METHOD AND APPARATUS FOR END-TO-END TASK-ORIENTED LATENT COMPRESSION WITH DEEP REINFORCEMENT LEARNING

Abstract

End-to-end task oriented latent compression using deep reinforcement learning (DRL) is performed by at least one processor and includes generating latent representations of an input image using a first neural network, wherein the latent representations is a sequence of latent signals, encoding the latent signals using a second neural network, generating a set of quantization keys based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to each of the latent signals using a third neural network, generating a set of dequantized numbers representing dequantized representations of the encoded latent signals based on the set of quantization keys using a fourth neural network, generating a reconstructed output based on the set of dequantized numbers, and performing a target task based on the reconstructed output using a fifth neural network.

Description

BACKGROUND

The international standardization organizations ISO/IEC/IEEE are actively searching for AI-based video coding technologies, especially focusing on technologies based on Deep Neural Neworks (DNNs). Various AhGs have been formed to investigate Neural Network Compression (NNR), Video Coding for Machine (VCM), Neural Network-based Video Coding (NNVC), etc. The Chinese AITISA and AVS also established corresponding expert groups to study standardization of similar technologies.

The process of an End-to-End Latent Representation Compression (E2ELRC) can be describe as follows. Given an input image or video sequence x, a DNN Latent Generator first computes a latent representation f, which is passed through a DNN Encoder to compute a compact representation y that is quantized into a discrete-valued quantized representation y. This discrete-valued representation y may be entropy encoded, losslessly, for easy storage and transmission. On the decoder side, the discrete-valued representation y may be recovered from lossless entropy decoding, and used as the input to a DNN Decoder to compute a reconstructed latent representation f. Then, a DNN Task Performer performs target tasks like detection, recognition, segmentation, etc. based on the reconstructed latent representation f. In other words, without the encoding and decoding processes (from latent representation f to the reconstructed latent representation f), the original DNN Latent Generator would compute the latent representation f, which will be directly used by the DNN Task Performer to perform the target tasks. Therefore, the reconstructed latent representation f can be seen as an altered version of the latent representation f. The goal of E2ELRC is to find an effective encoding-decoding mechanism so that the compact representation y is efficient for storage and transmission, and the recovered reconstructed latent representation f can preserve the original task performance.

Quantization is a core process in all compression standards and production, for images, videos, and latent features. Quantization is also one main source of compression quality loss, and improving quantization efficiency can bring large performance gain in image and video compression tasks.

SUMMARY

According to embodiments, a method of end-to-end task oriented latent image compression using deep reinforcement learning is performed by at least one processor and includes generating a plurality of latent representations of an input image using a first neural network, wherein the plurality of latent representations comprise a sequence of latent signals, encoding the plurality of latent representations using a second neural network, generating a set of quantization keys, using a third neural network, based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to the plurality of latent representations, generating a set of dequantized numbers representing dequantized representations of the encoded plurality of latent representations, based on the set of quantization keys, using a fourth neural network, generating a reconstructed output, based on the set of dequantized numbers, and performing a target task based on the reconstructed output using a fifth neural network.

According to embodiments, an apparatus for end-to-end task oriented latent image compression using deep reinforcement learning including at least one memory configured to store program code and at least one processor configured to read the program code and operate as instructed by the program code. The program code includes first generating code configured to cause the at least one processor to generate a plurality of latent representations of an input using a first neural network, wherein the plurality of latent representations comprise a sequence of latent signals, encoding code configured to cause the at least one processor to encode the plurality of latent representations using a second neural network, second generating code configured to cause the at least one processor to generate a set of quantization keys, using a third neural network, based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to the plurality of latent representations, third generating code configured to cause the at least one processor to generate a set of dequantized numbers representing dequantized representations of the encoded plurality of latent representations, based on the set of quantization keys, using a fourth neural network, decoding code configured to cause the at least one processor to decode a reconstructed output, based on the set of dequantized numbers, and performing code configured to cause the at least one processor to perform a target task based on the reconstructed output using a fifth neural network.

According to embodiments, a non-transitory computer-readable medium storing instructions for that, when executed by at least one processor for end-to-end task oriented latent image compression using deep reinforcement learning, cause the at least one processor to generate a plurality of latent representations using a first neural network, wherein the plurality of latent representations comprise a sequence of latent signals, encode the plurality of latent representations using a second neural network, generate a set of quantization keys, using a third neural network, based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to the plurality of latent representations, generate a set of dequantized numbers representing dequantized representations of the encoded plurality of latent representations, based on the set of quantization keys, using a fourth neural network, decode a reconstructed output, based on the set of dequantized numbers, and perform a target task based on the reconstructed output using a fifth neural network.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a diagram of an environment in which methods, apparatuses and systems described herein may be implemented, according to embodiments.

FIG. 2 is a block diagram of example components of one or more devices of FIG. 1.

FIG. 3 is a diagram of a dependent quantization (DQ) mechanism using two quantizers in a DQ design.

FIG. 4(a) is a state diagram of a hand-designed state machine illustrating the switching between the two quantizers in the DQ design.

FIG. 4(b) is a state table representing the state diagram of the hand-designed state machine of FIG. 4(a).

FIG. 5 is a block diagram of a general process of a latent representation compression (LRC) system.

FIG. 6 is a block diagram of an End-To-End Latent Representation Compression (E2ELRC) apparatus, during a test stage, according to embodiments.

FIG. 7 is a detailed block diagram of a DRL Quantization module from the test stage apparatus in FIG. 6, during a test stage, according to embodiments.

FIG. 8 is a detailed block diagram of a DRL Dequantization module from the test stage apparatus in FIG. 6, during a test stage, according to embodiments.

FIG. 9 is a workflow of the DRL Quantization module and the DRL

Dequantization module, during a training stage, according to embodiments.

FIG. 10 is a detailed workflow of a Memory Replay & Weight Update module, during a training stage, according to embodiments.

FIG. 11 is a flowchart of a method of End-To-End Latent Representation Compression (E2ELRC) using Deep Reinforcement Learning (DRL), according to embodiments.

FIG. 12 is a block diagram of an apparatus for End-To-End Latent Representation Compression (E2ELRC) using Deep Reinforcement Learning (DRL), according to embodiments.

DETAILED DESCRIPTION

Embodiments may relate to a framework of End-to-End Latent Representation Compression (E2ELRC) using Deep Reinforcement Learning (DRL). The method takes into consideration both the task performance and the compression efficiency, and optimizes the system jointly.

Instead of encoding and transmitting the original input images/videos, encoding and transmitting latent representations of the original inputs can bring benefits such as reduced transmission costs and improved privacy. For example, a surveillance system aiming at detecting abnormal vehicles does not need to watch the original video streams but only the extracted latent features necessary for the detection task. The VCM and DCM (the Chinese Data Coding for Machine) standards have been formed to investigate latent feature coding techniques to generate encoded latent features that are both efficient for storage and transmission and effective to perform machine vision or human vision tasks.

Traditional image and video coding standards use Dependent Quantization (DQ) or trellis-coded quantization with hand-designed quantization rules. DQ comprises of two quantizers Q₀ and Q₁ and a procedure for switching between them. FIG. 3 gives an example illustration of a DQ mechanism using quantizers Q₀ and Q₁ in the DQ design. The labels above the circles show the associated states and the labels below the circles show the associated quantization keys. On the decoder side, a reconstructed number x′ is determined by an integer key k multiplying a quantization step size A for either of the quantizers Q₀ or Q₁. The switching between quantizers Q₀ and Q₁ may be represented by a state machine with M=2^K DQ states, K≥2 (hence M ≥4), where each DQ state is associated with one of the quantizers Q₀ or Qi. The current DQ state is uniquely determined by the previous DQ state and the value of the current quantization key k_i. For encoding an input stream x₁, x₂, . . . the potential transitions between quantizers Q₀ and Q₁ may be illustrated by a trellis with 2^K DQ states. Thus, selecting the optimal sequence of quantization keys k₁, k₂, . . . is equivalent to finding the trellis path with the minimum Rate-Distortion (R-D) cost. The problem can be solved by the Viterbi algorithm.

