Ethernet® physical layer transceiver with nonlinear neural network equalizer

Abstract

A physical layer transceiver for connecting a host device to a wired channel medium includes a host interface for coupling to the host device, a line interface for coupling to the channel medium, a transmit path operably coupled to the host interface and the line interface, a receive path operably coupled to the line interface and the host interface, and an adaptive filter circuit operably coupled to at least one of the transmit path and the receive path for filtering signals on at least one of the transmit path and the receive path, the adaptive filter circuit including a nonlinear equalizer. The nonlinear equalizer may be a multilayer perceptron or a neural network equalizer based on a radial basis function, or may be a linear equalizer using a nonlinear activation function. The nonlinear equalizer may also have a front-end filter to reduce input complexity.

Description

[Technical Field]

[0001] [Cross-reference of related applications] This disclosure claims the interests of concurrently pending U.S. Provisional Patent Application No. 63 / 141,460, filed on 25 January 2021 by the same applicant, which is incorporated herein by reference in its entirety.

[0002] [Field of use] This disclosure relates to the use of nonlinear equalizers in physical layer transceivers. More specifically, this disclosure relates to the use of nonlinear neural network equalizers in the transmit and receive paths of physical layer transceivers, such as Ethernet® physical layer transceivers, and also for use in cancellation echo, near-end crosstalk, and far-end crosstalk. [Background technology]

[0003] The background information provided herein is intended to provide a general context for this disclosure. The inventors’ research described herein is not expressly or implicitly considered prior art to the subject matter of this disclosure, insofar as such research is described in this background section, as is the case with any description that would not be considered prior art at the time of filing.

[0004] Many integrated circuit devices, particularly "systems on a chip" (SoCs), include high-speed serial links between various device components (e.g., individual silicon dies in the SoC). Typical high-speed serial links of this type, commonly known as "SERDES" (serializers / deserializers), can suffer from significant nonlinearity or channel interference in the signal path, such as insertion loss, intersymbol interference (ISI), and nonlinearity such as scattering loss in the optical system, or crosstalk and jitter in the copper (i.e., wiring) system. To attempt to mitigate such channel interference, various forms of linear equalization are typically used at the receiver end of such links. However, linear equalization may not be sufficient to compensate for such nonlinearity, especially when the signal levels to be distinguished (e.g., voltage levels) in the data signal are close to each other and the signal-to-noise ratio (SNR) is low. [Overview of the Initiative]

[0005] According to an implementation of the subject matter of this disclosure, a physical layer transceiver for connecting a host device to a wired channel medium includes a host interface for coupling to the host device, a line interface for coupling to the wired channel medium, a transmit path operably coupled to the host interface and the line interface for encoding host data and driving the encoded host data to the wired channel medium, a receive path operably coupled to the line interface and the host interface for decoding data received from the wired channel medium and passing the decoded data to the host interface, and an adaptive filter circuit operably coupled to the transmit path and the at least one of the receive path for filtering signals on the transmit path and the receive path, the adaptive filter circuit including a nonlinear equalizer.

[0006] In a first implementation of such a physical layer transceiver, the adaptive filter circuit may include a non-linear equalizer inline in the transmission path and be configured to equalize the transmission signal.

[0007] In a second implementation of such a physical layer transceiver, the adaptive filter circuit may include a non-linear equalizer inline in the reception path and be configured to equalize the reception signal.

[0008] In a third implementation of the subject matter of the present disclosure, the adaptive filter circuit may include a non-linear echo cancellation circuit coupled to both the transmission path and the reception path and be configured to cancel echoes between the transmission path and the reception path.

[0009] According to a first aspect of the third implementation, the adaptive filter circuit may include a non-linear echo cancellation circuit operating in the analog region of the physical layer transceiver.

[0010] According to a second aspect of the third implementation, the adaptive filter circuit may include a non-linear echo cancellation circuit operating in the digital region of the physical layer transceiver.

[0011] According to a fourth aspect of the third implementation, the adaptive filter circuit may include a non-linear crosstalk cancellation circuit coupled to both the transmission path and the reception path to cancel at least one of (a) near-end crosstalk and (b) far-end crosstalk between the transmission path and the reception path.

[0012] A fourth implementation of such a physical layer transceiver may further include an adaptation circuit configured to compare the output of the adaptive filter circuit with known data and adapt the adaptive filter circuit based on a cost function to reduce an error in the output for subsequent iterations.

[0013] In a fifth implementation of such a physical layer transceiver, the adaptation circuit may be configured to adapt the adaptive filter circuit based on the cross entropy between each bit and the log-likelihood ratio corresponding to each bit.

[0014] In a sixth implementation of such a physical layer transceiver, the non-linear equalizer may include a neural network equalizer.

[0015] According to a first aspect of the sixth implementation, the neural network equalizer may include a multi-layer perceptron neural network equalizer.

[0016] According to a second aspect of the sixth implementation, the neural network equalizer may include a radial basis function neural network equalizer.

[0017] According to a third aspect of the sixth implementation, the neural network equalizer may be a low-complexity neural network equalizer including a front-end filter having a first number of inputs and a second number of outputs, where the second number is smaller than the first number, and a neural network filter having the output of the front-end filter as an input.

[0018] In a first instance of the third aspect of the sixth implementation, the front-end filter of the low-complexity neural network equalizer may include a finite impulse response filter to reduce the first number of inputs to the second number of inputs.

[0019] In a seventh implementation of such a physical layer transceiver, the non-linear equalizer may include a linear filter and a non-linear activation function.

[0020] According to a first aspect of the seventh implementation, the non-linear activation function may be a hyperbolic tangent function.

[0021] According to the first aspect of the seventh implementation, the nonlinear activation function may be a sigmoid function.

[0022] According to an implementation of the subject matter of this disclosure, a method for filtering interference in a physical layer transceiver for connecting a host device to a wired channel medium includes the steps of: performing a nonlinear equalization on at least one of the transmit path and the receive path to filter a signal on the transmit path and the receive path; and adapting the nonlinear equalizer based on the cross-entropy between the equalizer output and the data signal in the wired channel medium.

[0023] In a first implementation of such a method, the step of performing nonlinear equalization on at least one of the transmission path and the reception path may include the step of performing inline nonlinear equalization on the transmission path to equalize the transmission signals.

[0024] In a second implementation of such a method, the step of performing nonlinear equalization on at least one of the transmission path and the reception path may include the step of performing inline nonlinear equalization on the reception path to equalize the received signal.

[0025] In a third implementation of such a method, the step of performing nonlinear equalization may include the step of performing nonlinear echo cancellation between the transmission path and the reception path.

