Pith. sign in

REVIEW 4 major objections 3 minor 181 references

Efficient Deep Neural Networks

T0 review · 4 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read The dissertation's central claim: deep neural networks can be made efficient enough for edge deployment by attacking model design, data, hardware, and architecture search together, with reported gains from 10x to 421x.

desk verdict A useful compilation of already-published efficient-DNN work with a genuinely handy efficiency-metrics chapter, but the abstract's headline speedups are baseline-dependent and shouldn't be quoted without the comparison point. read the letter →

arxiv 1908.08926 v1 pith:DNC4YBFO submitted 2019-08-20 cs.CV

classification cs.CV
keywords efficientneuralnetworksedgecomputingobjectdetectionLiDARpointclouddomainadaptationarchitecturesearchmodel-hardwareco-designdataannotation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Deep learning's standard recipe — more compute, more data, bigger models — is unavailable on the edge, where phones and vehicles have tight power budgets and LiDAR annotation is exorbitantly expensive. This dissertation argues that efficiency can be attacked on four independent fronts, and that attacking all four makes edge deep learning practical. It reports compact detectors and segmenters that match baseline accuracy at over 10x lower cost; an annotation tool that cuts LiDAR labeling time 6.2x; a training pipeline that lets models trained on simulated LiDAR match real-data training; a hardware-software co-design built around a zero-cost 'shift' operator that reaches 11.6x faster inference than prior accelerators; and a differentiable architecture search that finds competitive networks at 421x lower search cost. The unifying claim is that each of the three success factors can be substituted by engineering: better architectures replace big models, better tools and simulation replace big datasets, and co-designed silicon replaces brute compute.

What carries the argument

Four mechanisms carry the argument, each replacing an expensive step with a cheaper surrogate: the ConvDet layer (a fully convolutional detection head that replaces fully connected layers), the spherical projection of LiDAR point clouds onto dense 2D grids (letting ordinary 2D CNNs process 3D sensor data), the shift operator (a zero-FLOP, zero-parameter memory movement that replaces spatial convolution and leaves only $1\times1$ convolutions in the network), and the differentiable relaxation of architecture search into training a stochastic super net whose discrete layer choices become differentiable weights. Each converts an apparently necessary cost — parameter count, annotation labor, hardware operations, search compute — into something that can be reduced by design rather than accepted as fixed.

What would settle it

An independent reimplementation of SqueezeDet and Faster R-CNN on the same KITTI split, input resolution, and GPU, matched for mean average precision, could measure the true speed and energy margins; if they do not approach the reported 19.7x speedup and 35.2x energy saving at matched accuracy, the model-efficiency claim is weakened. The same re-run logic applies to the annotation tool (6.2x), the shift-based accelerator (11.6x), and the architecture search (421x), each against its own stated baseline.

Watch

Extended reading notes

Core claim

The paper's central claim is that deep neural networks can be made efficient enough for edge deployment by working at four levels rather than one: model design, data, hardware, and design automation. On the model level, SqueezeDet unifies region proposal and classification into a single convolutional detection layer, ConvDet, which replaces the fully connected detection head of prior single-stage detectors with a layer that has orders of magnitude fewer parameters; it matches Faster R-CNN accuracy on the KITTI benchmark while running at tens of frames per second on a desktop GPU with about 35x lower energy per frame. SqueezeSeg does the same for LiDAR point-cloud segmentation by projecting the cloud onto a spherical grid and feeding it to a compact SqueezeNet-style CNN, reaching over 100 frames per second. On the data level, the LATTE tool combines sensor fusion, one-click annotation, and Kalman-filter tracking to cut LiDAR annotation time 6.2x with better label quality, and SqueezeSegV2 adds a Context Aggregation Module, learned intensity rendering, geodesic correlation alignment, and progressive domain calibration so that models trained purely on simulated data approach the accuracy of models trained on real scans. On the hardware level, the shift operator moves data instead of multiplying it, eliminating spatial convolutions entirely and leaving networks composed only of 1x1 convolutions, which lets a purpose-built accelerator deliver 11.6x faster inference. On the design level, DNAS makes architecture search differentiable by training a stochastic super net whose discrete choices are smoothed into differentiable weights, discovering models with state-of-the-art accuracy-efficiency tradeoffs at 421x lower search cost than prior search methods.

Load-bearing premise

The reported gains — 10x model speedup, 6.2x annotation speedup, 11.6x inference speedup, and 421x search-cost reduction — are measured against specific baselines on specific datasets and hardware, and the dissertation assumes those comparisons are representative of what a strong competitor would achieve.

Editorial extensions

If this is right

  • Edge applications like autonomous driving, augmented reality, and IoT can run sophisticated vision models in real time: the compact networks reported here run at roughly 30 to over 100 frames per second on desktop GPUs and are small enough to be quantized for embedded processors.
  • LiDAR-based datasets become much cheaper to build: the annotation tool cuts labeling time 6.2x with higher label quality, while the domain-adaptation pipeline makes simulated data a viable substitute, relaxing the requirement for expensive real-world scans.
  • Hardware for neural networks can be dramatically simpler: because shift-based networks contain only $1\times1$ convolutions, a compute unit dedicated to that single operator sustains the 11.6x inference speedup, suggesting that co-designed silicon can track network design closely.
  • Model design can be automated: DNAS finds accurate, efficient networks at 421x lower search cost than earlier search methods, making per-device, per-task architecture search practical where it was previously prohibitive.
  • The four levels compose: each targets a different bottleneck, so the gains multiply — a compact model, trained on cheap simulated data, searched automatically for a target accelerator, is the dissertation's picture of practical edge AI.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the efficiency numbers generalize beyond their specific baselines, the practical consequence is that model size and FLOPs are not destiny: the same task accuracy can be bought with 10x to 400x less compute, shifting the edge-AI bottleneck from inference hardware to data pipelines and deployment engineering.
  • The shift operator's premise — that spatial mixing can be done by memory movement instead of multiplication — suggests a testable extension to other expensive layers such as attention mechanisms, where a similar 'move then mix' decomposition might cut cost without retraining.
  • A natural extension of DNAS would be to search jointly over architecture, precision, and data strategy for a fixed target device; the dissertation treats the four efficiency axes separately, and a combined formulation could expose interactions among them.
  • The domain-adaptation result's implicit promise is that synthetic training data plus unlabeled target data may eventually replace most real annotation; whether that holds at scale depends on whether the learned intensity rendering and correlation alignment continue to close the gap as objects, weather, and sensor configurations diversify.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. This dissertation, based on the author's prior published work, claims efficiency gains at four levels: model efficiency (SqueezeDet, SqueezeSeg), data efficiency (LATTE, SqueezeSegV2 with synthetic-data adaptation), hardware efficiency (Shift/Synetgy co-design), and design efficiency (DNAS). It reports large quantitative gains such as more than 10x faster detection, 6.2x faster LiDAR annotation, 11.6x faster inference on a co-designed accelerator, and 421x lower NAS cost. Each chapter presents a method, experiments on external benchmarks (KITTI, ImageNet, CIFAR), and comparisons against prior systems.

Significance. If the claims are accepted with their designated baselines, the work provides practically relevant techniques for deploying deep neural networks on resource-constrained platforms, and the Chapter 2 discussion of theoretical versus practical efficiency metrics is a useful contribution in its own right. The dissertation also ships open-source code for SqueezeDet and SqueezeSeg, and the ablations in Chapter 6 support component-level attributions. The main caveat is that the headline multipliers are baseline- and protocol-dependent; they are not intrinsic properties of the models or algorithms alone.

major comments (4)
  1. [Abstract and §3.4, Table 3.6] The central claim that the model-efficiency work is 'more than 10x faster' with 'lower energy' is an artifact of the chosen two-stage baselines. Table 3.6 reports YOLO at 25.8 FPS on the same TITAN X GPU, so SqueezeDet+ is only about 1.2x faster (32.1 vs. 25.8 FPS) and SqueezeDet is about 2.2x faster (57.2 vs. 25.8 FPS), with a roughly 5x energy advantage over YOLO for SqueezeDet rather than 35x. The abstract and Section 3.5 should state explicitly that the 19.7x speedup and 35.2x energy reduction are measured against the two-stage Faster R-CNN baselines of [3], so the numbers are not read as model-intrinsic properties.
  2. [§3.4, Table 3.6 and Fig. 3.7] The energy-efficiency comparison rests on a heuristic: GPU power is sampled with nvidia-smi at 0.1 s intervals and averaged over the middle third of the working period, and energy per frame is computed as average power divided by inference speed. This measures only GPU power rather than full-system energy, does not account for the idle-to-working transition, and no variance or repeated measurements are reported. Since the '35x lower energy' figure is part of the abstract, this measurement should be calibrated or the claim should be softened to an approximate GPU-power estimate.
  3. [§5.4, Table 5.1] The 6.2x annotation speedup is based on nine volunteers annotating 150 frames (1,116 instances), with the baseline and each feature tested on the same frames, but Table 5.1 reports no per-subject variance or statistical intervals for time or operation count. Human-subject timing is noisy, and without error bars the claimed speedup and the relative ordering of sensor fusion, one-click annotation, and tracking are not fully substantiated; standard deviations across subjects or per-condition ranges should be reported.
  4. [Abstract and Chapter 8] The '421x lower computational cost' claim for DNAS is not accompanied in the presented text by a precise definition of the cost metric (e.g., GPU-hours, same search space, same hardware, same evaluation protocol). NAS search-cost ratios are highly sensitive to the baseline search algorithm, implementation, and reward/evaluation choices; the exact comparison protocol and the corresponding table should be stated so the 421x figure is falsifiable.
minor comments (3)
  1. [§1.3, §3.4, §4.4] There are several typos: 'chpater' should be 'chapter' in Section 1.3, 'receptively' should be 'respectively' in Section 3.4, and 'Nividia' should be 'Nvidia' in Section 4.4.
  2. [List of Tables and §3.4] The list of tables in the front matter orders Table 3.6 before Table 3.5, while the body references Table 3.5 first; the ordering should be made consistent.
  3. [§5.1] In Section 5.1, the sentence beginning 'For example, a 2D bounding box can be determined...' would be clearer if the contrast between 2D and 3D annotation operations were separated into its own paragraph.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the dissertation's efficiency claims are empirical comparisons against external baselines and datasets, not fits or definitions that reduce to their inputs.