Traditionally, the state machine is hand designed empirically. FIG. 4 gives an example of the hand-designed state machine used in the VVC standard with four states. Specifically, FIG. 4(a) is a state diagram of the hand-designed state machine. FIG. 4(b) is a state table representing the state diagram of the hand-designed state machine.

There are three major limitations of the traditional DQ method. First, only two quantizers are used. If the number of quantizers are increased, the bit consumption in encoding the numbers can be reduced. Second, hand-designing the state-machine is not optimal and too expensive to include a large number of DQ states. Increasing the number of quantizers requires increasing the number of DQ states, which can improve the quantization efficiency, but will result in a state machine too complicated to be hand-designed. Finally, the method of key generation and number reconstruction is heuristically designed manually, which is also not optimal. Searching for other better methods requires domain expertise and can be too expensive to be manually designed.

Accordingly, embodiments of the present disclosure may relate to learning-based quantization that is learned by the DRL mechanism. Embodiments may flexibly support various types of quantization methods (e.g., uniform quantization, codebook-based quantization, or deep learning based quantization), and learns the optimal quantizer in a data-driven manner. In addition, embodiments may relate to the entire E2ELRC process jointly, where the DNN Encoder, DNN Decoder, the learning-based quantization methods, the DNN Latent Generator, and the DNN Task Performer may be jointly optimized to provide improved data adaptive compression results.

FIG. 1 is a diagram of an environment 100 in which methods, apparatuses and systems described herein may be implemented, according to embodiments.

As shown in FIG. 1, the environment 100 may include a user device 110, a platform 120, and a network 130. Devices of the environment 100 may interconnect via wired connections, wireless connections, or a combination of wired and wireless connections.

The user device 110 includes one or more devices capable of receiving, generating, storing, processing, and/or providing information associated with platform 120. For example, the user device 110 may include a computing device (e.g., a desktop computer, a laptop computer, a tablet computer, a handheld computer, a smart speaker, a server, etc.), a mobile phone (e.g., a smart phone, a radiotelephone, etc.), a wearable device (e.g., a pair of smart glasses or a smart watch), or a similar device. In some implementations, the user device 110 may receive information from and/or transmit information to the platform 120.

The platform 120 includes one or more devices as described elsewhere herein. In some implementations, the platform 120 may include a cloud server or a group of cloud servers. In some implementations, the platform 120 may be designed to be modular such that software components may be swapped in or out. As such, the platform 120 may be easily and/or quickly reconfigured for different uses.

In some implementations, as shown, the platform 120 may be hosted in a cloud computing environment 122. Notably, while implementations described herein describe the platform 120 as being hosted in the cloud computing environment 122, in some implementations, the platform 120 may not be cloud-based (i.e., may be implemented outside of a cloud computing environment) or may be partially cloud-based.

The cloud computing environment 122 includes an environment that hosts the platform 120. The cloud computing environment 122 may provide computation, software, data access, storage, etc. services that do not require end-user (e.g., the user device 110) knowledge of a physical location and configuration of system(s) and/or device(s) that hosts the platform 120. As shown, the cloud computing environment 122 may include a group of computing resources 124 (referred to collectively as “computing resources 124” and individually as “computing resource 124”).

The computing resource 124 includes one or more personal computers, workstation computers, server devices, or other types of computation and/or communication devices. In some implementations, the computing resource 124 may host the platform 120. The cloud resources may include compute instances executing in the computing resource 124, storage devices provided in the computing resource 124, data transfer devices provided by the computing resource 124, etc. In some implementations, the computing resource 124 may communicate with other computing resources 124 via wired connections, wireless connections, or a combination of wired and wireless connections.

As further shown in FIG. 1, the computing resource 124 includes a group of cloud resources, such as one or more applications (“APPs”) 124-1, one or more virtual machines (“VMs”) 124-2, virtualized storage (“VSs”) 124-3, one or more hypervisors (“HYPs”) 124-4, or the like.

The application 124-1 includes one or more software applications that may be provided to or accessed by the user device 110 and/or the platform 120. The application 124-1 may eliminate a need to install and execute the software applications on the user device 110. For example, the application 124-1 may include software associated with the platform 120 and/or any other software capable of being provided via the cloud computing environment 122. In some implementations, one application 124-1 may send/receive information to/from one or more other applications 124-1, via the virtual machine 124-2.

The virtual machine 124-2 includes a software implementation of a machine (e.g., a computer) that executes programs like a physical machine. The virtual machine 124-2 may be either a system virtual machine or a process virtual machine, depending upon use and degree of correspondence to any real machine by the virtual machine 124-2. A system virtual machine may provide a complete system platform that supports execution of a complete operating system (“OS”). A process virtual machine may execute a single program, and may support a single process. In some implementations, the virtual machine 124-2 may execute on behalf of a user (e.g., the user device 110), and may manage infrastructure of the cloud computing environment 122, such as data management, synchronization, or long-duration data transfers.

The virtualized storage 124-3 includes one or more storage systems and/or one or more devices that use virtualization techniques within the storage systems or devices of the computing resource 124. In some implementations, within the context of a storage system, types of virtualizations may include block virtualization and file virtualization. Block virtualization may refer to abstraction (or separation) of logical storage from physical storage so that the storage system may be accessed without regard to physical storage or heterogeneous structure. The separation may permit administrators of the storage system flexibility in how the administrators manage storage for end users. File virtualization may eliminate dependencies between data accessed at a file level and a location where files are physically stored. This may enable optimization of storage use, server consolidation, and/or performance of non-disruptive file migrations.

The hypervisor 124-4 may provide hardware virtualization techniques that allow multiple operating systems (e.g., “guest operating systems”) to execute concurrently on a host computer, such as the computing resource 124. The hypervisor 124-4 may present a virtual operating platform to the guest operating systems, and may manage the execution of the guest operating systems. Multiple instances of a variety of operating systems may share virtualized hardware resources.

The network 130 includes one or more wired and/or wireless networks. For example, the network 130 may include a cellular network (e.g., a fifth generation (5G) network, a long-term evolution (LTE) network, a third generation (3G) network, a code division multiple access (CDMA) network, etc.), a public land mobile network (PLMN), a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a telephone network (e.g., the Public Switched Telephone Network (PSTN)), a private network, an ad hoc network, an intranet, the Internet, a fiber optic-based network, or the like, and/or a combination of these or other types of networks.

The number and arrangement of devices and networks shown in FIG. 1 are provided as an example. In practice, there may be additional devices and/or networks, fewer devices and/or networks, different devices and/or networks, or differently arranged devices and/or networks than those shown in FIG. 1. Furthermore, two or more devices shown in FIG. 1 may be implemented within a single device, or a single device shown in FIG. 1 may be implemented as multiple, distributed devices. Additionally, or alternatively, a set of devices (e.g., one or more devices) of the environment 100 may perform one or more functions described as being performed by another set of devices of the environment 100.

FIG. 2 is a block diagram of example components of one or more devices of FIG. 1.

A device 200 may correspond to the user device 110 and/or the platform 120. As shown in FIG. 2, the device 200 may include a bus 210, a processor 220, a memory 230, a storage component 240, an input component 250, an output component 260, and a communication interface 270.