[0026] In a fourth implementation of such a method, the step of performing nonlinear equalization may include the step of performing nonlinear crosstalk cancellation to cancel at least one of (a) near-end crosstalk and (b) far-end crosstalk between the transmitting path and the receiving path.

[0027] In a fifth implementation of such a method, the step of performing nonlinear equalization may include the step of applying a nonlinear activation function and the step of performing linear filtering.

[0028] According to the first aspect of the fifth implementation, the step of applying a nonlinear activation function may include the step of applying a hyperbolic tangent function.

[0029] According to the second aspect of the fifth implementation, the step of applying a nonlinear activation function may include the step of applying a sigmoid function.

[0030] A sixth implementation of such a method may further include a step of applying initial filtering of the equalization inputs before the step of performing the nonlinear equalization in order to reduce complexity by reducing the number of inputs for the nonlinear equalization.

[0031] According to the first aspect of the sixth implementation, the step of applying initial filtering may include the step of applying finite impulse response filtering. [Brief explanation of the drawing]

[0032] Further features, properties, and various advantages of this disclosure will become apparent upon consideration of the following detailed description, which is to be interpreted in conjunction with the accompanying drawings, where similar reference numerals refer to the same parts throughout.

[0033] [Figure 1] This is a representation of a physical layer transceiver that may incorporate an implementation of the subject matter of this disclosure.

[0034] [Figure 2] This is a representation of a specific implementation of a physical layer transceiver incorporating the subject matter of this disclosure.

[0035] [Figure 3] This is a plot of the exclusive OR function in Cartesian coordinate space, illustrating the problem solved by the implementation of the subject matter of this disclosure.

[0036] [Figure 4] Figure 2 is a plot of the transformation of the exclusive OR function to different coordinate spaces, illustrating a solution based on the implementation of the subject matter of this disclosure.

[0037] [Figure 5] This is a diagram of a first implementation of a nonlinear equalizer that may be used in accordance with the subject matter of this disclosure.

[0038] [Figure 6] This is a diagram of a second implementation of a nonlinear equalizer incorporating the subject matter of this disclosure.

[0039] [Figure 7] This is a diagram of a third implementation of a nonlinear equalizer incorporating the subject matter of this disclosure.

[0040] [Figure 8] This is a diagram of a fourth implementation of a nonlinear equalizer incorporating the subject matter of this disclosure.

[0041] [Figure 9] This is a diagram of a fifth implementation of a nonlinear equalizer incorporating the subject matter of this disclosure.

[0042] [Figure 10] This is a diagram of a sixth implementation of a nonlinear equalizer incorporating the subject matter of this disclosure.

[0043] [Figure 11] A general implementation of a class of low-complexity nonlinear neural network filters in accordance with the subject matter of this disclosure is shown.

[0044] [Figure 12] Figure 11 shows a first implementation of a nonlinear equalizer in a class of low-complexity nonlinear neural network filters commonly shown.

[0045] [Figure 13] Figure 11 shows a second implementation of a nonlinear equalizer in a class of low-complexity nonlinear neural network filters commonly shown.

[0046] [Figure 14] Figure 11 shows a third implementation of a nonlinear equalizer in a class of low-complexity nonlinear neural network filters commonly represented.

[0047] [Figure 15] Figure 14 shows an alternative representation of the implementation of a low-complexity nonlinear neural network filter.

[0048] [Figure 16] Figure 11 shows a fourth implementation of a nonlinear equalizer in a class of low-complexity nonlinear neural network filters commonly shown.

[0049] [Figure 17] Figure 11 shows a fifth implementation of a nonlinear equalizer in a class of low-complexity nonlinear neural network filters commonly presented.

[0050] [Figure 18] Figure 17 is a graphic diagram of a nonlinear function that can be equalized using a low-complexity nonlinear neural network filter.

[0051] [Figure 19] Figure 11 shows a sixth implementation of a nonlinear equalizer in the class of low-complexity nonlinear neural network filters commonly shown.

[0052] [Figure 20] This is a flowchart illustrating a method for implementing the subject matter of this disclosure.

[0053] [Figure 21] This is a flowchart illustrating part of the method shown in Figure 20. [Modes for carrying out the invention]

[0054] As described above, integrated circuit devices may include high-speed SERDES links between various device components. Typical SERDES links may suffer from significant nonlinearity or channel faults in the signal path, such as insertion loss, intersymbol interference (ISI), and nonlinearity such as scattering loss in the optical system, or crosstalk and jitter in copper (i.e., wired) systems. To attempt to address such channel faults, various forms of linear equalization are typically used at the receiver end of such links.

[0055] However, especially in Ethernet physical layer transceivers (PHYs), linear equalization may not be sufficient to compensate for such nonlinearities, because the signal levels (e.g., voltage levels) to be distinguished in the data signal may be close to each other. For example, in contrast to typical non-zero return (NRZ) signaling which uses two levels to represent "0" and "1", SERDES in SoC devices may represent four possible 2-bit codes ("00", "01", "10", "11") using 4-level pulse amplitude modulation (PAM4) signaling which has four voltage levels but the maximum voltage swing is the same as NRZ signaling. Furthermore, Ethernet signaling may use even higher modulation such as 8-level pulse amplitude modulation (PAM8) or 16-level pulse amplitude modulation (PAM16) or even higher levels of pulse amplitude modulation. Thus, instead of one threshold within a voltage range dividing two signal levels, there may be 15 (or more) thresholds within a voltage range dividing as many as 16 (or more) signal levels. Linear equalization may not be sufficient to accurately assign received samples near the threshold between levels to the correct transmitted bits or codes when the thresholds are close together and the signal-to-noise ratio is low.

[0056] Furthermore, in Ethernet-type signaling, there can be many signal sources on the channel that contribute to various forms of interference, particularly echoes, near-end crosstalk, and far-end crosstalk.

[0057] In the implementation of the subject matter of this disclosure, nonlinear equalization is used to compensate for nonlinearity in the PHY channel, as well as to cancel echoes, near-end crosstalk, and far-end crosstalk, thereby reducing the bit error rate (BER). Different implementations may use different types of nonlinear equalizers.

[0058] Conceptually, a linear equalizer performs the separation of samples for assignment to one level or another by drawing substantially straight lines between groups of samples plotted in a two-dimensional (e.g., (x,y)) space. In channels with insufficient linearity, or where levels are too close to each other, there may be no straight lines that can be drawn between samples from different levels on such a plot. A nonlinear equalizer effectively remaps samples into different nonlinear (e.g., radial or polar) spaces where samples from different levels can be separated by straight lines or other smooth curves.