full rationale

The primary claims of this dissertation are measured performance ratios: SqueezeDet is compared on KITTI against Faster R-CNN variants and YOLO (Tables 3.2 and 3.6), SqueezeSeg is benchmarked on KITTI and on Drive PX2 hardware, LATTE is evaluated by human annotation time against a stripped-down baseline, SqueezeSegV2's domain adaptation is tested on KITTI after training on GTA-LiDAR, and DNAS search cost is compared to previously published NAS costs. None of these comparisons is defined in terms of the quantity being predicted. The dissertation does summarize the author's own prior papers, and it cites several works by the same research group, but the load-bearing evidence in each chapter is an externally measurable result: speed in frames per second, energy in joules per frame, annotation time per instance, IoU on KITTI, and GPU-hours for search. There is no fitted parameter that is renamed as a prediction: the learned intensity rendering and domain calibration in Chapter 6 use unlabeled real data, but the reported accuracy is measured on labeled KITTI data after adaptation, and the adaptation procedure does not use the test labels. The headline multipliers are baseline-dependent, and a skeptic could argue that some baselines are favorable, but baseline selection is a fairness or correctness concern, not a circularity concern. No equation in the manuscript reduces one of the claimed results to its own definition, and no 'uniqueness theorem' or ansatz is imported from a self-citation to force the conclusion. I therefore find no circular step under the required standard of quoting a specific reduction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central claims rest on conventional deep learning assumptions plus domain-specific approximations (spherical projection, bounding-box to point labels, synthetic-to-real transfer). The free parameters are standard hyperparameters chosen by validation; they do not, by themselves, make the claims circular.

free parameters (6)
  • SqueezeDet loss weights = λ_bbox=5, λ+_conf=75, λ-_conf=100
    Chosen empirically on KITTI validation; the central detection loss balances localization, confidence, and classification.
  • SqueezeSeg CRF kernel scales = σ_alpha, σ_beta, σ_gamma (values not given)
    Selected empirically in Chapter 4; the CRF refinement's strength depends on these.
  • Focal loss gamma = 2
    Set in SqueezeSegV2 to down-weight easy background points.
  • Geodesic loss weight = 10
    Weight λ in the domain adaptation loss, chosen in Chapter 6.
  • LATTE clustering thresholds = ε=0.3m, minPts=20 in SqueezeSeg instance experiments
    DBSCAN hyperparameters for one-click annotation and instance segmentation.
  • Intensity rendering regions = n=10
    Number of intensity bins in the learned rendering network.
assumptions (4)
  • domain assumption Point-wise labels can be derived from 3D bounding boxes
    Used to convert KITTI 3D boxes into segmentation targets (Chapters 4, 5, 6). This is an approximation that may mislabel points inside boxes.
  • domain assumption LiDAR point clouds can be represented by a spherical 2D grid without losing key information
    Equation 4.1 projects points to a 64x512 grid; this assumes the discretization retains enough structure for segmentation.
  • domain assumption Simulated GTA-V data plus domain adaptation can substitute for real data
    Chapter 6 relies on this to train on synthetic LiDAR and test on KITTI.
  • standard math Standard deep learning training assumptions (backprop, SGD, batch norm) are valid
    All chapters rely on conventional training.
invented entities (2)
  • Shift operation independent evidence
    purpose: Replaces spatial convolutions with parameter-free memory moves plus 1x1 convolutions
    Evaluated on CIFAR-10/100, ImageNet, face verification, and style transfer; the results provide external evidence.
  • Context Aggregation Module (CAM) independent evidence
    purpose: Reduces sensitivity to LiDAR dropout noise by pooling context
    Tested on KITTI in Chapter 6, improving IoU; this is an empirical component.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Deep Neural Networks." pith.science (2026). https://pith.science/paper/DNC4YBFO

@misc{pith2026190808926,
  author       = {Pith},
  title        = {Pith review of: Efficient Deep Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DNC4YBFO}},
  note         = {Machine review of arXiv:1908.08926}
}
read the original abstract

The success of deep neural networks (DNNs) is attributable to three factors: increased compute capacity, more complex models, and more data. These factors, however, are not always present, especially for edge applications such as autonomous driving, augmented reality, and internet-of-things. Training DNNs requires a large amount of data, which is difficult to obtain. Edge devices such as mobile phones have limited compute capacity, and therefore, require specialized and efficient DNNs. However, due to the enormous design space and prohibitive training costs, designing efficient DNNs for different target devices is challenging. So the question is, with limited data, compute capacity, and model complexity, can we still successfully apply deep neural networks? This dissertation focuses on the above problems and improving the efficiency of deep neural networks at four levels. Model efficiency: we designed neural networks for various computer vision tasks and achieved more than 10x faster speed and lower energy. Data efficiency: we developed an advanced tool that enables 6.2x faster annotation of a LiDAR point cloud. We also leveraged domain adaptation to utilize simulated data, bypassing the need for real data. Hardware efficiency: we co-designed neural networks and hardware accelerators and achieved 11.6x faster inference. Design efficiency: the process of finding the optimal neural networks is time-consuming. Our automated neural architecture search algorithms discovered, using 421x lower computational cost than previous search methods, models with state-of-the-art accuracy and efficiency.

Figures

Figures reproduced from arXiv: 1908.08926 by the authors.