The bus 210 includes a component that permits communication among the components of the device 200. The processor 220 is implemented in hardware, firmware, or a combination of hardware and software. The processor 220 is a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), a microprocessor, a microcontroller, a digital signal processor (DSP), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or another type of processing component. In some implementations, the processor 220 includes one or more processors capable of being programmed to perform a function. The memory 230 includes a random access memory (RAM), a read only memory (ROM), and/or another type of dynamic or static storage device (e.g., a flash memory, a magnetic memory, and/or an optical memory) that stores information and/or instructions for use by the processor 220.

The storage component 240 stores information and/or software related to the operation and use of the device 200. For example, the storage component 240 may include a hard disk (e.g., a magnetic disk, an optical disk, a magneto-optic disk, and/or a solid state disk), a compact disc (CD), a digital versatile disc (DVD), a floppy disk, a cartridge, a magnetic tape, and/or another type of non-transitory computer-readable medium, along with a corresponding drive.

The input component 250 includes a component that permits the device 200 to receive information, such as via user input (e.g., a touch screen display, a keyboard, a keypad, a mouse, a button, a switch, and/or a microphone). Additionally, or alternatively, the input component 250 may include a sensor for sensing information (e.g., a global positioning system (GPS) component, an accelerometer, a gyroscope, and/or an actuator). The output component 260 includes a component that provides output information from the device 200 (e.g., a display, a speaker, and/or one or more light-emitting diodes (LEDs)).

The communication interface 270 includes a transceiver-like component (e.g., a transceiver and/or a separate receiver and transmitter) that enables the device 200 to communicate with other devices, such as via a wired connection, a wireless connection, or a combination of wired and wireless connections. The communication interface 270 may permit the device 200 to receive information from another device and/or provide information to another device. For example, the communication interface 270 may include an Ethernet interface, an optical interface, a coaxial interface, an infrared interface, a radio frequency (RF) interface, a universal serial bus (USB) interface, a Wi-Fi interface, a cellular network interface, or the like.

The device 200 may perform one or more processes described herein. The device 200 may perform these processes in response to the processor 220 executing software instructions stored by a non-transitory computer-readable medium, such as the memory 230 and/or the storage component 240. A computer-readable medium is defined herein as a non-transitory memory device. A memory device includes memory space within a single physical storage device or memory space spread across multiple physical storage devices.

Software instructions may be read into the memory 230 and/or the storage component 240 from another computer-readable medium or from another device via the communication interface 270. When executed, software instructions stored in the memory 230 and/or the storage component 240 may cause the processor 220 to perform one or more processes described herein. Additionally, or alternatively, hardwired circuitry may be used in place of or in combination with software instructions to perform one or more processes described herein. Thus, implementations described herein are not limited to any specific combination of hardware circuitry and software.

The number and arrangement of components shown in FIG. 2 are provided as an example. In practice, the device 200 may include additional components, fewer components, different components, or differently arranged components than those shown in FIG. 2. Additionally, or alternatively, a set of components (e.g., one or more components) of the device 200 may perform one or more functions described as being performed by another set of components of the device 200.

A method and an apparatus for a general process of a latent representation compression (LRC) system will now be described in detail with reference to FIG. 5 of the embodiment.

FIG. 5 is a block diagram of an apparatus for a general process of a latent representation compression (LRC) system.

As shown in FIG. 5, the apparatus of the general process includes a DNN Latent Generation module 510, a DNN Encoding module 520, a Quantization module 530, an Entropy Encoding module 540, an Entropy Decoding module 550, a Dequantization module 560, and a DNN Decoding module 570.

Let X denote an input (image, video, audio, or other types of data). The DNN Latent Generation module 510 generates a latent representation F by using a DNN Latent Generator. The latent representation F can be serialized into a sequence of coding signals, F=f₁, f₂, . . . , where a signal f_t can be generally represented as a 4D tensor of size (h, w, c, d). For each signal f_t, the DNN Encoding module 520 computes a DNN encoded representation y_t based on the signal f_t, by using a DNN Encoder. Then, the Quantization module 530 generates a quantized representation y_t based on the encoded representation y_t by using a Quantizer. After that, the Entropy Encoding module 540 encodes the quantized representation y_t into a compact representation y_t for easy storage and transmission, by using an Entropy Encoder. Then, on the decoder side, after receiving the compact representation y_t, the Entropy Decoding module 550 recovers a decoded representation y′_t based on the compact representation y_t, by using an Entropy Decoder. The lossless entropy coding method may be used by the Entropy Encoder and Entropy Decoder, and result in the decoded representation y′_t being equal to the quantized representation y′_t (i.e. y′_t=y_t). Then, the Dequantization module 560 computes a dequantized representation y′_t based on the decoded representation y′_t, by using a Dequantizer. The DNN Decoding module 570 then generates a reconstructed latent representation f_t based on the dequantized representation y′_t, by using an DNN Decoder. Finally, the Perform DNN Task module 580 performs the target task based on the recovered reconstructed latent representation f_t, by using a DNN Task Performer.

The overall goal of the LRC system is to minimize a joint loss L_LRC (f_t, y_t, f_t) that takes into account two aspects: minimizing a Rate-Distortion (R-D) loss so that the quantized representation y_t will have little bit consumption (reflected by a rate loss R_LRC(y_t)) and the reconstructed latent representation f_t is close to the original f_t (reflected by a distortion loss D_LRC(f_t, f_t)); and minimizing a task prediction loss T_LRC(f_t) so that the reconstructed latent representation f_t may perform the original target task well. The joint loss L_LRC(f_t, y_t, f_t) may be computed according to the following equation:

L _LRC(f_t, y_t, f_t), βT_LRC(f_t)+λD_LRC(f_t, f_t)+R_LRC(y_t)   (1)

The distortion loss D_LRC (f_t, f_t) measures the reconstruction error, such as the PSNR and/or SSIM metric. The rate loss R_LRC(y_t) is related to the bit rate of the quantized representation y_t. The hyperparameters β and λ balance the importance of different loss terms.

Since the quantization/dequantization operations are generally not differentiable, the Quantizer/Dequantizer are optimized separately from the DNN Encoder, DNN Decoder, DNN Latent Generato, and DNN Task Performer. For example, previous methods assume linear quantization and approximate a differentiable rate loss R_LRC(y) through entropy estimation, so that the DNN Encoder, DNN Decoder, DNN Latent Generator, and DNN Task Performer may be learned through back-propagation.

Embodiments propose an E2ELRC method, where the DNN Encoder, DNN Decoder, DNN Latent Generator, and DNN Task Performer, as well as the Quantizer and Dequantizer, are jointly learned. Specifically, Deep Reinforcement Learning (DRL) is exploited to combine the optimization of the DNN Encoder, DNN Decoder, DNN Latent Generator, DNN Task Performer, and the optimization of the Quantizer and Dequantizer. The proposed E2ELRC framework is general and broad to accommodate different types of quantization methods and different types of DNN Encoder, DNN Decoder, DNN Latent Generator, and DNN Task Performer network architectures.

A method and an apparatus of an End-to-End Latent Representation Compression (E2ELRC) system using Deep Reinforcement Learning (DRL) will now be described in detail.

FIG. 6 is a block diagram of an E2ELRC apparatus, during a test stage, according to embodiments.

As shown in FIG. 6, the E2ELRC test apparatus includes a DNN Latent Generation module 610, a DNN Encoding module 620, a DRL Quantization module 630, an Entropy Encoding module 640, an Entropy Decoding module 650, a DRL Dequantization module 660, a DNN Decoding module 670, and a Perform DNN Task module 680.