[0059] Nonlinear equalizers implemented in this disclosure may be somewhat complex. For example, a nonlinear equalizer may have more or fewer variables or taps, and the complexity is proportional to the number of variables. In addition, a nonlinear equalizer that operates at the bit level, i.e., separately for each bit of a code (e.g., 2 bits / code in PAM4 signaling) rather than for the entire code, may be less complex than a nonlinear equalizer that operates at the code level. In any case, greater complexity leads to better performance, all other considerations being equal. However, greater complexity may also require a larger device area and / or power consumption.

[0060] The types of nonlinear equalizers that may be used in accordance with the subject matter of this disclosure may include multilayer perceptron neural network (MLPNN) equalizers and low-complexity multilayer perceptron neural network (RC-MLPNN) equalizers, as well as radial basis function neural network (RBFNN) equalizers and low-complexity radial basis function neural network (RC-RBFNN) equalizers, as described in more detail below.

[0061] The performance of a nonlinear equalizer can be affected by the cost function used for the equalizer's adaptation. For example, according to the implementation of the subject matter of this disclosure, a nonlinear equalizer may use one of several different cost functions for adaptation, including either a least mean squared error (MMSE or MSE) cost function or a cross-entropy (CE) based cost function. While CE-based cost functions may yield better results than MMSE cost functions, they are more complex than MMSE cost functions.

[0062] Therefore, according to the implementation of the subject matter of this disclosure, the choice of which form of nonlinear equalizer to use and which cost function to use may be a trade-off between complexity (and thus cost) and performance.

[0063] The subject matter of this disclosure can be better understood by referring to Figures 1 through 21.

[0064] Figure 1 shows the structure of a physical layer transceiver 100 that may incorporate an implementation of the subject matter of this disclosure, in the context of a communication channel such as an Ethernet network channel. The physical layer transceiver 100 may include a transmitter path / channel 101 and a receiver path / channel 102 for data flowing between a host device 170 and a wired channel medium (e.g., a cable) 180. The host interface 171 connects the transmitter path / channel 101 and the receiver path / channel 102 to the host device 170, while the medium-dependent interface (MDI) 181 connects the transmitter path / channel 101 and the receiver path / channel 102 to the channel medium 180.

[0065] Figure 2 shows details of a specific implementation 190 of the physical layer transceiver 100 relating to the subject matter of this disclosure. In the transmitter path / channel 101 of the transceiver 190, an encoder 140, for example, a forward error correction (FEC) encoder, may be used to encode the transmit data bits 146 that are to be transmitted, and subsequently, a pulse shaping circuit 141 manipulates the time-domain characteristics of the transmit waveform so that the signal timing information can be easily extracted at the receiver. A transmit equalizer 142 may be provided to remove undesirable components or to recapture signal components that spread to adjacent codes before transmission. The equalized output is converted from digital to analog for transmission by a transmit digital-to-analog converter 143 and then driven onto the channel medium 103 by a transmit driver 144 via a hybrid coupler and transformer 145 (functioning as an MDI 181).

[0066] In the receiver path / channel 102 of the transceiver 190, data may be received from the channel medium 180 in the hybrid coupler and transformer 145 and transmitted to the analog front end 151 of the receiver 102, and then to the analog-to-digital converter (ADC) 152. The equalizer 153 may include one or more equalizers to remove interference. The output of the equalizer block 153 is sliced ​​in the slicer 154 and provided to a decoder 155, such as a forward error correction (FEC) decoder, which outputs the received data bits 156.

[0067] An analog echo canceller 161 may be provided in 112 between the analog domains of the transmission path 101 and the reception path 102. A digital echo canceller 162 may be provided in 122 between the digital domains of the transmission path 101 and the reception path 102. A crosstalk canceller 163 capable of filtering near-end crosstalk, far-end crosstalk, or both may also be provided in 122 between the digital domains of the transmission path 101 and the reception path 102.

[0068] According to the implementation of the subject matter of this disclosure, any one or more of the transmitter equalizer 142, receiver equalizer 153, analog echo canceller 161, digital echo canceller 162, and crosstalk canceller 163 may be based on a nonlinear filter, in particular a nonlinear neural network filter. Suitable nonlinear neural network filters are described in concurrently pending U.S. Patent Application No. 17 / 248,658, filed on 2 February 2021 by the same applicant, and concurrently pending U.S. Patent Application No. 17 / 648,831, filed concurrently herein, which are incorporated herein by reference in their entirety.

[0069] The adaptive function 164 can adapt various nonlinear equalizers 142, 153, 161, 162, and 163 by comparing the log-likelihood ratio 165 output by the equalizer 153 with the output data bits 156, or, during training mode, with the training bits 166.

[0070] The purpose of implementing equalization for a channel is to correct for the various interference sources mentioned above, thereby effectively moving samples that are on the wrong side of the threshold to the correct side. Linear equalization effectively plots the samples in two-dimensional (x,y) space and draws a straight line between the samples to indicate where the threshold should be. However, in channels with nonlinearity, there may not be a straight line that precisely separates the samples that can be drawn on the two-dimensional plot. In such cases, nonlinear equalization can be used. Nonlinear equalization can effectively remap the samples to a different space (e.g., with a different scale or coordinate system) where a straight line that precisely separates the samples exists.

[0071] Alternatively, a nonlinear equalizer may remap the samples to a space where some smooth curve exists other than a straight line that precisely divides the samples. For example, a nonlinear equalizer may remap the samples to polar or radial space in which the samples are grouped into circular or annular bands that can be divided by circles or ellipses.

[0072] The advantages of nonlinear equalization over linear equalization in nonlinear channels can be seen in the simplified examples shown in Figures 3 and 4, where the signals to be equalized are subjected to an exclusive OR (XOR or

number

number

[0073] However, radial basis functions

number

number

number

number

[0074] As will be discussed below, various types of nonlinear equalizers are available. The type of nonlinear equalizer used may be adaptable to accommodate changes in the channel state. Various forms of the cost function may be used to adapt in order to reduce errors in subsequent iterations.