Figure 1.1
Figure 1.1. Three factors for the success of deep neural networks. [PITH_FULL_IMAGE:figures/full_fig_p015_1_1.png] view at source ↗
Figure 1.2
Figure 1.2. For edge-based applications, the three factors for the success of deep learning [PITH_FULL_IMAGE:figures/full_fig_p016_1_2.png] view at source ↗
Figure 1.3
Figure 1.3. An overview of this thesis. analyze the problems in LiDAR point cloud annotation: low resolution, complex annotation operation, and temporal correlation. To solve these problems, we built a new annotation tool that improves annotation efficiency by 6.2x. In Chapter 6, we discuss a more radical strategy to leverage simulated data to train neural networks and adapt the model to the real world. By improving the model s… view at source ↗
Figures from the paper (58 more)
Figure 2.1
Figure 2.1. Figure 2.1: The memory hierarchy of computer architecture. The top-level is CPU and [PITH_FULL_IMAGE:figures/full_fig_p021_2_1.png]
Figure 2.2
Figure 2.2. Figure 2.2: An illustration of a convolutional layer. The computation of this layer is illus [PITH_FULL_IMAGE:figures/full_fig_p023_2_2.png]
Figure 3.1
Figure 3.1. Figure 3.1: SqueezeDet detection pipeline. A convolutional neural network extracts a feature [PITH_FULL_IMAGE:figures/full_fig_p032_3_1.png]
Figure 3.2
Figure 3.2. Figure 3.2: Bounding box transformation. Each grid center has [PITH_FULL_IMAGE:figures/full_fig_p034_3_2.png]
Figure 3.3
Figure 3.3. Figure 3.3: Comparing RPN, ConvDet and the detection layer of YOLO [119]. Activations are represented as blue cubes and layers (and their parameters) are represented as orange ones. Activation and parameter dimensions are also annotated. The second fully connected layer in YOLO …
Figure 3.4
Figure 3.4. Figure 3.4: Example of detection errors. time complexity of NMS with respect to the number of bounding boxes. Thus we only kept the top 64 predictions to feed into NMS. An interesting question to ask is, how does the number of bounding boxes kept affect recall? We tested this wi…
Figure 3.5
Figure 3.5. Figure 3.5: Overall recall vs Nobj for SqueezeDet and SqueezeDet+ models. We also tried to re-scale the input image by 1.5X and 0.75X. The squeezeDet and SqueezeDet+ models achieved the best recall of 0.91 and 0.92 respectively with all bounding boxes. SqueezeDet with 1.5X image…
Figure 3.6
Figure 3.6. Figure 3.6: Model size vs. mean average precision for car detection. Each point on this [PITH_FULL_IMAGE:figures/full_fig_p042_3_6.png]
Figure 3.7
Figure 3.7. Figure 3.7: GPU power measured by nvidia-smi. Here we plot the power measurement curve of 3 models, SqueezeDet, SqueezeDet+, and Faster R-CNN + VGG16 model. We normalize the working period of 3 models to the same range of [0, 1]. We divide the working period evenly into 3 parts …
Figure 4.1
Figure 4.1. Figure 4.1: An example of SqueezeSeg segmentation results. Our predicted result is on the [PITH_FULL_IMAGE:figures/full_fig_p048_4_1.png]
Figure 4.2
Figure 4.2. Figure 4.2: LiDAR Projections. Note that each channel reflects structural information in [PITH_FULL_IMAGE:figures/full_fig_p048_4_2.png]
Figure 4.3
Figure 4.3. Figure 4.3: Network structure of SqueezeSeg. The input to SqueezeSeg is a 64 × 512 × 5 tensor as described in the previous section. We ported layers (conv1a to fire9 ) from SqueezeNet for feature extraction. SqueezeNet used max-pooling to down-sample intermediate feature maps in…
Figure 4.4
Figure 4.4. Figure 4.4: Structure of a FireModule (left) and a fireDeconv (right). Conditional Random Field With image segmentation, label maps predicted by CNN models tend to have blurry bound￾aries. This is due to the loss of low-level details in down-sampling operations such as max- [PI…
Figure 4.5
Figure 4.5. Figure 4.5: Conditional Random Field (CRF) as an RNN layer. [PITH_FULL_IMAGE:figures/full_fig_p054_4_5.png]
Figure 4.6
Figure 4.6. Figure 4.6: Left: Image of a game scene from GTA-V. Right: A LiDAR point cloud corre [PITH_FULL_IMAGE:figures/full_fig_p055_4_6.png]
Figure 4.7
Figure 4.7. Figure 4.7: Fixing distribution of noise in synthesized data [PITH_FULL_IMAGE:figures/full_fig_p056_4_7.png]
Figure 4.8
Figure 4.8. Figure 4.8: Visualization of SqueezeSeg’s prediction on a projected LiDAR depth map. For [PITH_FULL_IMAGE:figures/full_fig_p059_4_8.png]
Figure 5.1
Figure 5.1. Figure 5.1: A screenshot of LATTE. Best viewed in color. [PITH_FULL_IMAGE:figures/full_fig_p063_5_1.png]
Figure 5.2
Figure 5.2. Figure 5.2: Challenges of annotating LiDAR point clouds. (a) LiDAR point clouds have low [PITH_FULL_IMAGE:figures/full_fig_p064_5_2.png]
Figure 5.3
Figure 5.3. Figure 5.3: The sensor-fusion pipeline of LATTE. A LiDAR point cloud is projected onto [PITH_FULL_IMAGE:figures/full_fig_p067_5_3.png]
Figure 5.4
Figure 5.4. Figure 5.4: We use sensor fusion to help annotators confirm the category of a selected object. [PITH_FULL_IMAGE:figures/full_fig_p068_5_4.png]
Figure 5
Figure 5. Figure 5: , ideally, drawing a 3D bounding box requires 1 operation to locate the object, 3 [PITH_FULL_IMAGE:figures/full_fig_p068_5.png]
Figure 5.5
Figure 5.5. Figure 5.5: A comparison of drawing a 3D bounding box, a top-view 2D bounding box, and [PITH_FULL_IMAGE:figures/full_fig_p069_5_5.png]
Figure 5.6
Figure 5.6. Figure 5.6: The one-click annotation pipeline of LATTE. For a given LiDAR point cloud, [PITH_FULL_IMAGE:figures/full_fig_p069_5_6.png]
Figure 5.7
Figure 5.7. Figure 5.7: The tracking pipeline of LATTE. Annotators label a bounding box in the initial [PITH_FULL_IMAGE:figures/full_fig_p073_5_7.png]
Figure 5.8
Figure 5.8. Figure 5.8: Distribution of bounding boxes by class, box area, and orientation in our test [PITH_FULL_IMAGE:figures/full_fig_p074_5_8.png]
Figure 5.9
Figure 5.9. Figure 5.9: Visualization of bounding box annotations from our volunteers, our ground truth, [PITH_FULL_IMAGE:figures/full_fig_p076_5_9.png]
Figure 6.1
Figure 6.1. Figure 6.1: An example of domain shift. The point clouds are projected onto a spherical surface for visualization (cars in red, pedestrians in blue). Our domain adaptation pipeline improves the segmentation from (c) to (d) while trained on synthetic data. and therefore reduces o…
Figure 6.2
Figure 6.2. Figure 6.2: The network structure of the proposed SqueezeSegV2 model for road-object [PITH_FULL_IMAGE:figures/full_fig_p082_6_2.png]
Figure 6.3
Figure 6.3. Figure 6.3: Structures of the Context Aggregation Module. The module starts with a max [PITH_FULL_IMAGE:figures/full_fig_p083_6_3.png]
Figure 6.4
Figure 6.4. Figure 6.4: We feed a random tensor to a convolutional filter, one with CAM before a 3 [PITH_FULL_IMAGE:figures/full_fig_p084_6_4.png]
Figure 6.5
Figure 6.5. Figure 6.5: The framework of the proposed unsupervised domain adaptation method for [PITH_FULL_IMAGE:figures/full_fig_p086_6_5.png]
Figure 6.6
Figure 6.6. Figure 6.6: Rendered v.s. ground truth intensity in the KITTI dataset. [PITH_FULL_IMAGE:figures/full_fig_p087_6_6.png]
Figure 6.7
Figure 6.7. Figure 6.7: Segmentation result comparison between SqueezeSeg [162] and our SqueezeSegV2 [PITH_FULL_IMAGE:figures/full_fig_p089_6_7.png]
Figure 6.8
Figure 6.8. Figure 6.8: Segmentation result comparison before and after domain adaptation (red: car, [PITH_FULL_IMAGE:figures/full_fig_p089_6_8.png]
Figure 6.8
Figure 6.8. Figure 6.8: From the results, we have the following observations. (1) Models trained on the [PITH_FULL_IMAGE:figures/full_fig_p090_6_8.png]
Figure 7.1
Figure 7.1. Figure 7.1: Illustration of a shift operation followed by a 1x1 convolution. The shift operation [PITH_FULL_IMAGE:figures/full_fig_p093_7_1.png]
Figure 7.2
Figure 7.2. Figure 7.2: Illustration of (a) spatial convolutions, (b) depth-wise convolutions and (c) shift. [PITH_FULL_IMAGE:figures/full_fig_p094_7_2.png]
Figure 7.3
Figure 7.3. Figure 7.3: Illustration of the Conv-Shift-Conv CSC module and the Shift-Conv-Shift-Conv (SC2 ) module. Based on the analysis in previous sections, we propose a module using shift operations as shown in [PITH_FULL_IMAGE:figures/full_fig_p097_7_3.png]
Figure 7.4
Figure 7.4. Figure 7.4: This figure shows that ShiftResNet family members are significantly more efficient [PITH_FULL_IMAGE:figures/full_fig_p100_7_4.png]
Figure 7.5
Figure 7.5. Figure 7.5: Tradeoff curves further to the top left are more efficient, with higher accuracy [PITH_FULL_IMAGE:figures/full_fig_p102_7_5.png]
Figure 7.6
Figure 7.6. Figure 7.6: Style transfer results using shift operators [PITH_FULL_IMAGE:figures/full_fig_p105_7_6.png]
Figure 7
Figure 7. Figure 7: shows different types of operators in ShuffleNetV2 and their FLOP contributions. [PITH_FULL_IMAGE:figures/full_fig_p106_7.png]
Figure 7.7
Figure 7.7. Figure 7.7: ShuffleNetV2 blocks vs. DiracDeltaNet blocks [PITH_FULL_IMAGE:figures/full_fig_p107_7_7.png]
Figure 7.8
Figure 7.8. Figure 7.8: Additive Skip Connections vs. Concatenative Skip Connections. Rectangles [PITH_FULL_IMAGE:figures/full_fig_p108_7_8.png]
Figure 7.9
Figure 7.9. Figure 7.9: Types of operators and their FLOPs in ShuffleNetV2. [PITH_FULL_IMAGE:figures/full_fig_p108_7_9.png]
Figure 7.10
Figure 7.10. Figure 7.10: Transpose based shuffle (ShuffleNetV2) vs. our HW efficient shuffle [PITH_FULL_IMAGE:figures/full_fig_p109_7_10.png]
Figure 7.11
Figure 7.11. Figure 7.11: Progressive quantization schedule [PITH_FULL_IMAGE:figures/full_fig_p112_7_11.png]
Figure 7
Figure 7. Figure 7: shows the overall accelerator architecture design. Our accelerator, highlighted [PITH_FULL_IMAGE:figures/full_fig_p113_7.png]
Figure 7.12
Figure 7.12. Figure 7.12: Accelerator architecture of synetgy [PITH_FULL_IMAGE:figures/full_fig_p114_7_12.png]
Figure 8.1
Figure 8.1. Figure 8.1: Illustration of manual design and reinforcement learning based neural architec [PITH_FULL_IMAGE:figures/full_fig_p118_8_1.png]
Figure 8.2
Figure 8.2. Figure 8.2: Differentiable neural architecture search (DNAS) for ConvNet design. DNAS [PITH_FULL_IMAGE:figures/full_fig_p119_8_2.png]
Figure 8.3
Figure 8.3. Figure 8.3: Illustration of a stochastic super net. Nodes represent data tensors and edges [PITH_FULL_IMAGE:figures/full_fig_p123_8_3.png]
Figure 8.4
Figure 8.4. Figure 8.4: One layer of a super net for mixed-precision quantization of a ConvNet. Nodes [PITH_FULL_IMAGE:figures/full_fig_p126_8_4.png]
Figure 8.5
Figure 8.5. Figure 8.5: Visualization of all searched architectures for ResNet110 on CIFAR10 dataset. [PITH_FULL_IMAGE:figures/full_fig_p129_8_5.png]
Figure 8.6
Figure 8.6. Figure 8.6: The block structure of the micro-architecture search space. Each candidate block [PITH_FULL_IMAGE:figures/full_fig_p133_8_6.png]
Figure 8.7
Figure 8.7. Figure 8.7: Visualization of searched architectures. We use colored boxes to denote blocks [PITH_FULL_IMAGE:figures/full_fig_p137_8_7.png]
Figure 8.8
Figure 8.8. Figure 8.8: Comparison of operator runtime on two devices. Runtime is in micro-second (us). [PITH_FULL_IMAGE:figures/full_fig_p140_8_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

