessing, and can therefore be used unchanged for any language. Evidence of its generality and power is provided by data from a recent international Arabic recognition competition, where it outperformed all entries (91.4% accuracy compared to 87.2% for the competition winner) despite the fact that neither author understands a word of Arabic. 1

3 0.51240283 160 nips-2008-On Computational Power and the Order-Chaos Phase Transition in Reservoir Computing

Author: Benjamin Schrauwen, Lars Buesing, Robert A. Legenstein

Abstract: Randomly connected recurrent neural circuits have proven to be very powerful models for online computations when a trained memoryless readout function is appended. Such Reservoir Computing (RC) systems are commonly used in two ﬂavors: with analog or binary (spiking) neurons in the recurrent circuits. Previous work showed a fundamental difference between these two incarnations of the RC idea. The performance of a RC system built from binary neurons seems to depend strongly on the network connectivity structure. In networks of analog neurons such dependency has not been observed. In this article we investigate this apparent dichotomy in terms of the in-degree of the circuit nodes. Our analyses based amongst others on the Lyapunov exponent reveal that the phase transition between ordered and chaotic network behavior of binary circuits qualitatively differs from the one in analog circuits. This explains the observed decreased computational performance of binary circuits of high node in-degree. Furthermore, a novel mean-ﬁeld predictor for computational performance is introduced and shown to accurately predict the numerically obtained results. 1

4 0.50980699 148 nips-2008-Natural Image Denoising with Convolutional Networks

Author: Viren Jain, Sebastian Seung

5 0.45373356 56 nips-2008-Deep Learning with Kernel Regularization for Visual Recognition

Author: Kai Yu, Wei Xu, Yihong Gong

6 0.4379209 121 nips-2008-Learning to Use Working Memory in Partially Observable Environments through Dopaminergic Reinforcement

7 0.42287338 28 nips-2008-Asynchronous Distributed Learning of Topic Models

8 0.40085819 124 nips-2008-Load and Attentional Bayes

9 0.3944881 222 nips-2008-Stress, noradrenaline, and realistic prediction of mouse behaviour using reinforcement learning

10 0.369589 172 nips-2008-Optimal Response Initiation: Why Recent Experience Matters

11 0.3684141 68 nips-2008-Efficient Direct Density Ratio Estimation for Non-stationarity Adaptation and Outlier Detection

12 0.36825997 188 nips-2008-QUIC-SVD: Fast SVD Using Cosine Trees

13 0.35569745 83 nips-2008-Fast High-dimensional Kernel Summations Using the Monte Carlo Multipole Method

14 0.34305173 231 nips-2008-Temporal Dynamics of Cognitive Control

15 0.3245329 248 nips-2008-Using matrices to model symbolic relationship

16 0.32452977 61 nips-2008-Diffeomorphic Dimensionality Reduction

17 0.32123595 187 nips-2008-Psychiatry: Insights into depression through normative decision-making models

18 0.30567154 29 nips-2008-Automatic online tuning for fast Gaussian summation

19 0.30462256 77 nips-2008-Evaluating probabilities under high-dimensional latent variable models

20 0.29688829 30 nips-2008-Bayesian Experimental Design of Magnetic Resonance Imaging Sequences

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(4, 0.014), (6, 0.031), (7, 0.062), (12, 0.028), (28, 0.096), (57, 0.078), (59, 0.014), (63, 0.512), (77, 0.027), (78, 0.012), (83, 0.031)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.82002878 3 nips-2008-A Massively Parallel Digital Learning Processor

Author: Hans P. Graf, Srihari Cadambi, Venkata Jakkula, Murugan Sankaradass, Eric Cosatto, Srimat Chakradhar, Igor Dourdanovic

2 0.78337485 212 nips-2008-Skill Characterization Based on Betweenness

Author: Ozgur Simsek, Andre S. Barreto

Abstract: We present a characterization of a useful class of skills based on a graphical representation of an agent’s interaction with its environment. Our characterization uses betweenness, a measure of centrality on graphs. It captures and generalizes (at least intuitively) the bottleneck concept, which has inspired many of the existing skill-discovery algorithms. Our characterization may be used directly to form a set of skills suitable for a given task. More importantly, it serves as a useful guide for developing incremental skill-discovery algorithms that do not rely on knowing or representing the interaction graph in its entirety. 1

