Distributed learning requires a frequent communication of neural network update data. For this, we present a set of new compression tools, jointly called differential neural network coding (dNNC). dNNC is specifically tailored to efficiently code incremental neural network updates and includes tools for federated BatchNorm folding (FedBNF), structured and unstructured sparsification, tensor row skipping, quantization optimization and temporal adaptation for improved context-adaptive binary arithmetic coding (CABAC). Furthermore, dNNC provides a new parameter update tree (PUT) mechanism, which allows to identify updates for different neural network parameter sub-sets and their relationship in synchronous and asynchronous neural network communication scenarios. Most of these tools have been included into the standardization process of the NNC standard (ISO/IEC 15938-17) edition 2. We benchmark dNNC in multiple federated and split learning scenarios using a variety of NN models and data including vision transformers and large-scale ImageNet experiments: It achieves compression efficiencies of 60% in comparison to the NNC standard edition 1 for transparent coding cases, i.e., without degrading the inference or training performance. This corresponds to a reduction in the size of the NN updates to less than 1% of their original size. Moreover, dNNC reduces the overall energy consumption required for communication in federated learning systems by up to 94%.

Federated learning (FL) scenarios inherently generate a large communication overhead by frequently transmitting neural network updates between clients and server. To minimize the communication cost, introducing sparsity in conjunction with differential updates is a commonly used technique. However, sparse model updates can slow down convergence speed or unintentionally skip certain update aspects, e.g., learned features, if error accumulation is not properly addressed. In this work, we propose a new scaling method operating at the granularity of convolutional filters which 1) compensates for highly sparse updates in FL processes, 2) adapts the local models to new data domains by enhancing some features in the filter space while diminishing others and 3) motivates extra sparsity in updates and thus achieves higher compression ratios, i.e., savings in the overall data transfer. Compared to unscaled updates and previous work, experimental results on different computer vision tasks (Pascal VOC, CIFAR10, Chest X-Ray) and neural networks (ResNets, MobileNets, VGGs) in uni-, bidirectional and partial update FL settings show that the proposed method improves the performance of the central server model while converging faster and reducing the total amount of transmitted data by up to 377×.

This paper presents an improved probability estimation scheme for the entropy coder of Incremental Neural Network Coding (INNC), which is currently under standardization in ISO/IEC MPEG. More specifically, the paper first analyzes the compression performance of INNC and how the bitstream size relates to the neural network (NN) layers. For the layers requiring the most bits, it analyzes the coded NN weight updates and their temporal dependencies. Major finding is that the probability of a significant (i.e., non-zero) update for a weight can depend considerably on whether the weight has been updated before. Based on this finding, the paper proposes a new probability estimation scheme: Depending on whether a significant update has been received before (i.e., based on the weight’s history), the entropy coder models the probability for a current significant update differently. This scheme achieves a bitstream size reduction of about 2% and 1% in a transfer and a federated learning scenario, respectively, without any accuracy loss or significant complexity increase. Therefore, MPEG adopted our history dependent significance probability (HDSP) scheme to its emerging standard for INNC.

This paper presents a CNN to reduce the encoding time of a VVC-based intra-picture encoder. For encoding a 32 x 32 block, the CNN estimates two partitioning parameters that restrict the allowed coding block width and height. To estimate them such that the encoder skips testing inefficient partitioning modes, we train the CNN as follows: First, we generate training data by encoding sequences without the CNN. While encoding, we test all combinations of the two parameters for each 32 x 32 block and store the resulting Lagrangian rate-distortion-time (RDT) cost. We use the recorded cost to derive the loss function when training the CNN. Consequently, the CNN is trained such that it minimizes the Lagrangian RDT cost. Our CNN reduces the encoding time by 50% with a bit rate increase of 0.9%, which outperforms existing CNN-based approaches. Our generic training approach could also be applied for other encoder parameters.

To provide an encoder and a decoder that improve coding efficiency by dividing blocks into non-rectangular parts.SOLUTION: An encoder separates a block 200 into two regions 202a and 202b called wedgelet partitions by one line segment 201. A line segment separating the regions is determined by a start point S and an end point E, both of which are set on the boundary of the block. Partition information is stored in the form of a partition pattern to use wedgelet partitions (areas 202a and 202b).SELECTED DRAWING: Figure 4

This paper presents two new methods for fast VVC intra-picture encoding. Both are based on an approach that uses a CNN for blockadaptive parameter estimation. The parameters restrict the multitype-tree (MTT) partitionings tested by the encoder. The methods aim for an improvement of the approach by further constraints with additional parameters. Adding parameters increases the time required for training data generation exponentially. This raises the question which parameters to add and how. To explore further partitioning restrictions, the first method adds parameters controlling the block sizes the MTT can start from. Although this leads to four parameters, we can exploit that some of their combinations are invalid. To investigate whether testing fewer prediction and transform modes is feasible, the second method adds a single parameter that restricts their number jointly. The paper evaluates hypothetical and actual encoding time reductions for VTM-10.2. The first method outperforms our other and other existing method: The encoding time decreases by 50% with a bit rate increase of 0.7%.