181 extracted references · 22 canonical work pages

  1. [3]

    Shallow Networks for High-Accuracy Road Object-Detection

    Khalid Ashraf et al. “Shallow Networks for High-Accuracy Road Object-Detection”. In: arXiv:1606.01561 (2016)

  2. [1]

    The Vapnik-Chervonenkis dimension: Information versus com- plexity in learning

    Yaser S Abu-Mostafa. “The Vapnik-Chervonenkis dimension: Information versus com- plexity in learning”. In: Neural Computation 1.3 (1989), pp. 312–317

  3. [2]

    Efficient Interactive Annotation of Segmentation Datasets with Polygon-RNN++

    David Acuna et al. “Efficient Interactive Annotation of Segmentation Datasets with Polygon-RNN++”. In: (2018)

  4. [4]

    Label Refinery: Improving ImageNet Classification through Label Progression

    Hessam Bagherinezhad et al. “Label Refinery: Improving ImageNet Classification through Label Progression”. In: arXiv preprint arXiv:1805.02641 (2018)

  5. [5]

    Estimating or propagating gradients through stochastic neurons for conditional computation

    Yoshua Bengio, Nicholas L´ eonard, and Aaron Courville. “Estimating or propagating gradients through stochastic neurons for conditional computation”. In: arXiv preprint arXiv:1308.3432 (2013)

  6. [6]

    FINN-R: An End-to-End Deep-Learning Framework for Fast Exploration of Quantized Neural Networks

    Michaela Blott et al. FINN-R: An End-to-End Deep-Learning Framework for Fast Exploration of Quantized Neural Networks . 2018. arXiv: 1809.04570 [cs.AR]

  7. [7]

    Unsupervised pixel-level domain adaptation with gen- erative adversarial networks

    Konstantinos Bousmalis et al. “Unsupervised pixel-level domain adaptation with gen- erative adversarial networks”. In: CVPR. 2017, pp. 3722–3731

  8. [8]

    A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection

    Zhaowei Cai et al. “A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection”. In: ECCV. 2016

Show all 181 references
  1. [9]

    Fast LIDAR-based Road Detection Using Fully Convolutional Neural Networks

    Luca Caltagirone et al. “Fast LIDAR-based Road Detection Using Fully Convolutional Neural Networks.” In: Intelligent Vehicles Symposium (IV), 2017 IEEE . IEEE. 2017, pp. 1019–1024

  2. [10]

    Annotating object instances with a polygon-rnn

    Lluis Castrejon et al. “Annotating object instances with a polygon-rnn”. In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 2017, pp. 5230–5238

  3. [11]

    Deeplab: Semantic image segmentation with deep con- volutional nets, atrous convolution, and fully connected crfs

    Liang-Chieh Chen et al. “Deeplab: Semantic image segmentation with deep con- volutional nets, atrous convolution, and fully connected crfs”. In: arXiv preprint arXiv:1606.00915 (2016)

  4. [12]

    All You Need is a Few Shifts: Designing Efficient Convolu- tional Neural Networks for Image Classification

    Weijie Chen et al. “All You Need is a Few Shifts: Designing Efficient Convolu- tional Neural Networks for Image Classification”. In:arXiv preprint arXiv:1903.05285 (2019). BIBLIOGRAPHY 136

  5. [13]

    Multi-view 3d object detection network for autonomous driving

    Xiaozhi Chen et al. “Multi-view 3d object detection network for autonomous driving”. In: arXiv preprint arXiv:1611.07759 (2016)

  6. [14]

    Detnas: Neural architecture search on object detection

    Yukang Chen et al. “Detnas: Neural architecture search on object detection”. In: arXiv preprint arXiv:1903.10979 (2019)

  7. [15]

    cuDNN: Efficient Primitives for Deep Learning

    Sharan Chetlur et al. “cuDNN: Efficient Primitives for Deep Learning”. In: arXiv:1410.0759 (2014)

  8. [16]

    PACT: Parameterized Clipping Activation for Quantized Neu- ral Networks

    Jungwook Choi et al. “PACT: Parameterized Clipping Activation for Quantized Neu- ral Networks”. In: arXiv preprint arXiv:1805.06085 (2018)

  9. [17]

    Xception: Deep Learning with Depthwise Separable Convolutions

    Fran¸ cois Chollet. “Xception: Deep Learning with Depthwise Separable Convolutions”. In: arXiv preprint arXiv:1610.02357 (2016)

  10. [18]

    Visual Wake Words Dataset

    Aakanksha Chowdhery et al. “Visual Wake Words Dataset”. In: arXiv preprint arXiv:1906.05721 (2019)

  11. [19]

    Domain adaptation for visual applications: A comprehensive sur- vey

    Gabriela Csurka. “Domain adaptation for visual applications: A comprehensive sur- vey”. In: arXiv:1702.05374 (2017)

  12. [20]

    Histograms of Oriented Gradients for Human Detec- tion

    Navneet Dalal and Bill Triggs. “Histograms of Oriented Gradients for Human Detec- tion”. In: CVPR. 2005

  13. [21]

    Imagenet: A large-scale hierarchical image database

    Jia Deng et al. “Imagenet: A large-scale hierarchical image database”. In: 2009 IEEE conference on computer vision and pattern recognition . Ieee. 2009, pp. 248–255

  14. [22]

    HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precision

    Zhen Dong et al. “HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precision”. In: arXiv preprint arXiv:1905.03696 (2019)

  15. [23]

    CARLA: An Open Urban Driving Simulator

    Alexey Dosovitskiy et al. “CARLA: An Open Urban Driving Simulator”. In: Proceed- ings of the 1st Annual Conference on Robot Learning . 2017, pp. 1–16

  16. [24]

    On the segmentation of 3D LIDAR point clouds

    Bertrand Douillard et al. “On the segmentation of 3D LIDAR point clouds”. In: Robotics and Automation (ICRA), 2011 IEEE International Conference on . IEEE. 2011, pp. 2798–2805

  17. [25]

    On the segmentation of 3D LIDAR point clouds

    Bertrand Douillard et al. “On the segmentation of 3D LIDAR point clouds”. In: ICRA. 2011, pp. 2798–2805

  18. [26]

    Dutta, A

    A. Dutta, A. Gupta, and A. Zissermann. VGG Image Annotator (VIA) . http : / / www.robots.ox.ac.uk/~vgg/software/via/. 2016

  19. [27]

    A Density-based Algorithm for Discovering Clusters a Density- based Algorithm for Discovering Clusters in Large Spatial Databases with Noise

    Martin Ester et al. “A Density-based Algorithm for Discovering Clusters a Density- based Algorithm for Discovering Clusters in Large Spatial Databases with Noise”. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. KDD’96. Portland, ...