As part of an encoding process, given an input signal X, the DNN Latent Generation module 610 generates the latent representation F by using the DNN Latent Generator. The latent representation F is serialized into the sequence of coding signals, F=f₁, f₂, . . . , where each signal f_t is a 4D tensor of size (h,w,c,d). The DNN Encoding module 620 computes the DNN encoded representation y_t based on the signal f_t, by using the DNN Encoder. The DNN encoded representation y_t can be viewed as a stream of numbers, y_t=y_t,1, y_t,2 . . . . For a batch of m numbers Y_t,i=. . . , y_t,i−1, y_t,i, the DRL Quantization module 630 computes a batch of Quantization Keys (QKs) K_t,i=. . . , k_t,i−1, k_t,i, each QK k_t,l corresponding to each of the encoded representations y_t,l, by using an DRL Quantizer. For a 1-size batch (m=1), numbers are processed one by one, individually. When m>1, numbers are quantized in an organized manner. The numbers may also be organized in different orders. For example, the number may be organized block-wise to preserve the relative location information. Then, the system sends the QKs K_t,i to a decoding process and goes on to process the next batch of numbers Y_t,i+1. Optionally, the QKs K_t,i will be further compressed by the Entropy Encoding module 640 (preferably in a lossless way) for easy storage and transmission.

As part of the decoding process, after receiving the QKs K_t,i, if the received QKs are entropy encoded, the Entropy Decoding module 650 is applied to obtain the entropy decoded QKs K_t,i=. . . , k_t,i−1, k_t,i. Then, the DRL Dequantization module 660 recovers a batch of dequantized numbers Y′_t,i=. . . , y′_t,i−1, y′_t,i by using an DRL Dequantizer, which is a batch in the whole steam of the dequantized representation y′_t. Then, the DNN Decoding module 670 generates the reconstructed output f_t based on the dequantized representation y′_t, by using the DNN Decoder. Finally, the Perform DNN Task module 680 performs the target task based on the recovered reconstructed output f_t, by using the DNN Task Performer. Note that the Entropy Encoding module 640 and Entropy Decoding module 650 are optional and marked by a dotted line in FIG. 6. In an example embodiment, when the Entropy Encoding module 640 and Entropy Decoding module 650 are used, the embodiment takes lossless entropy coding methods, and therefore, therefore the entropy decoded QKs and the QKs computed by the DRL Quantization module 630 are the same (i.e. K_t,i=K_t,i). Therefore, hereafter the same notation (K_t,i) will be used for both the QKs computed by the encoding process and decoding process.

The DRL Quantizer and the DRL Dequantizer in FIG. 6 use learning-based quantization methods. FIG. 7 and FIG. 8 describe a detailed workflow of the DRL Quantization module 630 and the DRL Dequantization module 660, respectively.

As shown in FIG. 7, the DRL Quantization module 630 includes a Compute Key module 710 and a State Prediction module 720.

As part of the encoding process, given the batch of m numbers Y_t,i=. . . , y_t,i−1, y_t,i, according to a batch of previous Quantization States (QSs) S_t,i−1=. . . , s_t,i−2, s_t,i−1, each QS s_t,i−1 corresponding to each of the encoded representations y_t,l, the Compute Key module 710 computes the QKs K_t,i=. . . , k_t,i−1, k_t,i, each QK k_t,1 corresponding to each of the encoded representations y_t,l, by using a Key Generator. Then, the State Prediction module 720 computes the current QSs S_t,i=. . . , s_t,i−1, s_t,i by using a State Predictor.

Given the previous QSs the Key Generator computes the QKs using a quantization method. This quantization method can be a predetermined rule-based method like uniform quantization with a fixed step size, where QK k_t,i is the integer that can best reconstruct the corresponding encoded representation y_t,i as the multiplication of the QK k_t,i with the quantization step size. This quantization method can also be a statistic model like k-means where QK k_t,i is the index of the cluster whose centroid can best reconstruct the encoded representation y_t,i. This disclosure does not put any restrictions on the specific quantization methods used as the Key Generator.

Given the previous QSs S_t,i−1 and the current QKs K_t,i, the State Prediction module 720 computes the current QS s_t,i. In an example embodiment, only the latest QS s_t,i−1 is used by the State Prediction module 720, which is attached to each of the m QKs to form a pair, and all the m pairs are stacked together to form an input matrix of size (m, 2). In another example embodiment, each QK and the corresponding QS form a pair (k_t,l, s_t,l−1), and the m pairs are stacked together to form an input matrix of size (m, 2). The State Prediction module 720 computes the current QS s_t,i based on a State Predictor, which uses a learning-based model to support transition among an arbitrary number of possible states the QS can take. In embodiments, the learning-based model is trained through the Deep Q-Learning (DQN) algorithm, which will be described in detail later.

As shown in FIG. 8, the DRL Dequantization module 660 includes the State Prediction module 720 and a Reconstruction module 810.

As part of the decoding process, after receiving the QKs K_t,i=. . . , k_t,i−1, k_t,i, the State Prediction module 720 computes the current QS s_t,i by using the State Predictor in the same way the encoding process computes the current QS s_t,i, based on the input QKs K_t,i and previous QSs S_t,i−1=. . . , s_t,i−2, s_t,i−1. Then, the Reconstruction module 810 computes the batch of dequantized numbers Y′_t,i=. . . , y′_t,i−1, y′_t,i based on the QKs K_t,i and QSs S_t,i−1, by using a Reconstructor. The Reconstructor uses a dequantization method that corresponds to the quantization method used in the Key Generator. For example, when the quantization method is predetermined rule-based method like uniform quantization with a fixed step size, the dequantization method is also predetermined rule-based such as computing the dequantized number y′_t,i as the multiplication of the QK k_t,i with the quantization step size. When the quantization method is a statistic model like k-means, the dequantization method may be the centroid indexed by the QK k_t,i. This disclosure does not put any restrictions on the specific dequantization methods used as the Reconstructor.

The State Predictor is an action-value mapping function f (a_j, ν_j|K_t,i, S_t,i−1) between an action a_j and an output Q-value ν_j associated with the action, j=1, . . . , J (assuming we have J possible actions in total), given the QKs K_t,i and QSs S_t,i−1. Each action a_j corresponds to a possible state that QS s_t,i can take. Given the current QKs K_t,i and QSs the State Predictor computes the Q-values ν_j of all possible actions a_j, and selects the optimal action a_i* with the optimal Q-value ν_i* . The state corresponding to the optimal action a_i* is the QS s_i the system selects. The Q-value is designed to measure the target compression performance associated with the sequence of actions. Therefore, selecting the optimal action gives the optimal target compression performance.

The Deep Q-learning mechanism, specifically the DQN algorithm, is used as the training method in embodiments. DQN is an off-policy DRL method, which finds an optimal action selection policy for any given finite Markov Decision Process by learning the action-value mapping function to assign a reward Q-value to an action. A policy is a rule that the system follows in selecting actions. Given a current status, the learning agent may choose from a set of candidate actions, which result in different reward values. By experiencing various status and trying out various actions being at various status, the learning agent learns overtime to optimize the rewards so that it can behave optimally in the future at any given status it is in.

Specifically, a DNN is used as the State Predictor, which acts as a function approximator to estimate the action-value mapping function f (a_j, v_j|K_t,i, S_t,i−1). The State Predictor DNN typically comprises of a set of convolutional layers followed by one or multiple fully connected layers. This disclosure does not put any restrictions on the specific network architectures of the State Predictor.

The training process of the DRL Quantization module 630 and DRL Dequantization module 660 according to embodiments will now be described. An overall workflow of the training process is illustrated in FIG. 9.

As shown in FIG. 9, the E2ELRC system training apparatus includes the DNN Latent Generation module 610, the DNN Encoding module 620, the DNN Decoding module 670, the Perform Task module 680, the Compute Key module 710, the State Prediction module 720, the Reconstruction module 810, a Compute Distortion module 910, a Compute Rate module 920, a Compute Reward module 930, a Memory Replay & Weight Update module 940, a Compute LRC Distortion module 950, a Compute LRC Rate module 960, and an LRC Weight Update module 970.