Versatile Video Coding (VVC) is the most recent and efficient video-compression standard of ITU-T and ISO/IEC. It follows the principle of a hybrid, block-based video codec and offers a high flexibility to select a coded representation of a video. While encoders can exploit this flexibility for compression efficiency, designing algorithms for fast encoding becomes a challenging problem. This problem has recently been attacked with data-driven methods that train suitable neural networks to steer the encoder decisions. On the other hand, an optimized and fast VVC software implementation is provided by Fraunhofer’s Versatile Video Encoder VVenC. The goal of this paper is to investigate whether these two approaches can be combined. To this end, we exemplarily incorporate a recent CNN-based approach that showed its efficiency for intra-picture coding in the VVC reference software VTM to VVenC. The CNN estimates parameters that restrict the multi-type tree (MTT) partitioning modes that are tested in rate-distortion optimization. To train the CNN, the approach considers the Lagrangian rate-distortion-time cost caused by the parameters. For performance evaluation, we compare the five operational points reachable with the VVenC presets to operational points that we reach by using the CNN jointly with the presets. Results show that the combination of both approaches is efficient and that there is room for further improvements.

This paper presents a new method for fast VVC intra-picture encoding using a CNN. The CNN operates on the original samples of $\mathbf{32}\times \mathbf{32}$ blocks. Given a current block, it derives for each of the block's multi-type trees (MTTs), which are nested in quad-tree (QT) nodes, a parameter pair. The parameter pairs constrain the minimum width and height of the sub-blocks in their MTTs. This enables the CNN to control the number of tested MTT splits with fine granularity. To skip modes while maintaining the rate-distortion (RD) performance, we train the CNN considering the Lagrangian rate-distortion-time (RDT) cost caused by the derived parameters. First, we generate training data by encoding; when reaching a quad-tree node in a $\mathbf{32}\times \mathbf{32}$ block, we encode the associated MTT with varying parameter pair values and record the resulting the RD and time cost. Then, when the CNN outputs parameters in training, we estimate the related RDT cost of the $\mathbf{32}\times \mathbf{32}$ block using the recorded data. For this, we model the dependency between RDT cost and the parameters by emulating the encoder's RD optimization process. This way, we train the CNN while considering the RDT cost with an accuracy that is sufficient to outperform existing approaches. The approach achieves an encoding time reduction of 50% with a bit rate increase of only 0.7% for VTM-10.2.

This paper presents a thresholding method for processing the predicted samples in the state-of-the-art High Efficiency Video Coding (HEVC) standard. The method applies an integer-based approximation of the discrete cosine transform to an extended prediction block and sets transform coefficients beneath a certain threshold to zero. Transforming back into the sample domain yields the improved prediction signal. The method is incorporated into a software implementation that is conforming to the HEVC standard and applies to both intra and inter predictions. Consequently, bit-rate savings ranging from 2.3% to 8.0% have been measured in terms of the Bjøntegaard-Delta bit rate (BD-rate).

The High Efficiency Video Coding (HEVC) standard has recently been extended to support efficient representation of multiview video and depth-based 3D video formats. The multiview extension, MV-HEVC, allows efficient coding of multiple camera views and associated auxiliary pictures, and can be implemented by reusing single-layer decoders without changing the block-level processing modules since block-level syntax and decoding processes remain unchanged. Bit rate savings compared with HEVC simulcast are achieved by enabling the use of inter-view references in motion-compensated prediction. The more advanced 3D video extension, 3D-HEVC, targets a coded representation consisting of multiple views and associated depth maps, as required for generating additional intermediate views in advanced 3D displays. Additional bit rate reduction compared with MV-HEVC is achieved by specifying new block-level video coding tools, which explicitly exploit statistical dependencies between video texture and depth and specifically adapt to the properties of depth maps. The technical concepts and features of both extensions are presented in this paper.

3D video systems transmit depth maps in order to render synthesized views (SVs) at a receiver. To anticipate this purpose when processing a depth map, a sender-side depth processing algorithm (DPA), e.g. a depth encoder, can also render the SVs, compute their SV distortion (SVD), and adapt to it. This requires a low-complexity algorithm as computational resources are usually limited. We propose such an algorithm in this paper. First, we discuss a measure that relates a depth change to an SVD change using rendering. Then, we present an optimized process combining basic rendering steps, as warping, occlusion handling, interpolation, hole filling, and blending. Furthermore, we analyze which parts of an SV are affected by a depth change and modify the process to re-render only them. The resulting algorithm is significantly less complex than an unoptimized rendering-based variant and quantifies the SVD more accurately than existing estimation methods. The algorithm is used by the 3D-High Efficiency Video Coding reference software encoder as the main method for distortion computation and can also be used by other DPAs.