  20. [28]

    The Pascal Visual Object Classes (VOC) Challenge

    Mark Everingham et al. “The Pascal Visual Object Classes (VOC) Challenge”. In: IJCV (2010). BIBLIOGRAPHY 137

  21. [29]

    Object detection with discriminatively trained part- based models

    Pedro F. Felzenszwalb et al. “Object detection with discriminatively trained part- based models”. In: PAMI (2010)

  22. [30]

    Scenic: a language for scenario specification and scene gen- eration

    Daniel J Fremont et al. “Scenic: a language for scenario specification and scene gen- eration”. In: Proceedings of the 40th ACM SIGPLAN Conference on Programming Language Design and Implementation . ACM. 2019, pp. 63–78

  23. [31]

    Domain-adversarial training of neural networks

    Yaroslav Ganin et al. “Domain-adversarial training of neural networks”. In: JMLR 17.1 (2016), pp. 2096–2030

  24. [32]

    A Neural Algorithm of Artistic Style

    Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. “A Neural Algorithm of Artistic Style”. In: CoRR abs/1508.06576 (2015). arXiv: 1508.06576 . url: http: //arxiv.org/abs/1508.06576

  25. [33]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. “Are we ready for autonomous driving? the kitti vision benchmark suite”. In: Computer Vision and Pattern Recog- nition (CVPR), 2012 IEEE Conference on . IEEE. 2012, pp. 3354–3361

  26. [34]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. “Are we ready for autonomous driving? the kitti vision benchmark suite”. In: CVPR. 2012, pp. 3354–3361

  27. [35]

    Deep reconstruction-classification networks for unsuper- vised domain adaptation

    Muhammad Ghifary et al. “Deep reconstruction-classification networks for unsuper- vised domain adaptation”. In: ECCV. 2016, pp. 597–613

  28. [36]

    Domain generalization for object recognition with multi- task autoencoders

    Muhammad Ghifary et al. “Domain generalization for object recognition with multi- task autoencoders”. In: ICCV. 2015, pp. 2551–2559

  29. [37]

    SqueezeNext: Hardware-Aware Neural Network Design

    Amir Gholami et al. “SqueezeNext: Hardware-Aware Neural Network Design”. In: arXiv preprint arXiv:1803.10615 (2018)

  30. [38]

    Fast R-CNN

    Ross Girshick. “Fast R-CNN”. In: ICCV. 2015

  31. [39]

    Deformable Part Models are Convolutional Neural Networks

    Ross B. Girshick et al. “Deformable Part Models are Convolutional Neural Networks”. In: CVPR. 2015

  32. [40]

    Rich feature hierarchies for accurate object detection and semantic segmentation

    Ross B. Girshick et al. “Rich feature hierarchies for accurate object detection and semantic segmentation”. In: CVPR. 2014

  33. [41]

    Supplementary Material: Rich feature hierarchies for accurate object detection and semantic segmentation

    Ross B. Girshick et al. “Supplementary Material: Rich feature hierarchies for accurate object detection and semantic segmentation”. In: CVPR. 2014

  34. [42]

    Software-Hardware Codesign for Efficient Neural Network Ac- celeration

    Kaiyuan Guo et al. “Software-Hardware Codesign for Efficient Neural Network Ac- celeration”. In: IEEE Micro 37.2 (2017), pp. 18–25

  35. [43]

    Ms-celeb-1m: Challenge of recognizing one million celebrities in the real world

    Yandong Guo et al. “Ms-celeb-1m: Challenge of recognizing one million celebrities in the real world”. In: Electronic Imaging 2016.11 (2016), pp. 1–6

  36. [44]

    Single path one-shot neural architecture search with uniform sam- pling

    Zichao Guo et al. “Single path one-shot neural architecture search with uniform sam- pling”. In: arXiv preprint arXiv:1904.00420 (2019)

  37. [45]

    The unreasonable effectiveness of data

    Alon Halevy, Peter Norvig, and Fernando Pereira. “The unreasonable effectiveness of data”. In: (2009). BIBLIOGRAPHY 138

  38. [46]

    Deep Compression: Compressing DNNs with Pruning, Trained Quantization and Huffman Coding

    S. Han, H. Mao, and W. Dally. “Deep Compression: Compressing DNNs with Pruning, Trained Quantization and Huffman Coding”. In: ICLR. 2016

  39. [47]

    Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding

    Song Han, Huizi Mao, and William J Dally. “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding”. In: arXiv preprint arXiv:1510.00149 (2015)

  40. [48]

    Achieving human parity on automatic chinese to english news translation

    Hany Hassan et al. “Achieving human parity on automatic chinese to english news translation”. In: arXiv preprint arXiv:1803.05567 (2018)

  41. [49]

    Deep Residual Learning for Image Recognition

    Kaiming He et al. “Deep Residual Learning for Image Recognition”. In: arXiv:1512.03385 (2015)

  42. [50]

    Deep residual learning for image recognition

    Kaiming He et al. “Deep residual learning for image recognition”. In: Proceedings of the IEEE conference on computer vision and pattern recognition . 2016, pp. 770–778

  43. [51]

    Identity mappings in deep residual networks

    Kaiming He et al. “Identity mappings in deep residual networks”. In: European con- ference on computer vision . Springer. 2016, pp. 630–645

  44. [52]

    Mask r-cnn

    Kaiming He et al. “Mask r-cnn”. In: Proceedings of the IEEE international conference on computer vision . 2017, pp. 2961–2969

  45. [53]

    Addressnet: Shift-based primitives for efficient convolutional neural networks

    Yihui He et al. “Addressnet: Shift-based primitives for efficient convolutional neural networks”. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE. 2019, pp. 1213–1222

  46. [54]

    AMC: AutoML for Model Compression and Acceleration on Mobile Devices

    Yihui He et al. “AMC: AutoML for Model Compression and Acceleration on Mobile Devices”. In: Proceedings of the European Conference on Computer Vision (ECCV) . 2018, pp. 784–800

  47. [55]

    LIDAR-based 3D object perception

    Michael Himmelsbach et al. “LIDAR-based 3D object perception”. In: Proceedings of 1st international workshop on cognition for technical systems . Vol. 1. 2008

  48. [56]

    CyCADA: Cycle-Consistent Adversarial Domain Adaptation

    Judy Hoffman et al. “CyCADA: Cycle-Consistent Adversarial Domain Adaptation”. In: ICML. 2018

  49. [58]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications

    Andrew G Howard et al. “Mobilenets: Efficient convolutional neural networks for mobile vision applications”. In: arXiv preprint arXiv:1704.04861 (2017)

  50. [59]

    Searching for mobilenetv3

    Andrew Howard et al. “Searching for mobilenetv3”. In: arXiv preprint arXiv:1905.02244 (2019)

  51. [60]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. “Squeeze-and-excitation networks”. In: CVPR. 2018, pp. 7132–7141

  52. [61]

    Densely Connected Convolutional Networks

    Gao Huang et al. “Densely Connected Convolutional Networks.” In: CVPR. Vol. 1

  53. [62]

    Rethinking the inception architecture for computer vision

    Gary B Huang et al. “Rethinking the inception architecture for computer vision”. In: ECCV Workshops. 2016. BIBLIOGRAPHY 139

  54. [63]

    The apolloscape dataset for autonomous driving

    Xinyu Huang et al. “The apolloscape dataset for autonomous driving”. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops . 2018, pp. 954–960

  55. [64]

    DenseNet: Implementing Efficient ConvNet Descriptor Pyra- mids

    Forrest N. Iandola et al. “DenseNet: Implementing Efficient ConvNet Descriptor Pyra- mids”. In: arXiv:1404.1869 (2014)

  56. [66]

    SqueezeNet: AlexNet-level accuracy with 50x fewer param- eters and ¡0.5MB model size

    Forrest N. Iandola et al. “SqueezeNet: AlexNet-level accuracy with 50x fewer param- eters and ¡0.5MB model size”. In: arXiv:1602.07360 (2016)

  57. [67]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe and Christian Szegedy. “Batch normalization: Accelerating deep network training by reducing internal covariate shift”. In: ICML. 2015, pp. 448–456

  58. [68]

    Categorical reparameterization with gumbel- softmax

    Eric Jang, Shixiang Gu, and Ben Poole. “Categorical reparameterization with gumbel- softmax”. In: arXiv preprint arXiv:1611.01144 (2016)

  59. [69]

    Caffe: Convolutional Architecture for Fast Feature Embedding

    Yangqing Jia et al. “Caffe: Convolutional Architecture for Fast Feature Embedding”. In: arXiv:1408.5093 (2014)

  60. [70]

    Accelerating low bit-width convolutional neural networks with em- bedded FPGA

    Li Jiao et al. “Accelerating low bit-width convolutional neural networks with em- bedded FPGA”. In: Field Programmable Logic and Applications (FPL), 2017 27th International Conference on. IEEE. 2017, pp. 1–4

  61. [71]

    Perceptual Losses for Real-Time Style Transfer and Super-Resolution

    Justin Johnson, Alexandre Alahi, and Fei-Fei Li. “Perceptual Losses for Real-Time Style Transfer and Super-Resolution”. In:CoRR abs/1603.08155 (2016). arXiv: 1603. 08155. url: http://arxiv.org/abs/1603.08155

  62. [72]

    Driving in the Matrix: Can Virtual Worlds Re- place Human-Generated Annotations for Real World Tasks?

    Matthew Johnson-Roberson et al. “Driving in the Matrix: Can Virtual Worlds Re- place Human-Generated Annotations for Real World Tasks?” In:CoRR abs/1610.01983 (2016). url: http://arxiv.org/abs/1610.01983

  63. [73]

    Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks?