Let State(t_s−1) be the current State Predictor, let Key(t_k−1) denote the current Key Generator, let Recon(t_r−1) be the current Reconstructor, let Enc (t_e−1) be the current DNN Encoder, let Dec(t_d−1) be the current DNN Decoder, let Latent(t_l−1) be the current DNN Latent Generator, and let Task(t_t−1) be the current DNN Task Performer. t_s, t_k, t_r, t_e, t_l and t_t may be different, so that the State Predictor, the Key Generator, the Reconstructor, the DNN Encoder, the DNN Decoder, the DNN Latent Generator, and the DNN Task Performer may be updated at different times with different updating frequencies.

Given the training input X, the DNN Latent Generation module 610 computes the sequence of latent signals F=f₁, f₂, . . . , using the current DNN Latent Generator Latent(t_l−1). For each signal f_t, the DNN Encoding module 620 uses the current DNN Encoder Enc(t_e−1) to compute the DNN encoded representation y_t=y_t,1, y_t,2 . . . . For the batch of m numbers Y_t,i=. . . , y_t,i−1, y_t,i, according to the previous QSs S_t,i−1=. . . , s_t,i−2, s_t,i−1, the Compute Key module 710 computes the QKs K_t,i=. . . , k_t,i−1, k_t,i, by using the current Key Generator Key(t_k−1). The batch size and the way the numbers are organized are the same as the test stage. Then, the State Prediction module 720 uses the current State Predictor State(t_s−1) to compute the current QS s_t,i, based on the previous QSs S_t,i−1 and the current QKs K_t,i. The input of the State Prediction module 720 is also the same as the test stage. Then, the Reconstruction module 810 uses the current Reconstructor Recon(t_r−1) to compute the batch of dequantized numbers Y′_t,i=. . . , y′_t,i−1, y′_t,i based on the QKs K_t,i and QSs S_t,i−1. Finally, the DNN Decoding module 670 generates a reconstructed z_t based on the dequantized number y′_t, by using the current DNN Decoder Dec(t_d−1).

In the training process, the State Predictor selects the optimal action a_i* using an ϵ-greedy method. Specifically, after the current State Predictor State(t_s−1) computes the Q-values ν_j of all possible actions a_j, with probability c (a number between 0 and 1), a random action will be selected as the optimal action a_i*, and with probability (1−ϵ), the optimal action a_i* with the optimal Q-value ν_i* will be selected.

The Compute Distortion module 910 computes a distortion loss to D(Y_t,i, Y′_t,i) to measure the difference between the original DNN encoded representation Y_t,i and the decoded representation Yt _i For example, the distortion loss can be the average of the L_k-norm, e.g., L₁-norm as Mean Absolute Error and L₂-norm as Mean Square Error, of the difference between the corresponding elements in the encoded representation Y_t,i and the decoded representation Y′_t,i:

D(Y_t,i, Y′_t,i)=avg _l=i−m+1ⁱ∥y_t,l−y′_t,l∥^k   (2)

At the same time, the Compute Rate module 920 computes a rate loss R(K_t,i) to measure the bit consumption of the quantized representation, i.e., the computed QKs K_t,i that are sent from the Encoder to Decoder. There are multiple ways to compute the rate loss. For example, the QKs may be compressed using any lossless entropy coding method and the actual bit count of the compressed bitstream may be obtained as the rate loss.

For an adjacent batch of numbers Y_t,i and Y_t,i+1, based on the distortion D(Y_t,i, Y′_t,i) and D(Y_t,i+1, Y′_t,i+1), and the rate loss R(K_t,i) and R(K_t,i+1), the Compute Reward module 930 computes a reward ϕ (Y_t,i+1, K_t,i+1, Y′_t,i+1). The reward ϕ(Y_t,i+1, K_t,i+1, Y′_t,i+1) measures the reward the State Predictor can get by taking the optimal action a_i* given the current QKs K_t,i and QSs S_t,i−1 according to the following equation:

ϕ(Y_t,i+1, K_t,i+1, Y′_t,i+1)=D(Y_t,i+1, Y′_t,i+1)+αR(K_t,i+1)   (3)

where a is a hyperparameter to balance the rate loss and distortion in the reward. An experience E{ϕ(Y_t,i+1, K_t,i+1, Y′_t,i+1), α*_i, ν*_i, Y_t,i, S_t,i−1, K_t,i}, i.e., selecting action α_i* with associated Q-value ν_i* based on QKs K_t,i and QSs S_t,i−1 and then obtaining the reward ϕ(Y_t,i+1, K_t,i+1, Y′_t,i+1), is added into a Replay Memory. The Replay Memory usually has a maximum storage limit and once it reaches its limit, the oldest experience will be replaced by the latest one.

When it is time to update the State Predictor, the Key Generator, and the Reconstructor, the system samples a batch of experiences from the Replay Memory, and uses these sampled experiences to update the model parameters in the Memory Replay & Weight Update module 940. FIG. 10 is a detailed workflow of the Memory Replay & Weight Update module 940 during the training stage.

As shown in FIG. 10, the Memory Replay & Weight Update module 940 includes the Compute Key module 710, the State Prediction module 720, the Reconstruction module 810, the Compute Distortion module 910, the Compute Rate module 920, and the Compute Reward module 930, a Sample Experience module 1001, a Compute Loss module 1002, and a Weight Update module 1003.

During the training stage, a Target State Predictor State^T, a Target Key Generator Key^T, and a Target Reconstructor Recon^T are maintained and have exactly the same model structure as the State Predictor, the Key Generator, and the Reconstructor, respectively. The only difference is the model parameters, such as the DNN weight coefficients of the State Predictor, or the k-means model parameter of the Key Generator when k-means quantization is used, or the DNN weight coefficients of the Key Generator when quantization is based on deep clustering. These model parameters are cloned from the corresponding State Predictor, Key Generator and Reconstructor at every T_s, T_k and T_r parameter updating cycles.

During each parameter updating cycle, the Sample Experience module 1001 samples a set of experiences from the Replay Memory (E{ϕ(Y_t,l+1, K_t,l+1, Y′_t,l+1), α*_l, ν*_l, Y_t,l, S_t,l−1, K_t,l}). The State Prediction module 720, for each experience E{ϕ(Y_t,l+1, K_t,l+1, Y′_t,l+1), α*_l, ν*_l, Y_t,l, S_t,l−1, K_t,l}, uses the Target State Predictor State^T to predict a target QS ŝ_t,l based on the QKs Y_t,l and QSs S_t,l−1 in the experience. Based on the target QS ŝ_t,l, the Target Key Generator Key^T computes a target key {circumflex over (K)}_t,l+1 in the Compute Key module 710. Based on the target key {circumflex over (K)}_t,l+1 and the target QSs Ŝ_t,l, the Target Reconstructor Recon^T can compute a batch of target dequantized numbers Ŷ′_t,l+1=. . . , y′_t,l, ŷ′_t,l+1 in the Reconstruction module 810. Then, the Compute Distortion module 910 computes a target distortion D(Y_t,l+1, Ŷ′_t,l+1) between the original representation Y_t,l+1 in the experience and the decoded representation Ŷ′_t,l+1. The Compute Rate module 920 computes a target rate loss R({circumflex over (K)}_t,l+1) based on the target key {circumflex over (K)}_t,l+1. A target reward ϕ(Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1) is then computed in the Compute Reward module 930 as:

ϕ(Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1)=D(Y_t,l+1, Ŷ′_t,l+1)+αR({circumflex over (K)}_t,l+1)   (4)

Then, the Compute Loss module 1002 computes a target reward T(α*_l+1, Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1, Ŝ_t,l) as:

T(α*_l+1, Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1, Ŝ_t,l)=ϕ(Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1)+γmax_j{circumflex over (Q)}(α*_l+1, {circumflex over (K)}_t,l+1, Ŝ_t,l)   (5)

where {circumflex over (Q)}(α*_l+1, {circumflex over (K)}_t,l+1, Ŝ_t,l) is the Q-value predicted by the Target State Predictor State^T for action α*_j given the QKs {circumflex over (K)}_t,l+1 and QSs Ŝ_t,l. The hyperparameter γ is the discount rate valued between 0 and 1, which determines how important the system weights long-term rewards against short-term ones. The smaller the discount rate, the system weights less on long-term rewards but cares only for the short-term rewards. Then a target loss L(α*_l+1, ν*_l, Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1, Ŝ_t,l) is computed, based on the target reward T(α*_l+1, Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1, Ŝ_t,l) and the Q-value ν*_l from the experience, e.g., L_k-norm of the difference between the two rewards:

L(α*_l+1, ν*_l, Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1, Ŝ_t,l)=∥T(α*_l, Y_t,l, S_t,l−1)−ν*_l∥^k   (6)

Then, the Weight update module 1003 computes a gradient of the target loss, which is back-propagated to update the weight parameters of the DNNs of the State Predictor into State(ts). The gradient of the target loss may also be used in combination with the optimization objectives of the learning-based Key Generator and Reconstructor to update the Key Generator Key(t_k) and the Reconstructor Recon(t_r). For example, in a case where the Key Generator and Reconstructor use quantization methods based on deep clustering, weight parameters of the DNNs for the Key Generator and Reconstructor are updated through back-propagation. When other learning based methods are used for quantization, the model parameters are learned by optimizing an objective function, and the target loss L(α*_l+1, ν*_l, Y_t,l+1, {circumflex over (K)}_t,l+1, Ŷ′_t,l+1, Ŝ_t,l) may be weighted and added to the optimization objective function as additional regularization terms to update the model parameters. As mentioned before, the State Predictor, the Key Generator, and the Reconstructor may be updated at different time stamps.

For every T_s, T_k and T_r iterations, the weight parameters of the State Predictor, the Key Generator, and the Reconstructor will be cloned to the Target State Predictor State^T, the Target Key Generator Key^T, and the Target Reconstructor Recon^T, respectively.

Embodiments use the Replay Memory, the Target State Predictor, the Target Key Generator, and the Target Reconstructor to stabilize the training process. The Replay Memory can only have one latest experience, which equals to not having a Replay Memory. Also, T_s, T_k and T_r may all be equal to 1 so that the Target State Predictor, the Target Key Generator, and the Target Reconstructor will be updated for every iteration, which equals to not having another set of Target State Predictor, Target Key Generator, and Target Reconstructor.

As for the entire E2ELRC system (described in FIG. 9) for each input X, the DNN Latent Generation module 610 uses the current DNN Latent Generator Latent(t_l−1) to compute the sequence of latent signals F=f₁, f₂, . . . , . For each signal f_t, the DNN Encoding module 620 uses the current DNN Encoder Enc(te-1) to compute the DNN encoded representation y_t=y_t,1, y_t,2, . . . . Through the DRL Quantization module 630 and the DRL Dequantization module 660, the dequantized representations y′_t=y′_t,1, y′_t,2, . . . are generated. Then, the DNN Decoding module 670 generates the reconstructed latent representation f_t based on the dequantized representation y′_t by using the current DNN Decoder Dec(t_d−1). Finally, the Perform DNN Task module 680 performs the target task based on the reconstructed latent representation f_t by using the current DNN Task Performer Task(t_t−1) and computes a task prediction loss T_LRC(f_t) based on the training labels (e.g., classification or regression loss of the original task).

Then, the Compute LRC Distortion module 950 computes a latent representation distortion loss D_LRC(f_t, f_t) to measure the error introduced by the latent representation compression process, such as the PSNR and/or SSIM related metrics. The Compute LRC Rate module 960 computes a latent compression rate loss R_LRC(y_t), for example, by non-parametric density estimation based on the quantized representation y_t (i.e., the QKs k_t,1, k_t,2, . . . that are stored and transmitted to the decoding process) with a uniform density or normal density. Then, the overall joint loss L_LRC(f_t, y_t, f_t) may be computed as:

L _LRC(f_t, y_t, f_t)=βT_LRC(f_t)+λD_LRC(f_t, f_t)+R_LRC(y_t)   (7)

The hyperparameters β and λ and balance the importance of different loss terms.

Then, the LRC Weight Update module 970 computes a gradient of the joint loss(e.g., by summing up the gradient of the joint loss over several input data), which may update the weight parameters of the DNN Encoder, the DNN Decoder, the DNN Latent Generator, and the DNN Task Performer into Enc(t_e), Dec(t_d), Latent(t_l) and Task(t_t), respectively, through back-propagation.

In embodiments, the DNN Latent Generator and the DNN Task Performer are pre-trained (denoted by Latent(0) and Task(0) respectively), by omitting encoding/decoding processes. In such a pre-training process, given a pre-training input X, the DNN Latent Generation module 610 computes the latent representation F, which is directly used by the Perform DNN Task module 680. The task prediction loss T_LRC(f_t) can then be computed, whose gradients are back-propagated to learn the DNN Latent Generator and the DNN Task Performer.

Also, in embodiments, the DNN Encoder and DNN Decoder are pre-trained (denoted by Enc(0) and Dec(0) respectively), by assuming the uniform quantization method and estimating the latent compression rate loss R_LRC(y_t) by an entropy estimation model. In such a pre-training process, given a pre-training latent signal f_t, the DNN Encoder computes representation y_t, which is further used by the entropy estimation model to compute the latent compression rate loss R_LRC(y_t). The DNN Decoder then computes the output (reconstructed latent representation f_t) based on the representation y_t. The latent distortion loss D_LRC(f_t, f_t) may then be computed and an R-D loss obtained as:

λD_LRC(f_t, f_t)+R_LRC(y_t),   (8)

whose gradient may be used to update the DNN Encoder and DNN Decoder through back-propagation.

When the pre-trained DNN Encoder, DNN Decoder, DNN Latent Generator and DNN Task Performer are deployed, the training process described in embodiments of FIG. 9 and FIG. 10 train the DRL Quantizer and DRL Dequantizer to cope with the DNN Encoder, DNN Decoder, DNN Latent Generator and DNN Task Performer to improve the quantization performance. The described training process may also update the DNN Encoder, DNN Decoder, DNN Latent Generator and DNN Task Performer according to the current training data so that the entire latent compression system can adaptively improve the total compression performance and task performance. The update of the DNN Encoder, DNN Decoder, DNN Latent Generator and DNN Task Performer may happen offline or online and may be permanent or temporary data dependent.

Similarly, after deployed, the State Predictor, the Key Generator, and the Reconstructor in the DRL Quantizer and DRL Dequantizer may also be updated offline or online and may be permanently or temporarily data dependent. For example, in the case of video-based tasks, some or all of the DNN Encoder, DNN Decoder, DNN Latent Generator, DNN Task Performer State Predictor, Key Generator, and Reconstructor may be updated based on the first few frames. But these updates will not be recorded to influence computation for future videos. Such updates may also be accumulated to a certain amount based on which modules may be updated permanently to be applied to future videos. In terms of parameter updates, part of the model parameters of a DNN may be frozen and only the remaining parameters updated. This disclosure does not put any restrictions on which DNN models to update or which part of the weight parameters to update in the DNN models.

FIG. 11 is a flowchart of a method of end-to-end latent representation compression using deep reinforcement learning, according to embodiments.