[0075] One type of adaptive function that can be used is the least mean squared error (MMSE), where the mean squared error (MSE) is equal to the equalized signal (Y) and the ideal signal.

number

[0076] Another type of adaptive function that can be used is the cross-entropy (CE) between the training bit and its log-likelihood ratio (LLR). In particular, the cost function circuit may be configured to calculate a cross-entropy value that represents the difference between the probability distribution of the detected bit values ​​(a function of the LLR signal) and the probability distribution of the training bit values. The cost function circuit then adapts the equalizer by setting the equalizer parameters (e.g., one or more coefficients of the equalizer's filter taps) to values ​​corresponding to the minimum cross-entropy value among the calculated cross-entropy value and one or more previously calculated cross-entropy values, thereby reducing the bit error rate of the channel. As with MSE equalization, the equalizer may first be adapted in the training mode, where ideal signal values ​​are available. Later, during runtime operation, the detected output values ​​of the equalized channel should be close enough to the ideal values ​​to be used for adaptation. Specifically, if an arbitrary forward error correction code (FEC) decoder (e.g., a Reed-Solomon (RS) decoder or a low-density parity check (LDPC) decoder) is available after the equalizer, a successfully decoded frame from the output of the FEC decoder may be used for adaptation.

[0077] LLR can be defined as the relationship between the probability that a bit is "0" (P0) and the probability that a bit is "1" (P1).

number

[0078]

number

[0079] The gradient of cross-entropy with respect to LLR can be calculated by substituting P0 and P1 into the cross-entropy equation.

number

[0080] LLR is cross-entropy (i.e.,

number

number

[0081] A negative LLR means that bit=0 has a higher probability than bit=1, while a positive LLR means that bit=1 has a higher probability than bit=0. In these equations, P0 and P1 are probabilities and therefore positive values, and α is the adaptive bandwidth, which is also positive. Thus, when the true bit=0, an adaptation using cross-entropy will make the negative LLR more negative, and when the true bit=1, an adaptation using cross-entropy will make the positive LLR more positive. Therefore, a cross-entropy-based adaptation maximizes the magnitude of the LLR and is thus a maximum-likelihood adaptation, which reduces the BER. Thus, an equalizer adaptation to minimize cross-entropy also minimizes the BER.

[0082] Assuming there is a general computation graph from parameter X → Y → LLR → CE such that parameter X affects the value of output Y, which in turn affects LLR, and from which cross-entropy can be calculated, the cross-entropy gradient can be expressed with respect to other parameters:

number

[0083] One suitable implementation of a nonlinear filter that may be used in accordance with the subject matter of this disclosure is the nonlinear equalizer 401 shown in Figure 5. The nonlinear equalizer 401 is a multilayer perceptron neural network 402 that provides an equalized signal (Y) 411 from input digitized samples 421 that are delayed in 431 and combined in the multilayer perceptron neural network 402.

[0084] As shown in Figure 5, the multilayer perceptron neural network 441 includes at least one hidden layer 450 of hidden nodes 451. Although only one hidden layer 450 is shown in this figure, the multilayer perceptron neural network equalizer in the implementation of the subject matter of this disclosure may have multiple hidden layers (not shown). Similarly, while Figure 5 shows four hidden nodes 451 in the hidden layer 450, each hidden layer in the multilayer perceptron neural network equalizer in the implementation of the subject matter of this disclosure may have more or fewer hidden nodes 451, reflecting the number of parameters (filter tap coefficients).

[0085] Each hidden node 451 multiplies a delay sample 421 (shown as only one of the delays 431 being coupled to node 451 to avoid congestion in the diagram; however, each delay 431 is coupled to node 451) by a parameter (filter tap coefficient; not shown), and then sums the filter taps (Σ). Each hidden node 451 then applies a nonlinear activation function (e.g., hyperbolic tangent activation function) to its calculated sum.

number

[0086] The hidden node 451 may receive input not only from the feedforward delay 431 but also from the feedback delay 461, which represents the sample 460 of the previous code decision 412 that was fed back from the slicer 402, and this may help in mitigating intercode interference.

[0087] The aforementioned parameters of the nonlinear equalizer 401 are adapted based on the output Y. One method for adapting the parameters of the nonlinear equalizer 401 is to use an ideal sample derived from the training code 469.

number

[0088] Instead of the multilayer perceptron neural network 401, the implementation 500 (Figure 6) relating to the subject matter of this disclosure may include a low-complexity multilayer perceptron neural network 501 operating on an input digitized sample 521. The low-complexity multilayer perceptron neural network 501 includes two feedforward filters 542, 543, which may be, for example, finite impulse response (FIR) filters. The slicer 502 mitigates intersymbol interference from previous codes by providing an output decision Y 512 that is fed back via a decision feedback equalizer (DFE) 544 and combined with the output of the second feedforward (e.g., FIR) filter 543. The low-complexity multilayer perceptron neural network 501 includes a nonlinear activation function 545 (other nonlinear activation functions may be used, for example, a hyperbolic tangent activation function).

number

[0089] Similar to the case of the nonlinear equalizer 401, the parameters of the nonlinear equalizer 501 are adapted based on the output Y. One method for adapting the parameters of the nonlinear equalizer 501 is to use an ideal sample derived from the training code 569.

number

[0090] However, as mentioned above, cross-entropy can serve as a more effective cost function for adapting the parameters of the nonlinear equalizer to minimize BER.

[0091] Figure 7 shows an implementation 600 of a nonlinear equalizer relating to the subject matter of this disclosure. The nonlinear equalizer 601 takes four individual equalized signals (Y) from input digitized samples 621 that are delayed in 631 and combined in a multilayer perceptron neural network 641. ij This is a multilayer perceptron neural network 641 that provides ;i=0,1;j=0,1)611. The softmax function implemented in circuit 602:

number

[0092] As shown in Figure 5, the multilayer perceptron neural network 641 includes at least one hidden layer 650 of hidden nodes 651. Although only one hidden layer 650 is shown in this figure, a multilayer perceptron neural network equalizer in an implementation of the subject matter of this disclosure may have multiple hidden layers (not shown). Similarly, while Figure 7 shows four hidden nodes 651 in a hidden layer 650, each hidden layer in a multilayer perceptron neural network equalizer in an implementation of the subject matter of this disclosure may have more or fewer hidden nodes 651, reflecting the number of parameters (filter tap coefficients).

[0093] Each hidden node 651 multiplies a delay sample (shown as only one of the delays 631 is coupled to node 651 to avoid congestion in the diagram; however, each delay 631 is coupled to node 651) by a parameter (filter tap coefficient; not shown), and then sums the filter taps (Σ). Each hidden node 651 then applies a nonlinear activation function (e.g., hyperbolic tangent activation function) to its calculated sum.

number

[0094] Hidden node 651 receives input not only from feedforward delay 631 but also from feedback delay 661, which represents a sample of the previous code decision 660 that has been fed back, in order to mitigate intersymbol interference.