    Matthew Johnson-Roberson et al. “Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks?” In: ICRA. 2017, pp. 746–753

  64. [74]

    Local Binary Convo- lutional Neural Networks

    Felix Juefei-Xu, Vishnu Naresh Boddeti, and Marios Savvides. “Local Binary Convo- lutional Neural Networks”. In: arXiv preprint arXiv:1608.06049 (2016)

  65. [75]

    Joint Training of Low-Precision Neural Network with Quantization Interval Parameters

    Sangil Jung et al. “Joint Training of Low-Precision Neural Network with Quantization Interval Parameters”. In: arXiv preprint arXiv:1808.05779 (2018)

  66. [76]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. “Adam: A method for stochastic optimization”. In: arXiv preprint arXiv:1412.6980 (2014)

  67. [77]

    Free supervision from video games

    Philipp Kr¨ ahenb¨ uhl. “Free supervision from video games”. In:CVPR. 2018, pp. 2955– 2964

  68. [78]

    Efficient inference in fully connected crfs with gaussian edge potentials

    Philipp Kr¨ ahenb¨ uhl and Vladlen Koltun. “Efficient inference in fully connected crfs with gaussian edge potentials”. In:Advances in neural information processing systems. 2011, pp. 109–117. BIBLIOGRAPHY 140

  69. [79]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Tech. rep. Citeseer, 2009

  70. [80]

    ImageNet Classification with Deep Convolutional Neural Networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. “ImageNet Classification with Deep Convolutional Neural Networks”. In: NIPS. 2012

  71. [81]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. “Imagenet classification with deep convolutional neural networks”. In: Advances in neural information processing systems. 2012, pp. 1097–1105

  72. [82]

    Maestro: A Memory-on-Logic Architecture for Coordinated Parallel Use of Many Systolic Arrays

    HT Kung et al. “Maestro: A Memory-on-Logic Architecture for Coordinated Parallel Use of Many Systolic Arrays”. In: ()

  73. [83]

    Research methods in human-computer interaction

    Jonathan Lazar, Jinjuan Heidi Feng, and Harry Hochheiser. Research methods in human-computer interaction. Morgan Kaufmann, 2017

  74. [84]

    Extremely low bit neural network: Squeeze the last bit out with admm

    Cong Leng et al. “Extremely low bit neural network: Squeeze the last bit out with admm”. In: arXiv preprint arXiv:1707.09870 (2017)

  75. [85]

    Vehicle detection from 3d lidar using fully convolutional network

    Bo Li, Tianlei Zhang, and Tian Xia. “Vehicle detection from 3d lidar using fully convolutional network”. In: arXiv preprint arXiv:1608.07916 (2016)

  76. [86]

    Adaptive Batch Normalization for practical domain adaptation

    Yanghao Li et al. “Adaptive Batch Normalization for practical domain adaptation”. In: PR 80 (2018), pp. 109–117

  77. [87]

    Tiny-dsod: Lightweight object detection for resource-restricted usages

    Yuxi Li et al. “Tiny-dsod: Lightweight object detection for resource-restricted usages”. In: arXiv preprint arXiv:1807.11013 (2018)

  78. [88]

    FP-BNN: Binarized neural network on FPGA

    Shuang Liang et al. “FP-BNN: Binarized neural network on FPGA”. In: Neurocom- puting 275 (2018), pp. 1072–1086

  79. [89]

    Temporal shift module for efficient video un- derstanding

    Ji Lin, Chuang Gan, and Song Han. “Temporal shift module for efficient video un- derstanding”. In: arXiv preprint arXiv:1811.08383 (2018)

  80. [90]

    Focal loss for dense object detection

    Tsung-Yi Lin et al. “Focal loss for dense object detection”. In: IEEE TPAMI (2018)

  81. [91]

    Microsoft COCO: Common Objects in Context

    Tsung-Yi Lin et al. “Microsoft COCO: Common Objects in Context”. In: CoRR abs/1405.0312 (2014). arXiv: 1405.0312. url: http://arxiv.org/abs/1405.0312

  82. [92]

    Progressive neural architecture search

    Chenxi Liu et al. “Progressive neural architecture search”. In: arXiv preprint arXiv:1712.00559 (2017)

  83. [93]

    Darts: Differentiable architecture search

    Hanxiao Liu, Karen Simonyan, and Yiming Yang. “Darts: Differentiable architecture search”. In: arXiv preprint arXiv:1806.09055 (2018)

  84. [94]

    Coupled generative adversarial networks

    Ming-Yu Liu and Oncel Tuzel. “Coupled generative adversarial networks”. In: NIPS. 2016, pp. 469–477

  85. [95]

    Ssd: Single shot multibox detector

    Wei Liu et al. “Ssd: Single shot multibox detector”. In: European conference on com- puter vision. Springer. 2016, pp. 21–37

  86. [96]

    MetaPruning: Meta Learning for Automatic Neural Network Chan- nel Pruning

    Zechun Liu et al. “MetaPruning: Meta Learning for Automatic Neural Network Chan- nel Pruning”. In: arXiv preprint arXiv:1903.10258 (2019). BIBLIOGRAPHY 141

  87. [97]

    Fully Convolutional Networks for Semantic Segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. “Fully Convolutional Networks for Semantic Segmentation”. In: CVPR. 2015

  88. [98]

    Learning transferable features with deep adaptation net- works

    Mingsheng Long et al. “Learning transferable features with deep adaptation net- works”. In: ICML. 2015, pp. 97–105

  89. [99]

    Shufflenet v2: Practical guidelines for efficient cnn architecture design

    Ningning Ma et al. “Shufflenet v2: Practical guidelines for efficient cnn architecture design”. In: arXiv preprint arXiv:1807.11164 (2018)

  90. [100]

    The concrete distribution: A continuous relaxation of discrete random variables

    Chris J Maddison, Andriy Mnih, and Yee Whye Teh. “The concrete distribution: A continuous relaxation of discrete random variables”. In:arXiv preprint arXiv:1611.00712 (2016)

  91. [101]

    TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems

    Mart´ ın Abadi et al. “TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems”. In: Google Technical Report (2015)

  92. [102]

    3d convolutional neural networks for land- ing zone detection from lidar

    Daniel Maturana and Sebastian Scherer. “3d convolutional neural networks for land- ing zone detection from lidar”. In: Robotics and Automation (ICRA), 2015 IEEE International Conference on. IEEE. 2015, pp. 3471–3478

  93. [103]

    Segmentation of 3D lidar data in non-flat urban environments using a local convexity criterion

    Frank Moosmann, Oliver Pink, and Christoph Stiller. “Segmentation of 3D lidar data in non-flat urban environments using a local convexity criterion”. In: Intelligent Ve- hicles Symposium, 2009 IEEE . IEEE. 2009, pp. 215–220

  94. [104]

    Segmentation of 3D lidar data in non-flat urban environments using a local convexity criterion

    Frank Moosmann, Oliver Pink, and Christoph Stiller. “Segmentation of 3D lidar data in non-flat urban environments using a local convexity criterion”. In: IV. 2009, pp. 215–220

  95. [105]

    Minimal-Entropy Correlation Alignment for Unsupervised Deep Domain Adaptation

    Pietro Morerio, Jacopo Cavazza, and Vittorio Murino. “Minimal-Entropy Correlation Alignment for Unsupervised Deep Domain Adaptation”. In: ICLR. 2018

  96. [106]

    Deep Face Recognition

    Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. “Deep Face Recognition.” In: BMVC. Vol. 1. 3. 2015, p. 6

  97. [107]

    Automatic differentiation in PyTorch

    Adam Paszke et al. “Automatic differentiation in PyTorch”. In: (2017)

  98. [108]

    Visual domain adaptation: A survey of recent advances

    Vishal M Patel et al. “Visual domain adaptation: A survey of recent advances”. In: IEEE SPM 32.3 (2015), pp. 53–69

  99. [109]

    Dark memory and accelerator-rich system optimization in the dark silicon era

    Ardavan Pedram et al. “Dark memory and accelerator-rich system optimization in the dark silicon era”. In: IEEE Design & Test 34.2 (2016), pp. 39–50

  100. [110]

    Efficient Neural Architecture Search via Parameter Sharing

    Hieu Pham et al. “Efficient Neural Architecture Search via Parameter Sharing”. In: arXiv preprint arXiv:1802.03268 (2018)

  101. [111]

    Boosting LiDAR-based semantic labeling by cross-modal train- ing data generation

    Florian Piewak et al. “Boosting LiDAR-based semantic labeling by cross-modal train- ing data generation”. In: Proceedings of the European Conference on Computer Vision (ECCV). 2018

  102. [112]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi et al. “Pointnet++: Deep hierarchical feature learning on point sets in a metric space”. In: NIPS. 2017, pp. 5099–5108. BIBLIOGRAPHY 142

  103. [113]

    Frustum pointnets for 3d object detection from rgb-d data

    Charles R Qi et al. “Frustum pointnets for 3d object detection from rgb-d data”. In: arXiv preprint arXiv:1711.08488 (2017)