3 0.77902639 25 nips-2008-An interior-point stochastic approximation method and an L1-regularized delta rule

Author: Peter Carbonetto, Mark Schmidt, Nando D. Freitas

Abstract: The stochastic approximation method is behind the solution to many important, actively-studied problems in machine learning. Despite its farreaching application, there is almost no work on applying stochastic approximation to learning problems with general constraints. The reason for this, we hypothesize, is that no robust, widely-applicable stochastic approximation method exists for handling such problems. We propose that interior-point methods are a natural solution. We establish the stability of a stochastic interior-point approximation method both analytically and empirically, and demonstrate its utility by deriving an on-line learning algorithm that also performs feature selection via L1 regularization. 1

4 0.68727994 78 nips-2008-Exact Convex Confidence-Weighted Learning

Author: Koby Crammer, Mark Dredze, Fernando Pereira

Abstract: Conﬁdence-weighted (CW) learning [6], an online learning method for linear classiﬁers, maintains a Gaussian distributions over weight vectors, with a covariance matrix that represents uncertainty about weights and correlations. Conﬁdence constraints ensure that a weight vector drawn from the hypothesis distribution correctly classiﬁes examples with a speciﬁed probability. Within this framework, we derive a new convex form of the constraint and analyze it in the mistake bound model. Empirical evaluation with both synthetic and text data shows our version of CW learning achieves lower cumulative and out-of-sample errors than commonly used ﬁrst-order and second-order online methods. 1

5 0.37155178 209 nips-2008-Short-Term Depression in VLSI Stochastic Synapse

Author: Peng Xu, Timothy K. Horiuchi, Pamela A. Abshire

Abstract: We report a compact realization of short-term depression (STD) in a VLSI stochastic synapse. The behavior of the circuit is based on a subtractive single release model of STD. Experimental results agree well with simulation and exhibit expected STD behavior: the transmitted spike train has negative autocorrelation and lower power spectral density at low frequencies which can remove redundancy in the input spike train, and the mean transmission probability is inversely proportional to the input spike rate which has been suggested as an automatic gain control mechanism in neural systems. The dynamic stochastic synapse could potentially be a powerful addition to existing deterministic VLSI spiking neural systems. 1

6 0.37011263 62 nips-2008-Differentiable Sparse Coding

7 0.36761072 75 nips-2008-Estimating vector fields using sparse basis field expansions

8 0.36740056 210 nips-2008-Signal-to-Noise Ratio Analysis of Policy Gradient Algorithms

9 0.36362177 37 nips-2008-Biasing Approximate Dynamic Programming with a Lower Discount Factor

10 0.36348274 141 nips-2008-Multi-Agent Filtering with Infinitely Nested Beliefs

11 0.36283436 175 nips-2008-PSDBoost: Matrix-Generation Linear Programming for Positive Semidefinite Matrices Learning

12 0.36247084 134 nips-2008-Mixed Membership Stochastic Blockmodels

13 0.35864425 196 nips-2008-Relative Margin Machines

14 0.35822827 16 nips-2008-Adaptive Template Matching with Shift-Invariant Semi-NMF

15 0.35284922 30 nips-2008-Bayesian Experimental Design of Magnetic Resonance Imaging Sequences

16 0.34787261 133 nips-2008-Mind the Duality Gap: Logarithmic regret algorithms for online optimization

17 0.34770149 208 nips-2008-Shared Segmentation of Natural Scenes Using Dependent Pitman-Yor Processes

18 0.34763998 192 nips-2008-Reducing statistical dependencies in natural signals using radial Gaussianization

19 0.34465951 2 nips-2008-A Convex Upper Bound on the Log-Partition Function for Binary Distributions

20 0.3427577 200 nips-2008-Robust Kernel Principal Component Analysis