[0095] Since the equalizer 601 provides a soft output in the form of an LLR, the output can be used with a further external decoder (not shown), which may be a low-density parity check (LDPC) decoder or a forward error correction (FEC) decoder such as a Reed-Solomon decoder.

[0096] The aforementioned parameters of the nonlinear equalizer 601 are obtained by grouping the training bits 671 to form the training code (

number

number

[0097] Figure 8 shows an implementation 700 relating to the subject matter of this disclosure, which has a low-complexity multilayer perceptron neural network 741 coupled with a decision feedback equalizer 742, and takes an equalized signal (Y) 711 derived from an input digitized sample 721 as input to a code decision (sym) 744 and the log-likelihood ratio (LLR) of the code decision. sym )745 is the target code value

number

number

number

number

[0098] The low-complexity multilayer perceptron neural network 741 includes two feedforward filters 746, 747, which could be, for example, finite impulse response (FIR) filters. It also includes a nonlinear activation function 748 (e.g., a hyperbolic tangent activation function).

number

[0099] The parameters for feedforward filters 746 and 747 are the output log-likelihood ratio (LLR). sym )745 and the “true” code obtained from true bits which may be training bits, or which may be the output of a further external decoder (not shown) during runtime may be adapted to minimize the cross-entropy between them. The cross-entropy adaptive circuit 760 takes the output log-likelihood ratio (LLR) as input. sym ) has 745. In training mode, the cross-entropy adaptive circuit 760 also takes a known training bit 761 as input, which then acts as the "true" bit that is grouped to obtain the true sign. The cross-entropy adaptive circuit 760 has the training bit (

number

[0100] Since neural network equalizers can decorrelate the bits of multibit codes, such as 2-bit codes in PAM4 codes, a further implementation 800 relating to the subject matter of this disclosure may be provided (Figure 9). Implementation 800 includes an MLPNN equalizer 841 which is similar to the MLPNN equalizer 541 in that it includes at least one hidden layer 850 of hidden node 851, where the delayed samples in 831 (shown as only one of the delayed 831s coupled to node 851 to avoid congestion in the drawing; however, each delayed 831 is coupled to node 851) are multiplied by a parameter (filter tap coefficient; not shown), and then the filter taps are summed (Σ). Each hidden node 851 then applies a nonlinear activation function (e.g., hyperbolic tangent activation function) to its calculated sum.

number

[0101] The MLPNN841 differs from the MLPNN541 in that its final layer 852 includes two nodes 853 and 854. The inputs are not simply summed as in layer 552 of the MLPNN541, but after summing, a nonlinear activation function is also applied that decorrelates two bits of each sign, different from the nonlinear activation function of node 851, with each node 853 and 854 providing one of the two bits. The nonlinear activation function of each node 853 and 854 is, instead of a hyperbolic tangent activation function,

number

[0102] Node 853 provides a probability prediction 863 (p(bit msb )) for the most significant bit of two bits in the symbol, and node 854 provides a probability prediction 864 (p(bit lsb )) for the least significant bit of the two bits of the symbol. The two probability predictions 863, 864 are then compared with a threshold of 0.5 in slicer 855 to obtain a bit prediction (e.g., bit = 0 if p < 0.5 and bit = 1 if p ≧ 0.5) for each bit in the symbol.

[0103] In an implementation where the signaling includes more than four levels (e.g., PAM8 or PAM16), there will be more bits (e.g., 3 or 4 bits each) per symbol. In such a case, there will be a corresponding number of nodes, not just two nodes 853, 854.

[0104] At 856, the individual bits are regrouped into symbols again, then fed back at 857, and converted to corresponding voltages at 858 for input to feedback delay 861 (e.g., in a 4-level signaling system such as PAM4, -1 for "00", -1 / 3 for "01", +1 / 3 for "10", and +1 for "11"), representing samples of the previously determined symbol by the feedback for the next input from feedforward delay 831 to mitigate inter-symbol interference.

[0105] Since implementation 800 operates at the bit level instead of the symbol level, cross-entropy adaptation circuit 870 also operates at the bit level and determines the cross-entropy based on the individual bit-level probabilities 863, 864 and training bits 871, or, at runtime, on the output 890 of an external decoder (e.g., FEC decoder; not shown), etc.

[0106] At the bit level, cross-entropy can first be determined by judging the log-likelihood ratio from the probability prediction as described above. P0 is p(bit) msb=0 ) and P1 is p(bit msb=1 Starting from the most significant bit, LLR(bit msb ) can be calculated. Then LLR(bit msb ) and the most significant bit of the training bits or external decoder bits, CE(bit msb ) can be calculated. Next, p(bit lsb=0 Let P0 be p(bit) and P1 be p(bit) lsb=1 By using ), LLR(bit lsb ) can be calculated. Then LLR(bit lsb ), and the least significant bit of the training bits or external decoder bits, CE(bit lsb ) can be calculated. The bit-level cross-entropy is CE(bit msb )+CE(bit lsb This is the sum of ).

[0107] Figure 10 shows an implementation 900 relating to the subject matter of this disclosure, which includes a low-complexity multilayer perceptron neural network 941 capable of decorrelating the bits of a multibit code, such as 2 bits in a PAM4 code, coupled to each of the decision feedback equalizers 942, 952 for each bit.

[0108] The low-complexity multilayer perceptron neural network 941 includes a first feedforward filter 946, which may be, for example, a finite impulse response (FIR) filter. A nonlinear activation function 945 (e.g., a hyperbolic tangent activation function) is also included.

number

[0109] The nonlinear activation functions 961 and 962, which are different from the nonlinear activation function 945, produce their respective equalized bit outputs Y msb 944 and Y lsb This applies to 954. The nonlinear activation functions 961 and 962 are used instead of the hyperbolic tangent activation function.

number

[0110] The nonlinear activation function 961 is a probability prediction p(bit) for the most significant bit of the two bits in the code. msb ) provides a nonlinear activation function 962 that provides a probability prediction p(bit) for the least significant bit of the two bits of the sign. lsb ) is provided. Each of the two probability predictions is then compared to a threshold of 0.5 in the respective slicers 955, 956 to obtain a bit prediction for each bit in the code (e.g., bit=0 if p<0.5, and bit=1 if p≧0.5).

[0111] In 970, the two bits are grouped into code 971, and then in 972, they are converted to the corresponding voltages (e.g., -1 for "00", -1 / 3 for "01", +1 / 3 for "9", and +1 for "11") for input to the decision feedback equalizer 942 in the most significant bit path and for input to the decision feedback equalizer 952 in the least significant bit path. To mitigate intersymbol interference from previous codes, the outputs of each decision feedback equalizer 942, 952 are combined with the outputs of their respective feedforward filters 947, 957 in 943, 953, respectively, to produce the respective equalized bit output Y msb 944 and Y lsb 954, that is, the inputs to the nonlinear activation functions 961 and 962 as described above, are obtained.