  104. [114]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi et al. “Pointnet: Deep learning on point sets for 3d classification and segmentation”. In: CVPR. 2017, pp. 77–85

  105. [115]

    Going deeper with embedded fpga platform for convolutional neural network

    Jiantao Qiu et al. “Going deeper with embedded fpga platform for convolutional neural network”. In: Proceedings of the 2016 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays. 2016, pp. 26–35

  106. [116]

    Language models are unsupervised multitask learners

    Alec Radford et al. “Language models are unsupervised multitask learners”. In: Ope- nAI Blog 1.8 (2019)

  107. [118]

    Xnor-net: Imagenet classification using binary convo- lutional neural networks

    Mohammad Rastegari et al. “Xnor-net: Imagenet classification using binary convo- lutional neural networks”. In: European Conference on Computer Vision . Springer. 2016, pp. 525–542

  108. [119]

    You Only Look Once: Unified, Real-Time Object Detection

    Joseph Redmon et al. “You Only Look Once: Unified, Real-Time Object Detection”. In: CVPR. 2016

  109. [120]

    Faster R-CNN: Towards Real-Time Object Detection with Re- gion Proposal Networks

    Shaoqing Ren et al. “Faster R-CNN: Towards Real-Time Object Detection with Re- gion Proposal Networks”. In: NIPS. 2015

  110. [121]

    Playing for benchmarks

    Stephan R Richter, Zeeshan Hayder, and Vladlen Koltun. “Playing for benchmarks”. In: ICCV. 2017, pp. 2232–2241

  111. [122]

    Playing for Data: Ground Truth from Computer Games

    Stephan R. Richter et al. “Playing for Data: Ground Truth from Computer Games”. In: ECCV. 2016, pp. 102–118

  112. [123]

    Playing for Data: Ground Truth from Computer Games

    Stephan R. Richter et al. “Playing for Data: Ground Truth from Computer Games”. In: European Conference on Computer Vision (ECCV) . Ed. by Bastian Leibe et al. Vol. 9906. LNCS. Springer International Publishing, 2016, pp. 102–118

  113. [124]

    ImageNet Large Scale Visual Recognition Challenge

    Olga Russakovsky et al. “ImageNet Large Scale Visual Recognition Challenge”. In: International Journal of Computer Vision (IJCV) 115.3 (2015), pp. 211–252. doi: 10.1007/s11263-015-0816-y

  114. [125]

    MobileNetV2: Inverted Residuals and Linear Bottlenecks

    Mark Sandler et al. “MobileNetV2: Inverted Residuals and Linear Bottlenecks”. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 2018, pp. 4510–4520

  115. [126]

    Fusing lidar and images for pedestrian detection using convolutional neural networks

    Joel Schlosser, Christopher K Chow, and Zsolt Kira. “Fusing lidar and images for pedestrian detection using convolutional neural networks”. In: Robotics and Automa- tion (ICRA), 2016 IEEE International Conference on . IEEE. 2016, pp. 2198–2205

  116. [127]

    Discovering neural nets with low Kolmogorov complexity and high generalization capability

    J¨ urgen Schmidhuber. “Discovering neural nets with low Kolmogorov complexity and high generalization capability”. In: Neural Networks 10.5 (1997), pp. 857–873. BIBLIOGRAPHY 143

  117. [128]

    Facenet: A unified em- bedding for face recognition and clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. “Facenet: A unified em- bedding for face recognition and clustering”. In: CVPR, pp. 815–823

  118. [129]

    Real-Time and Accurate Segmentation of 3-D Point Clouds Based on Gaussian Process Regression

    Myung-Ok Shin et al. “Real-Time and Accurate Segmentation of 3-D Point Clouds Based on Gaussian Process Regression”. In: IEEE Transactions on Intelligent Trans- portation Systems (2017)

  119. [130]

    Learning from simulated and unsupervised images through adversarial training

    Ashish Shrivastava et al. “Learning from simulated and unsupervised images through adversarial training”. In: CVPR. 2017, pp. 2242–2251

  120. [131]

    Mastering the game of go without human knowledge

    David Silver et al. “Mastering the game of go without human knowledge”. In: Nature 550.7676 (2017), p. 354

  121. [134]

    Very deep convolutional networks for large- scale image recognition

    Karen Simonyan and Andrew Zisserman. “Very deep convolutional networks for large- scale image recognition”. In: arXiv preprint arXiv:1409.1556 (2014)

  122. [135]

    Single-path nas: Designing hardware-efficient convnets in less than 4 hours

    Dimitrios Stamoulis et al. “Single-path nas: Designing hardware-efficient convnets in less than 4 hours”. In: arXiv preprint arXiv:1904.02877 (2019)

  123. [136]

    Throughput-optimized opencl-based fpga accelerator for large- scale convolutional neural networks

    Naveen Suda et al. “Throughput-optimized opencl-based fpga accelerator for large- scale convolutional neural networks”. In: Proceedings of the 2016 International Sym- posium on Field-Programmable Gate Arrays. ACM. 2016, pp. 16–25

  124. [137]

    Correlation alignment for unsupervised domain adaptation

    Baochen Sun, Jiashi Feng, and Kate Saenko. “Correlation alignment for unsupervised domain adaptation”. In: Domain Adaptation in Computer Vision Applications . 2017, pp. 153–171

  125. [138]

    Revisiting unreasonable effectiveness of data in deep learning era

    Chen Sun et al. “Revisiting unreasonable effectiveness of data in deep learning era”. In: Proceedings of the IEEE international conference on computer vision. 2017, pp. 843– 852

  126. [139]

    Going Deeper with Convolutions

    Christian Szegedy et al. “Going Deeper with Convolutions”. In: arXiv:1409.4842 (2014)

  127. [140]

    Rethinking the inception architecture for computer vision

    Christian Szegedy et al. “Rethinking the inception architecture for computer vision”. In: CVPR. 2016, pp. 2818–2826

  128. [141]

    Mnasnet: Platform-aware neural architecture search for mobile

    Mingxing Tan et al. “Mnasnet: Platform-aware neural architecture search for mobile”. In: arXiv preprint arXiv:1807.11626 (2018)

  129. [142]

    Stanley: The robot that won the DARPA Grand Challenge

    Sebastian Thrun et al. “Stanley: The robot that won the DARPA Grand Challenge”. In: Journal of field Robotics 23.9 (2006), pp. 661–692

  130. [143]

    Unbiased look at dataset bias

    Antonio Torralba and Alexei A Efros. “Unbiased look at dataset bias”. In: CVPR. 2011, pp. 1521–1528. BIBLIOGRAPHY 144

  131. [144]

    Adversarial discriminative domain adaptation

    Eric Tzeng et al. “Adversarial discriminative domain adaptation”. In: CVPR. 2017, pp. 2962–2971

  132. [145]

    Instance Normalization: The Missing Ingredient for Fast Stylization

    Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. “Instance Normalization: The Missing Ingredient for Fast Stylization”. In:CoRR abs/1607.08022 (2016). arXiv: 1607.08022. url: http://arxiv.org/abs/1607.08022

  133. [146]

    Learning Time/Memory-Efficient Deep Architec- tures with Budgeted Super Networks

    Tom Veniat and Ludovic Denoyer. “Learning Time/Memory-Efficient Deep Architec- tures with Budgeted Super Networks”. In: arXiv preprint arXiv:1706.00046 (2017)

  134. [147]

    Efficiently Scaling up Crowd- sourced Video Annotation

    Carl Vondrick, Donald Patterson, and Deva Ramanan. “Efficiently Scaling up Crowd- sourced Video Annotation”. In:International Journal of Computer Vision (). 10.1007/s11263- 012-0564-1, pp. 1–21. issn: 0920-5691. url: http://dx.doi.org/10.1007/s11263- 012-0564-1

  135. [148]

    LATTE: Accelerating LiDAR Point Cloud Annotation via Sensor Fusion, One-Click Annotation, and Tracking

    Bernie Wang et al. “LATTE: Accelerating LiDAR Point Cloud Annotation via Sensor Fusion, One-Click Annotation, and Tracking”. In: arXiv preprint arXiv:1904.09085 (2019)

  136. [149]

    LDLS: 3-D Object Segmentation Through Label Diffusion From 2-D Images

    Brian H Wang et al. “LDLS: 3-D Object Segmentation Through Label Diffusion From 2-D Images”. In: IEEE Robotics and Automation Letters 4.3 (2019), pp. 2902–2909

  137. [150]

    What could move? finding cars, pedestrians and bicyclists in 3d laser data

    Dominic Zeng Wang, Ingmar Posner, and Paul Newman. “What could move? finding cars, pedestrians and bicyclists in 3d laser data”. In: ICRA. 2012, pp. 4038–4044

  138. [151]

    PointSeg: Real-Time Semantic Segmentation Based on 3D LiDAR Point Cloud

    Yuan Wang et al. “PointSeg: Real-Time Semantic Segmentation Based on 3D LiDAR Point Cloud”. In: arXiv preprint arXiv:1807.06288 (2018)