In some implementations, one or more process blocks of FIG. 11 may be performed by the platform 120. In some implementations, one or more process blocks of FIG. 11 may be performed by another device or a group of devices separate from or including the platform 120, such as the user device 110.

As shown in FIG. 11, in operation 1101, the method includes generating a plurality of latent representations of an input using a first neural network. The plurality of latent representations may be a sequence of latent signals.

In operation 1102, the method includes encoding the plurality of latent representations using a second neural network.

In operation 1103, the method includes generating a set of quantization keys, using a third neural network, based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to the plurality of latent representations. A set of encoded quantization keys may also be generated by entropy encoding the set of quantization keys.

A set of current quantization states, based on the set of previous quantization states and the set of quantization keys, by training the third neural network. The third neural network is trained by computing q-values for all possible actions, randomly selecting an action as an optimal action with an optimal q-value, generating a reward of the selected optimal action, sampling a set of selected optimal actions, and updating weight parameters of the third neural network to minimize distortion loss.

In operation 1104, the method includes generating a set of dequantized numbers representing dequantized representations of the encoded plurality of latent representations, based on the set of quantization keys, using a fourth neural network. If the set of encoded quantization keys are generated, a set of decoded quantization keys may also be generated by entropy decoding the set of encoded quantization keys and the set of dequantized numbers are instead generated based on the set of decoded quantization keys.

The set of quantization keys generated in operation 1103 and the set of dequantized numbers generated in operation 1104 are quantized and dequantized, respectively, using a block-wise quantization/dequantization method, individual quantization/dequantization method, or a static quantization/dequantization model method. Further, the quantization method of the set of quantization keys and the dequantization method of the set of dequantized numbers are the same.

In operation 1105, the method includes generating a reconstructed output, based on the set of dequantized numbers.

In operation 1106, the method includes performing a target task based on the reconstructed output using a fifth neural network.

Instead, the target task may be performed based on the generated plurality of latent representations. A task prediction loss, based on the target task, may also be computed wherein the first neural network and the fifth neural network are trained by back-propagating a gradient of the task prediction loss and updating weight parameters of the first neural network and the fifth neural network.

Although FIG. 11 shows example blocks of the method, in some implementations, the method may include additional blocks, fewer blocks, different blocks, or differently arranged blocks than those depicted in FIG. 11. Additionally, or alternatively, two or more of the blocks of the method may be performed in parallel.

FIG. 12 is a block diagram of an apparatus for end-to-end latent representation compression using deep reinforcement learning, according to embodiments.

As shown in FIG. 12, the apparatus includes first generating code 1201, encoding code 1202, second generating code 1203, third generating code 1204, decoding code 1205, and performing code 1206.

The first generating code 1201 is configured to cause the at least one processor to generate a plurality of latent representations of an input using a first neural network, wherein the plurality of latent representations comprise a sequence of latent signals.

The encoding code 1202 is configured to cause the at least one processor to encode the plurality of latent representations using a second neural network.

The second generating code 1203 is configured to cause the at least one processor to generate a set of quantization keys, using a third neural network, based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to the plurality of latent representations.

Further, the operations of the apparatus may also include state generating code configured to cause the at least one processor to a set of current quantization states, based on the set of previous quantization states and the set of quantization keys, by training the third neural network. The third neural network is trained by computing q-values for all possible actions, randomly selecting an action as an optimal action with an optimal q-value, generating a reward of the selected optimal action, sampling a set of selected optimal actions, and updating weight parameters of the third neural network to minimize distortion loss.

The third generating code 1204 is configured to cause the at least one processor to generate a set of dequantized numbers representing dequantized representations of the encoded plurality of latent representations, based on the set of quantization keys, using a fourth neural network.

The set of quantization keys generated by the second generating code 1203 and the set of dequantized numbers generated by the third generating code 1204 may be quantized and dequantized, respectively, using a block-wise quantization/dequantization method, individual quantization/dequantization method, or a static quantization/dequantization model method. Further, the quantization method of the set of quantization keys and the dequantization method of the set of dequantized numbers are the same.

The decoding code 1205 is configured to cause the at least one processor to decode a reconstructed output, based on the set of dequantized numbers.

The performing code 1206 is configured to cause the at least one processor to perform a target task based on the reconstructed output using a fifth neural network.

Instead, the target task may be performed based on the generated plurality of latent representations. The apparatus of FIG. 12 may also include computing code configured to cause the at least one processor to compute a task prediction loss based on the target task, wherein the first neural network and the fifth neural network are trained by back-propagating a gradient of the task prediction loss and updating weight parameters of the first neural network and the fifth neural network

Although FIG. 12 shows example blocks of the apparatus, in some implementations, the apparatus may include additional blocks, fewer blocks, different blocks, or differently arranged blocks than those depicted in FIG. 12. Additionally, or alternatively, two or more of the blocks of the apparatus may be combined.

Embodiments relate to an End-to-End Latent Representation Compression (E2ELRC) that improves compression performance by optimizing the latent representation compression for performing the a target task as an entire system. This method provides the flexibility to adjust learning-based quantization and encoding methods, online or offline based on the current data, and support different types of learning-based quantization methods, including DNN-based or conventional model-based methods. The described method also provides a flexible and general framework that accommodates different DNN architectures and tasks.

The proposed methods may be used separately or combined in any order. Further, each of the methods (or embodiments) may be implemented by processing circuitry (e.g., one or more processors or one or more integrated circuits). In one example, the one or more processors execute a program that is stored in a non-transitory computer-readable medium.

The present disclosure provides illustration and description, but is not intended to be exhaustive or to limit the implementations to the precise form disclosed. Modifications and variations are possible in light of the present disclosure or may be acquired from practice of the implementations.

As used herein, the term component is intended to be broadly construed as hardware, firmware, or a combination of hardware and software.

It will be apparent that systems and/or methods, described herein, may be implemented in different forms of hardware, firmware, or a combination of hardware and software. The actual specialized control hardware or software code used to implement these systems and/or methods is not limiting of the implementations. Thus, the operation and behavior of the systems and/or methods were described herein without reference to specific software code—it being understood that software and hardware may be designed to implement the systems and/or methods based on the description herein.

Even though combinations of features are recited in the claims and/or disclosed in the specification, these combinations are not intended to limit the disclosure of possible implementations. In fact, many of these features may be combined in ways not specifically recited in the claims and/or disclosed in the specification. Although each dependent claim listed below may directly depend on only one claim, the disclosure of possible implementations includes each dependent claim in combination with every other claim in the claim set.

No element, act, or instruction used herein may be construed as critical or essential unless explicitly described as such. Also, as used herein, the articles “a” and “an” are intended to include one or more items, and may be used interchangeably with “one or more.” Furthermore, as used herein, the term “set” is intended to include one or more items (e.g., related items, unrelated items, a combination of related and unrelated items, etc.), and may be used interchangeably with “one or more.” Where only one item is intended, the term “one” or similar language is used. Also, as used herein, the terms “has,” “have,” “having,” or the like are intended to be open-ended terms. Further, the phrase “based on” is intended to mean “based, at least in part, on” unless explicitly stated otherwise.

Claims

1. A method of end-to-end task oriented latent image compression using deep reinforcement learning, the method being performed by at least one processor, and the method comprising: generating a plurality of latent representations of an input using a first neural network, wherein the plurality of latent representations comprise a sequence of latent signals;encoding the plurality of latent representations using a second neural network;generating a set of quantization keys, using a third neural network, based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to the plurality of latent representations;generating a set of dequantized numbers representing dequantized representations of the encoded plurality of latent representations, based on the set of quantization keys, using a fourth neural network;generating a reconstructed output, based on the set of dequantized numbers; andperforming a target task based on the reconstructed output using a fifth neural network.