[0112] Cross-entropy is calculated as follows, in the case of implementation 800: first, by determining the log-likelihood ratio from the probability prediction as described above, p(bit) msb ), p(bit msb ), and can be determined from the training bit 981 in the cross-entropy adaptive circuit 980 or the external decoder output 990. P0 is p(bit msb=0 ) and P1 is p(bit msb=1 Starting from the most significant bit, LLR(bit msb ) can be calculated. Then LLR(bit msb ) and the most significant bit of the training bits or external decoder bits, CE(bit msb ) can be calculated. Next, p(bit lsb=0 Let P0 be p(bit) and P1 be p(bit) lsb=1 By using ), LLR(bit lsb ) can be calculated. Then LLR(bit lsb ), and the least significant bit of the training bits or external decoder bits, CE(bit lsb ) can be calculated. The bit-level cross-entropy is CE(bit msb )+CE(bit lsb This is the sum of ).

[0113] Several implementations of nonlinear neural network filters that may be used in accordance with the subject matter of this disclosure, with reduced additional complexity, are shown in Figures 11 to 19.

[0114] Figure 11 shows a typical implementation 1000 of a low-complexity nonlinear neural network filter 1001 relating to the subject of this disclosure for equalizing two sets of inputs 111 and 121 from two signal sources on a wired medium 180 (this is merely an example, and there may be any number, i.e., one or more sets of input signals, as described below in relation to, for example, Figure 12). The low-complexity nonlinear neural network filter 1001 accepts inputs 111 and 121 of constant complexity, but reduces the complexity of inputs 111 and 121 by first filtering them through a front-end filter 1002 before filtering the low-complexity inputs 1011 and 1021 through a nonlinear filter circuit 1003. The reduction in the complexity of inputs 1011 and 1021 allows for a reduction in the complexity (measured by dimension) of the nonlinear filter circuit 1003, and therefore the complexity of the nonlinear neural network filter 1001 does not need to be reduced from the complexity of the filtered inputs 111 and 121.

[0115] The first implementation of the low-complexity nonlinear neural network filter 1100, shown in Figure 12, is based on the radial basis function nonlinear neural network filter 1101, along with a finite impulse response (FIR) based front-end filter 1102.

[0116] In the radial basis function nonlinear neural network filter 1101, digital samples from two inputs 1111 and 1121 are delayed by a delay line 1131 and combined in the radial basis function nonlinear neural network 1141. As seen in Figure 12, the radial basis function nonlinear neural network 1141 includes at least one hidden layer 1150 of hidden nodes 1151. Each hidden node 1151 operates on each delayed sample using a radial basis function, but to avoid congestion in the figure, it is shown as if only a portion of the delay in the delay line 1131 is coupled to each hidden node 1151. The outputs of the hidden layer 1150 are combined in 1152 (e.g., by addition) to provide the Y output 1103.

[0117] Each sample input in 1111 and 1121 adds a parameter or dimension to the radial basis function nonlinear neural network filter 1101, increasing the complexity of the filter. To reduce the complexity of the radial basis function nonlinear neural network filter 1101, a low-complexity nonlinear neural network filter 1100 includes a front-end filter 1102 that combines some of the inputs from the ADC outputs 111 and 121 to provide the reduced number of inputs 1111 and 1121 to the radial basis function nonlinear neural network filter 1101. As can be seen in Figure 12, in this implementation, the front-end filter 1102 uses FIR filtering (each line connecting delay 1112 to sum 1122 represents a multiplication of samples by coefficients (not shown) that form filter taps, and the taps are summed at 1122), for example, by combining every four input samples from the ADC outputs 111, 121 into one input sample 1111, 1121, thereby enabling a reduction in the complexity (measured by dimension) of the radial basis function nonlinear neural network filter 1101, and thus the complexity of the nonlinear neural network filter 1100 does not need to be reduced by the complexity of the filtered inputs 111, 121. The coefficients not visible may be parameters adapted using a backpropagation algorithm, which may be derived, for example, from the equation described above in relation to the cross-entropy gradient.

[0118] In the implementation shown in Figure 12, each set of input samples 111 and 121 is processed in separate parts of delay line 1112 and 1131. In this implementation, using two sets of input samples, each delay line is divided into two segments. However, more generally, the number of segments corresponds to the number of input sets. Therefore, in the case of a single input set, there will be only one segment (i.e., the delay line will not be segmented), but if there are three input sets, the delay line may be divided into three segments, and so on.

[0119] A second implementation of the low-complexity nonlinear neural network filter 1200, shown in Figure 13, is also based on the radial basis filter neural network filter 1201, along with a finite impulse response (FIR) based front-end filter 1202. As in the case of the front-end filter 1102, the front-end filter 1202 uses FIR filtering (each line connecting the delay 1212 to the radial basis function 1250 represents the multiplication of samples by coefficients (not shown; see the above description related to Figure 12) that form filter taps, for example, by combining every four input samples from the ADC outputs 111, 121 into one input sample 1211, 1221, thereby enabling a reduction in the complexity (measured by dimension) of the radial basis function nonlinear neural network filter 1201, and thus the complexity of the nonlinear neural network filter 1200 does not need to be reduced by the complexity of the filtered inputs 111, 121.

[0120] However, in this implementation, instead of being summed, the taps of delay line 1212 are directly input to the hidden node 1250 of the radial basis function nonlinear neural network filter stage 1201, and they are upstream of delay line 1231 in this implementation.

[0121] Again, using inputs 111 and 121 from the two sources, half of the delay line 1212 of the front-end filter 1202, 1213, is applied to input 111, while half of the delay line 1212 of the front-end filter 1202, 1214, is applied to input 121. Each input source 111 and 121 has one hidden node 1250 in the radial basis function nonlinear neural network filter stage 1201. The same applies to the delay line 1231 in the radial basis function nonlinear neural network filter stage 1201, which has separate halves 1232 and 1233 of the delay line 1231 applied to inputs 111 and 121, respectively. Here again, the delays 1231 form individual taps in the final FIR filter, which are combined in the summing node 1241 to obtain output Y.

[0122] A third implementation of the low-complexity nonlinear neural network filter 1300, shown in Figure 14, is based on a multilayer perceptron (MLP) nonlinear neural network filter 1302, along with a finite impulse response (FIR) based front-end filter 1301.