  139. [152]

    An Introduction to the Kalman Filter

    Greg Welch and Gary Bishop. An Introduction to the Kalman Filter. Tech. rep. Chapel Hill, NC, USA, 1995

  140. [153]

    A discriminative feature learning approach for deep face recog- nition

    Yandong Wen et al. “A discriminative feature learning approach for deep face recog- nition”. In: ECCV. 2016, pp. 499–515

  141. [154]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning

    Ronald J Williams. “Simple statistical gradient-following algorithms for connectionist reinforcement learning”. In: Machine learning 8.3-4 (1992), pp. 229–256

  142. [155]

    Face recognition in unconstrained videos with matched background similarity

    Lior Wolf, Tal Hassner, and Itay Maoz. “Face recognition in unconstrained videos with matched background similarity”. In: CVPR. 2011, pp. 529–534

  143. [156]

    Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search

    Bichen Wu et al. “Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search”. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019, pp. 10734–10742

  144. [157]

    Mixed Precision Quantization of ConvNets via Differentiable Neural Architecture Search

    Bichen Wu et al. “Mixed Precision Quantization of ConvNets via Differentiable Neural Architecture Search”. In: arXiv preprint arXiv:1812.00090 (2018)

  145. [158]

    Shift: A Zero FLOP, Zero Parameter Alternative to Spatial Con- volutions

    Bichen Wu et al. “Shift: A Zero FLOP, Zero Parameter Alternative to Spatial Con- volutions”. In: arXiv:1711.08141 (2017)

  146. [159]

    SqueezeDet: Unified, Small, Low Power Fully Convolutional Neu- ral Networks for Real-Time Object Detection for Autonomous Driving

    Bichen Wu et al. “SqueezeDet: Unified, Small, Low Power Fully Convolutional Neu- ral Networks for Real-Time Object Detection for Autonomous Driving.” In: CVPR Workshops. 2017, pp. 446–454. BIBLIOGRAPHY 145

  147. [160]

    Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving

    Bichen Wu et al. “Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving”. In: arXiv preprint arXiv:1612.01051 (2016)

  148. [161]

    Squeezeseg: Convolutional neural nets with recurrent crf for real- time road-object segmentation from 3d lidar point cloud

    Bichen Wu et al. “Squeezeseg: Convolutional neural nets with recurrent crf for real- time road-object segmentation from 3d lidar point cloud”. In: 2018 IEEE Interna- tional Conference on Robotics and Automation (ICRA) . IEEE. 2018, pp. 1887–1893

  149. [162]

    Squeezeseg: Convolutional neural nets with recurrent crf for real- time road-object segmentation from 3d lidar point cloud

    Bichen Wu et al. “Squeezeseg: Convolutional neural nets with recurrent crf for real- time road-object segmentation from 3d lidar point cloud”. In: ICRA. 2018

  150. [163]

    Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud

    Bichen Wu et al. “Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud”. In: arXiv preprint arXiv:1809.08495 (2018)

  151. [164]

    A Compact DNN: Approaching GoogLeNet-Level Accuracy of Classification and Domain Adaptation

    Chunpeng Wu et al. “A Compact DNN: Approaching GoogLeNet-Level Accuracy of Classification and Domain Adaptation”. In: CoRR abs/1703.04071 (2017). arXiv: 1703.04071. url: http://arxiv.org/abs/1703.04071

  152. [165]

    Subcategory-aware Convolutional Neural Networks for Object Pro- posals and Detection

    Yu Xiang et al. “Subcategory-aware Convolutional Neural Networks for Object Pro- posals and Detection”. In: arXiv:1604.04693 (2016)

  153. [167]

    Aggregated residual transformations for deep neural networks

    Saining Xie et al. “Aggregated residual transformations for deep neural networks”. In: arXiv preprint arXiv:1611.05431 (2016)

  154. [168]

    Netadapt: Platform-aware neural network adaptation for mobile applications

    Tien-Ju Yang et al. “Netadapt: Platform-aware neural network adaptation for mobile applications”. In: Energy 41 (2018), p. 46

  155. [169]

    Synetgy: Algorithm-hardware Co-design for ConvNet Accelerators on Embedded FPGAs

    Yifan Yang et al. “Synetgy: Algorithm-hardware Co-design for ConvNet Accelerators on Embedded FPGAs”. In: arXiv preprint arXiv:1811.08634 (2018)

  156. [170]

    Imagenet training in minutes

    Yang You et al. “Imagenet training in minutes”. In: Proceedings of the 47th Interna- tional Conference on Parallel Processing . ACM. 2018, p. 1

  157. [171]

    BDD100K: A diverse driving video database with scalable annotation tooling

    Fisher Yu et al. “BDD100K: A diverse driving video database with scalable annotation tooling”. In: arXiv preprint arXiv:1805.04687 (2018)

  158. [172]

    On compressing deep models by low rank and sparse decomposition

    Xiyu Yu et al. “On compressing deep models by low rank and sparse decomposition”. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017, pp. 7370–7379

  159. [173]

    A LiDAR Point Cloud Generator: from a Virtual World to Autonomous Driving

    Xiangyu Yue et al. “A LiDAR Point Cloud Generator: from a Virtual World to Autonomous Driving”. In: ICMR. 2018, pp. 458–464

  160. [174]

    Fast segmentation of 3D point clouds: A paradigm on LiDAR data for autonomous vehicle applications

    Dimitris Zermas, Izzat Izzat, and Nikolaos Papanikolopoulos. “Fast segmentation of 3D point clouds: A paradigm on LiDAR data for autonomous vehicle applications”. In: Robotics and Automation (ICRA), 2017 IEEE International Conference on. IEEE. 2017, pp. 5067–5073. BIBLIOGRAPHY 146

  161. [175]

    Joint face detection and alignment using multitask cascaded convolutional networks

    Kaipeng Zhang et al. “Joint face detection and alignment using multitask cascaded convolutional networks”. In: IEEE Signal Processing Letters 23.10 (2016), pp. 1499– 1503

  162. [176]

    LiSeg: Lightweight Road-object Semantic Segmentation In 3D LiDAR Scans For Autonomous Driving

    Wenquan Zhang et al. “LiSeg: Lightweight Road-object Semantic Segmentation In 3D LiDAR Scans For Autonomous Driving”. In: 2018 IEEE Intelligent Vehicles Sym- posium (IV). IEEE. 2018, pp. 1021–1026

  163. [178]

    Efficient L-Shape Fitting for Vehicle Detection Using Laser Scan- ners

    Xiao Zhang et al. “Efficient L-Shape Fitting for Vehicle Detection Using Laser Scan- ners”. In: 2017 IEEE Intelligent Vehicles Symposium . June 2017

  164. [179]

    ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices. arXiv 2017

    X Zhang et al. “ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices. arXiv 2017”. In: arXiv preprint arXiv:1707.01083 ()

  165. [180]

    Curriculum domain adaptation for semantic segmentation of urban scenes

    Yang Zhang, Philip David, and Boqing Gong. “Curriculum domain adaptation for semantic segmentation of urban scenes”. In: ICCV. 2017, pp. 2039–2049

  166. [181]

    Conditional random fields as recurrent neural networks

    Shuai Zheng et al. “Conditional random fields as recurrent neural networks”. In: Pro- ceedings of the IEEE International Conference on Computer Vision . 2015, pp. 1529– 1537

  167. [182]

    Dorefa-net: Training low bitwidth convolutional neural net- works with low bitwidth gradients

    Shuchang Zhou et al. “Dorefa-net: Training low bitwidth convolutional neural net- works with low bitwidth gradients”. In: arXiv preprint arXiv:1606.06160 (2016)

  168. [183]

    Trained ternary quantization

    Chenzhuo Zhu et al. “Trained ternary quantization”. In: arXiv preprint arXiv:1612.01064 (2016)

  169. [184]

    Training Compact Neural Networks with Binary Weights and Low Precision Activations

    Bohan Zhuang, Chunhua Shen, and Ian Reid. “Training Compact Neural Networks with Binary Weights and Low Precision Activations”. In:arXiv preprint arXiv:1808.02631 (2018)

  170. [185]

    Towards Effective Low-bitwidth Convolutional Neural Networks

    B. Zhuang et al. “Towards Effective Low-bitwidth Convolutional Neural Networks”. In: arXiv preprint arXiv:1711.00205 (2017)

  171. [186]

    Deep Unsupervised Convolutional Domain Adaptation

    Junbao Zhuo et al. “Deep Unsupervised Convolutional Domain Adaptation”. In: ACM MM. 2017, pp. 261–269

  172. [187]

    Neural architecture search with reinforcement learn- ing

    Barret Zoph and Quoc V Le. “Neural architecture search with reinforcement learn- ing”. In: arXiv preprint arXiv:1611.01578 (2016)

  173. [188]

    Learning transferable architectures for scalable image recogni- tion

    Barret Zoph et al. “Learning transferable architectures for scalable image recogni- tion”. In: arXiv preprint arXiv:1707.07012 2.6 (2017)

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.