2. The method of claim 1, further comprising computing a task prediction loss based on the target task, wherein the first neural network and the fifth neural network are trained by back-propagating a gradient of the task prediction loss and updating weight parameters of the first neural network and the fifth neural network.

3. The method of claim 1, wherein the target task is performed based on the generated plurality of latent representations.

4. The method of claim 1, further comprising: generating a set of encoded quantization keys by entropy encoding the set of quantization keys;generating a set of decoded quantization keys by entropy decoding the set of encoded quantization keys; andwherein the set of dequantized numbers are generated based on the set of decoded quantization keys.

5. The method of claim 1, further comprising: generating the set of quantization keys using at least one of a block-wise quantization method, an individual quantization method, and a static quantization model method; andgenerating the set of dequantized numbers using at least one of a block-wise dequantization method, an individual dequantization method, and a static dequantization model method.

6. The method of claim 5, wherein a quantization method of the set of quantization keys is same as a dequantization method of the set of dequantized numbers; wherein based on the set of of quantization keys using the block-wise quantization method as the quantization method, the set of dequantized numbers use the block-wise dequantization method as the dequantization method;wherein based on the set of of quantization keys using the individual quantization method as the quantization method, the set of dequantized numbers use the individual dequantization method as the dequantization method; andwherein based on the set of of quantization keys using the static quantization model method as the quantization method, the set of dequantized numbers use the static dequantization model method as the dequantization method.

7. The method of claim 1, further comprising generating a set of current quantization states, based on the set of previous quantization states and the set of quantization keys, by training the third neural network, wherein the third neural network is trained by computing q-values for all possible actions, randomly selecting an action as an optimal action with an optimal q-value, generating a reward of the selected optimal action, sampling a set of selected optimal actions, and updating weight parameters of the third neural network to minimize distortion loss.

8. An apparatus for end-to-end task oriented latent image compression using deep reinforcement learning, the apparatus comprising: at least one memory configured to store program code; andat least one processor configured to read the program code and operate as instructed by the program code, the program code comprising: first generating code configured to cause the at least one processor to generate a plurality of latent representations of an input using a first neural network, wherein the plurality of latent representations comprise a sequence of latent signals;encoding code configured to cause the at least one processor to encode the plurality of latent representations using a second neural network;second generating code configured to cause the at least one processor to generate a set of quantization keys, using a third neural network, based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to the plurality of latent representations;third generating code configured to cause the at least one processor to generate a set of dequantized numbers representing dequantized representations of the encoded plurality of latent representations, based on the set of quantization keys, using a fourth neural network;decoding code configured to cause the at least one processor to decode a reconstructed output, based on the set of dequantized numbers; andperforming code configured to cause the at least one processor to perform a target task based on the reconstructed output using a fifth neural network.

9. The apparatus of claim 8, the program code further comprising computing code configured to cause the at least one processor to compute a task prediction loss based on the target task, wherein the first neural network and the fifth neural network are trained by back-propagating a gradient of the task prediction loss and updating weight parameters of the first neural network and the fifth neural network.

10. The apparatus of claim 8, wherein the target task is performed based on the generated plurality of latent representations.

11. The apparatus of claim 8, the program code further comprising: encoding key code configured to cause the at least one processor to generate a set of encoded quantization keys by entropy encoding the set of quantization keys;decoding key code configured to cause the at least one processor to generate a set of decoded quantization keys by entropy decoding the set of encoded quantization keys; andwherein the set of dequantized numbers are generated based on the set of decoded quantization keys.

12. The apparatus of claim 8, the program code further comprising: fourth generating code configured to cause the at least one processor to generate the set of quantization keys using at least one of a block-wise quantization method, an individual quantization method, and a static quantization model method; andfifth generating code configured to cause the at least one processor to generate the set of dequantized numbers using at least one of a block-wise dequantization method, an individual dequantization method, and a static dequantization model method.

13. The apparatus of claim 12, wherein a quantization method of the set of quantization keys is same as a dequantization method of the set of dequantized numbers; wherein based on the set of of quantization keys using the block-wise quantization method as the quantization method, the set of dequantized numbers use the block-wise dequantization method as the dequantization method;wherein based on the set of of quantization keys using the individual quantization method as the quantization method, the set of dequantized numbers use the individual dequantization method as the dequantization method; andwherein based on the set of of quantization keys using the static quantization model method as the quantization method, the set of dequantized numbers use the static dequantization model method as the dequantization method.

14. The apparatus of claim 8, further comprising state generating code configured to cause the at least one processor to generate a set of current quantization states, based on the set of previous quantization states and the set of quantization keys, by training the third neural network, wherein the third neural network is trained by computing q-values for all possible actions, randomly selecting an action as an optimal action with an optimal q-value, generating a reward of the selected optimal action, sampling a set of selected optimal actions, and updating weight parameters of the third neural network to minimize distortion loss.

15. A non-transitory computer-readable medium storing instructions for that, when executed by at least one processor for end-to-end task oriented latent image compression using deep reinforcement learning, cause the at least one processor to: generate a plurality of latent representations using a first neural network, wherein the plurality of latent representations comprise a sequence of latent signals;encode the plurality of latent representations using a second neural network;generate a set of quantization keys, using a third neural network, based on a set of previous quantization states, wherein each quantization key in the set of quantization keys and each previous quantization state in the set of previous quantization states correspond to the plurality of latent representations;generate a set of dequantized numbers representing dequantized representations of the encoded plurality of latent representations, based on the set of quantization keys, using a fourth neural network;decode a reconstructed output, based on the set of dequantized numbers; andperform a target task based on the reconstructed output using a fifth neural network.

16. The non-transitory computer-readable medium of claim 15, wherein the instructions, when executed by the at least one processor, further cause the at least one processor to compute a task prediction loss based on the target task, wherein the first neural network and the fifth neural network are trained by back-propagating a gradient of the task prediction loss and updating weight parameters of the first neural network and the fifth neural network.

17. The non-transitory computer-readable medium of claim 15, wherein the target task is performed based on the generated plurality of latent representations.

18. The non-transitory computer-readable medium of claim 15, wherein the instructions, when executed by the at least one processor, further cause the at least one processor to: generate a set of encoded quantization keys by entropy encoding the set of quantization keys;generate a set of decoded quantization keys by entropy decoding the set of encoded quantization keys; andwherein the set of dequantized numbers are generated based on the set of decoded quantization keys.

19. The non-transitory computer-readable medium of claim 15, wherein the instructions, when executed by the at least one processor, further cause the at least one processor to: generate the set of quantization keys using at least one of a block-wise quantization method, an individual quantization method, and a static quantization model method;generate the set of dequantized numbers using at least one of a block-wise dequantization method, an individual dequantization method, and a static dequantization model method; andwherein a quantization method of the set of quantization keys is same as a dequantization method of the set of dequantized numbers,wherein based on the set of of quantization keys using the block-wise quantization method as the quantization method, the set of dequantized numbers use the block-wise dequantization method as the dequantization method,wherein based on the set of of quantization keys using the individual quantization method as the quantization method, the set of dequantized numbers use the individual dequantization method as the dequantization method, andwherein based on the set of of quantization keys using the static quantization model method as the quantization method, the set of dequantized numbers use the static dequantization model method as the dequantization method.

20. The non-transitory computer-readable medium of claim 15, wherein the instructions, when executed by the at least one processor, further cause the at least one processor to generate a set of current quantization states, based on the set of previous quantization states and the set of quantization keys, by training the third neural network, wherein the third neural network is trained by computing q-values for all possible actions, randomly selecting an action as an optimal action with an optimal q-value, generating a reward of the selected optimal action, sampling a set of selected optimal actions, and updating weight parameters of the third neural network to minimize distortion loss.