[0123] Typically, an MLP filter includes a delay line for the input samples, followed by at least one hidden layer where the samples are summed, and then, for example, a hyperbolic tangent function.

number

[0124] In the finite impulse response (FIR) based front-end filter 1301, the delay line 1331 is divided into a first portion 1332 that receives input 111 and a second portion 1333 that receives input 121. Each line connecting the delay 1312 to the sum 1322 represents a sample multiplication by coefficients (not shown; see the above description related to Figure 12) that form the FIR filter taps. The taps are summed by the sum portion of each hidden node 1350, which includes a sum function, and then in this implementation

number

[0125] In this implementation, the boundary between the front-end filter 1301 and the MLP nonlinear neural network filter 1302 operates through the hidden layer of the hidden node 1350, although this is not necessarily the case in all implementations.

[0126] In this implementation, the MLP nonlinear neural network filter 1302 is part of each of the one hidden nodes 1350

number

[0127] The low-complexity nonlinear neural network filter 1300 can be represented as the equivalent filter configuration 1400 shown in Figure 15. The low-complexity nonlinear neural network filter 1400 consists of four FIR filters 1401, 1402, 1403, and 1404, and two nonlinear activation functions 1405 and 1406 (each

number

[0128] FIR filters 1401 and 1402 form a finite impulse response (FIR) based front-end filter 1410, where FIR filter 1401 receives input 111 and FIR filter 1402 receives input 121. FIR filters 1403 and 1404 and nonlinear activation functions 1405 and 1406 form a low-complexity nonlinear neural network 1420. In the low-complexity nonlinear neural network 1420, activation function 1405 receives the output of FIR filter 1401 and, after nonlinear activation, passes those outputs to FIR filter 1403, while activation function 1406 receives the output of FIR filter 1402 and, after nonlinear activation, passes those outputs to FIR filter 1404. The outputs of FIR filters 1403 and 1404 are combined at an additive node 1408 to obtain output Y.

[0129] Another implementation of the low-complexity nonlinear neural network filter 1500 shown in Figure 16 is also based on a multilayer perceptron (MLP) nonlinear neural network filter 1502, along with a finite impulse response (FIR) based front-end filter 1501. In this implementation 1500, the finite impulse response (FIR) based front-end filter 1501 includes two FIR filters 1511 and 1521, each filtering the respective sets of inputs 111 and 121. The outputs of the FIR filters 1511 and 1521 are combined by an additive node 1531.

[0130] The output 1541 of the finite impulse response (FIR) based front-end filter 1501 is then filtered by a multilayer perceptron (MLP) nonlinear neural network filter 1502, which uses a nonlinear activation function 1512(

number

[0131] In the modified version 1600 of the low-complexity nonlinear neural network filter 1500 shown in Figure 17, a scalable bypass path 1601 is provided around the nonlinear neural network filter 1502. The scalable bypass path 1601 is controlled by a scaling factor g(1611). The FIR filter 1522 inherently includes similar scaling control. The provision of the scalable bypass path 1601 enables several operating modes. First, when g=0, the low-complexity nonlinear neural network filter 1600 operates identically to the low-complexity nonlinear neural network filter 1500. Second, by setting g=1 and setting the scaling factor of the FIR filter 1522 to 0, the low-complexity nonlinear neural network filter 1600 operates as a linear filter. This linear mode can be used as a "jump start" mode while the nonlinear portion of the filter is being applied.

[0132] In addition, the nonlinear function 1700 (in particular, one that is close to the linear function 1701) can be approximated as a series of linear functions 1702 with different slopes, as shown in Figure 18. By changing g to change the slope, the nonlinear function 1700 can be filtered using a linear finite impulse response (FIR) based front-end filter 1501, along with a nonlinear neural network filter 1502 that corrects for the difference between the segmented linear approximations and the actual nonlinear function.

[0133] A similar modification 1800 based on the low-complexity nonlinear neural network filter 1400 is shown in Figure 19. A scalable bypass path 1801 is provided around the nonlinear neural network filter 1420. The scalable bypass path 1801 is controlled by a scaling factor g(1811). The FIR filters 1403 and 1404 of the nonlinear neural network filter 1420 inherently include similar scaling control. By controlling g in 1811, the nonlinear neural network filter 1800 can operate in various modes in a manner similar to the nonlinear neural network filter 1600.

[0134] In each of the implementations shown, additional filter layers or stages may be added (not shown). For example, when the nonlinearity in the channel is severe, or when the interference length in the time domain is longer, then one or more nonlinear transformations may be required to separate the signals. Each nonlinear stage will transform its input into a different space at the output. After multiple transformations, the final space will result in the signal being able to be linearly separated. In the implementations of Figures 15 to 17 and 19, each additional layer may include an additional nonlinear activation function, followed by an additional FIR filter. Furthermore, as mentioned above in relation to the delay-line implementation, the delay line may be segmented into delay groups corresponding to different input groups. Similarly, in the activation function plus FIR filter implementation, an additional parallel set of activation functions and FIR filters may be provided, feeding into one or more common summing nodes.

[0135] It can be shown that the various implementations of the low-complexity nonlinear neural network filters described above, especially when adapted using cross-entropy, can provide performance nearly as good as that of non-low-complexity nonlinear neural network filters. However, the reduced complexity provides substantial savings in device area and power consumption.

[0136] A method 1900, which implements the subject matter of this disclosure, is illustrated in Figure 20.

[0137] Method 1900 begins in 1901, in which nonlinear equalization is performed on at least one of the transmit and receive paths in a physical layer transceiver for connecting a host device to a wired channel medium in order to filter the signal on at least one of the transmit and receive paths. In 1902, the nonlinear equalization is adapted based on the cross-entropy between the equalizer output and the data signal in the wired channel medium, and Method 1900 ends. However, as seen in 1903, optionally (indicated by dashed lines), initial filtering, such as finite impulse response (FIR) filtering, may be applied before the nonlinear equalization to reduce the complexity of the nonlinear equalization.

[0138] As can be seen in Figure 21, one implementation of performing nonlinear equalization in 1901 uses a nonlinear activation function, for example,

number

[0139] Therefore, it can be seen that a physical layer transceiver is provided that uses a nonlinear neural network equalizer for the transmission and / or reception paths, and / or for cancellation echo, near-end crosstalk, and far-end crosstalk.

[0140] As used herein and in the following claims, “one of A and B” shall be interpreted as “A or B.”

[0141] The above is merely an illustration of the principles of the present invention, and it should be noted that the present invention can be implemented in ways other than those described, which are presented for illustrative purposes only and not limiting purposes, and the present invention is limited only by the following claims.

Claims

1. A physical layer transceiver for connecting a host device to a wired channel medium, wherein the physical layer transceiver comprises: A host interface for connecting to the aforementioned host device; A line interface for coupling to the aforementioned wired channel medium; A transmission path operably coupled to the host interface and the line interface, including a circuit for encoding host data and driving the encoded host data to the wired channel medium; A receiving path, which includes a circuit operably coupled to the line interface and the host interface for decoding data received from the wired channel medium and passing the decoded data to the host interface; and An adaptive filter circuit operably coupled to at least one of the transmission and reception paths for filtering a signal on at least one of the transmission and reception paths, wherein the adaptive filter circuit includes a nonlinear equalizer inline on at least one of the transmission and reception paths and is configured to equalize at least one of the transmission and reception signals. A physical layer transceiver equipped with this feature.

2. The physical layer transceiver according to claim 1, wherein the adaptive filter circuit includes a nonlinear echo cancellation circuit coupled to both the transmit path and the receive path, and is configured to cancel echoes between the transmit path and the receive path.

3. The physical layer transceiver according to claim 2, wherein the adaptive filter circuit includes a nonlinear echo cancellation circuit operating in the analog region of the receiving path where the analog signal is transmitted.

4. The physical layer transceiver according to claim 2 or 3, wherein the adaptive filter circuit includes a nonlinear echo cancellation circuit operating in the digital region of the receiving path where a digital signal is transmitted.

5. The physical layer transceiver according to any one of claims 2 to 4, wherein the adaptive filter circuit includes a nonlinear crosstalk cancellation circuit coupled to both the transmit path and the receive path to cancel at least one of (a) near-end crosstalk and (b) far-end crosstalk between the transmit path and the receive path.

6. The physical layer transceiver according to any one of claims 1 to 5, further comprising an adaptive circuit configured to compare the output of the adaptive filter circuit with known data and to adapt the adaptive filter circuit based on a cost function to reduce errors in the output for subsequent iterations.

7. The physical layer transceiver according to claim 6, wherein the adaptive circuit is configured to adapt the adaptive filter circuit based on the cross-entropy between each bit and the log-likelihood ratio corresponding to each bit.

8. The physical layer transceiver according to any one of claims 1 to 7, wherein the nonlinear equalizer includes a neural network equalizer.

9. The physical layer transceiver according to claim 8, wherein the neural network equalizer includes a multilayer perceptron neural network equalizer.

10. The physical layer transceiver according to claim 8 or 9, wherein the neural network equalizer includes a radial basis function neural network equalizer.

11. The physical layer transceiver according to any one of claims 8 to 10, wherein the neural network equalizer is a low-complexity neural network equalizer comprising a front-end filter having a first number of inputs and a second number of outputs, wherein the second number is less than the first number, and a neural network filter having the output of the front-end filter as an input.

12. The physical layer transceiver according to claim 11, wherein the front-end filter of the low-complexity neural network equalizer includes a finite impulse response filter to reduce the first number of inputs to the second number of inputs.

13. The physical layer transceiver according to any one of claims 1 to 12, wherein the nonlinear equalizer includes a linear filter and a nonlinear activation function.

14. The physical layer transceiver according to claim 13, wherein the nonlinear activation function is a hyperbolic tangent function.

15. The physical layer transceiver according to claim 13, wherein the nonlinear activation function is a sigmoid function.

16. A physical layer transceiver for connecting a host device to a wired channel medium, wherein the physical layer transceiver comprises: A host interface for connecting to the aforementioned host device; A line interface for coupling to the aforementioned wired channel medium; A transmission path operably coupled to the host interface and the line interface, including a circuit for encoding host data and driving the encoded host data to the wired channel medium; A receiving path, operably coupled to the line interface and the host interface, includes a circuit for decoding data received from the wired channel medium and passing the decoded data to the host interface; Adaptive filter circuits operably coupled to at least one of the transmission and reception paths for filtering signals on at least one of the transmission and reception paths, the adaptive filter circuits including a nonlinear equalizer; and An adaptive circuit configured to compare the output of the adaptive filter circuit with known data and adapt the adaptive filter circuit based on the cross-entropy between each bit and the log-likelihood ratio corresponding to each bit. A physical layer transceiver equipped with this feature.

17. A method for filtering interference in a physical layer transceiver for connecting a host device to a wired channel medium, the method being: A step of performing nonlinear equalization on at least one of the transmission path and the reception path in order to filter the signal on at least one of the transmission path and the reception path; and Steps to adapt a nonlinear equalizer based on the cross-entropy between the equalizer output and the data signal in the wired channel medium. A method for filtering interference in a physical layer transceiver, comprising the following:

18. A method for filtering interference in a physical layer transceiver according to claim 17, wherein the step of performing nonlinear equalization on at least one of the transmission path and the reception path includes the step of performing inline nonlinear equalization on the transmission path to equalize the transmission signals.

19. A method for filtering interference in a physical layer transceiver according to claim 17 or 18, wherein the step of performing nonlinear equalization on at least one of the transmission path and the reception path includes the step of performing inline nonlinear equalization on the reception path to equalize the received signals.

20. A method for filtering interference in a physical layer transceiver according to any one of claims 17 to 19, wherein the step of performing nonlinear equalization includes the step of performing nonlinear echo cancellation between the transmission path and the reception path.

21. A method for filtering interference in a physical layer transceiver according to any one of claims 17 to 20, wherein the step of performing nonlinear equalization includes the step of performing nonlinear crosstalk cancellation to cancel at least one of (a) near-end crosstalk and (b) far-end crosstalk between the transmitting path and the receiving path.

22. A method for filtering interference in a physical layer transceiver according to any one of claims 17 to 21, wherein the step of performing nonlinear equalization includes the step of applying a nonlinear activation function and the step of performing linear filtering.

23. A method for filtering interference in a physical layer transceiver according to claim 22, wherein the step of applying a nonlinear activation function includes the step of applying a hyperbolic tangent function.

24. A method for filtering interference in a physical layer transceiver according to claim 22, wherein the step of applying a nonlinear activation function includes the step of applying a sigmoid function.

25. A method for filtering interference in a physical layer transceiver according to any one of claims 17 to 24, further comprising the step of applying initial filtering of the equalization inputs before the step of performing the nonlinear equalization in order to reduce complexity by reducing the number of inputs for the nonlinear equalization.

26. A method for filtering interference in a physical layer transceiver according to claim 25, wherein the step of applying initial filtering includes a step of applying finite impulse response filtering.

Images