Methods for determining and correcting noisy class labels

The combination of EM and DA algorithms addresses the challenge of noisy labels in neural networks by using a temperature-controlled free energy minimization to achieve smoother convergence and improved classification accuracy.

DE102023205014A1Pending Publication Date: 2024-12-05ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
DE102023205014
Authority / Receiving Office
DE · DE
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-30
Publication Date
2024-12-05

AI Technical Summary

Technical Problem

Neural networks face difficulties in classification tasks due to noisy observed labels, leading to local minima and stochasticity in parameter optimization, making it challenging to achieve global optima.

Method used

Combining the Expectation-Maximization (EM) algorithm with deterministic annealing (DA) to vary a temperature parameter in an outer loop, reformulating the log-likelihood maximization problem as a free energy minimization, and gradually reducing the temperature to track the global minimum.

Benefits of technology

This approach ensures smoother and more effective convergence to the global minimum, improving the accuracy of classification tasks by reducing dependence on initial model parameters and enhancing the robustness of neural networks to noisy labels.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

The disclosure relates to a method for determining a given hidden classification label z in the space of hidden classification labels Z using a predictive model P(z|x) for a given input x in the space of input data X based on a given observed noisy classification label y in the space of observed noisy classification labels Y using a probabilistic dependency P(y|z), wherein an expectation-maximization (EM) algorithm is applied to iteratively improve the predictive models P(z|x), P(y|z), wherein in an expectation (E) step a posterior distribution of the hidden classification labels z and the probabilistic dependency P(y|z) of the hidden classification labels z and the noisy classification labels y is determined, and in a maximization (M) step model parameter w of the predictive model P(z|x) and model parameter θ of the probabilistic dependency P(y|z) are determined. be adaptedwhere a Deterministic Annealing (DA) method is applied to selectively influence, in particular slow down, the rate of change of a loss function of the M-step by varying a temperature parameter 1 / β in an outer loop of the iterations of the E-step and the M-step.
Need to check novelty before this filing date? Find Prior Art

Description

State of the art

[0001] The availability of large datasets makes it possible to achieve excellent results with neural networks in classification tasks. However, the quality of the results is negatively affected by inaccurate class labels, which are noisy observed labels. In such cases, the correct label can be considered a hidden variable, and the noisy process can be considered a communication channel, a noisy channel, with unknown parameters.

[0002] The EM (Expectation-Maximization) algorithm is widely used for parameter estimation of statistical models with hidden variables. This algorithm provides an iterative procedure for obtaining approximate ML (Maximum Likelihood) estimates.

[0003] In EM algorithms in the context of noisy labels, in addition to observed and potentially noisy labels, a hidden variable is assumed that is not directly observed and corresponds to the true value of the label. The goal of the EM algorithm is to obtain an increasingly better predictive model for the hidden variable, i.e. the true value of the label, through an iteration of probabilistic inference of the hidden variable in the Expectation (E) step and changing the model parameters in the Maximization (M) step. If, for example, neural networks are used for the predictive model in the M-step, the optimization of the underlying parameters is usually very difficult in practice, which is essentially due to the many local minima of these networks and the inherent stochasticity of the corresponding learning algorithms. This is also described, for example, in J. Goldberger & E.Ben-Reuven, TRAINING DEEP NEURAL-NETWORKS USING A NOISE ADAPTATION LAYER, ICLR 2017 conference Paper. Disclosure of the invention

[0004] The present invention provides an improved solution to the problem described above with a method according to claim 1.

[0005] To solve the problem, a combination of an EM algorithm and a deterministic annealing (DA) method is proposed. Deterministic annealing results in a more "gentle" sequence of EM updates by varying a temperature parameter in an outer loop of the EM algorithm iterations, in the sense of a slower change in the loss function of the M-step. The DA method involves replacing a non-smooth, jagged cost function, whose minimization most likely results in a local minimum, with a cost function smoothed by high temperature, whose global minimum is comparatively easy to find. To find the global minimum, the temperature is slowly lowered, and the respective global minimum can be easily "tracked" because, with a slow temperature decrease, it is located not far from the global minimum of the previous temperature.

[0006] By combining DA methods with EM algorithms, the problem of maximizing the logarithmic likelihood is reformulated by minimizing the thermodynamic free energy, which is determined by the principle of maximum entropy and an analogy from statistical mechanics. The influence of model parameters is controlled by the temperature parameter, thus reducing the dependence on initial model parameters.

[0007] A temperature parameter in an outer loop of the iterations of the EM algorithm is initially set accordingly high and then slowly cooled down over the course of the iterations to find a global optimum.

[0008] The following procedure assumes that each input data x is assigned a true but hidden classification label z. For each input data x, a noisy classification label y is observed. Each noisy classification label y is a noisy version of each hidden classification label z.

[0009] The goal of the EM algorithm is to determine a set of model parameters such that the incomplete log-likelihood function is maximized: L(θ)=log∫p(y,z|θ)dz

[0010] The procedure of the EM algorithm consists in maximizing the auxiliary function, the so-called Q-function, over θ at each iteration: Q(θ,θ')=∫p(z|y,θ')logp(y,z|θ)dz

[0011] θ' is the parameter from the previous iteration, and θ is the parameter from the current iteration. p(z|y, θ) is the posterior probability distribution. The posterior probability distribution can be determined using Bayes' theorem as: p(z|y,θ)=p(y,z|θ)∑z'p(y,z'|θ)

[0012] The logarithmic probability for the training data is guaranteed to increase as the value of the Q-function is increased, namely Q(θ,θ')≥Q(θ,θ)⇒L(y|θ')≥L(y|θ).

[0013] Therefore, maximizing the value of the Q-function at each iteration maximizes the likelihood for the training data. The EM algorithm starts with an initial model parameter θ (0) and iterates between the following two steps: E-step Determining Q(θ,θ(k)) and determining the posterior probability distribution p(z|y, θ). M-step θ(k+1)=argmaxθQ(θ,θ(k)) where k denotes the number of iterations. According to the method, the log-likelihood is increased in each iteration or at least remains the same, whereby after inference in the E-step, the Q-function is equal to the (new) log-likelihood.

[0014] To apply the DA method to the EM algorithm, the problem of maximizing the log-likelihood function is reformulated as the problem of minimizing a free energy function: Fβ(θ)=−1βlog∫p(y,z|θ)βdz where 1 / β is the temperature. For β=1, the negative free energy −Fβ(θ) the log-likelihood function L(θ). To solve this minimization problem, a new posterior distribution is introduced using Jensen's inequality: Fβ(θ)=−1βlog∫ƒ(z|y,θ')p(y,z|θ) βƒ(z|y,θ')dz≤−1β∫ƒ(z|y,θ')logp(y,z|θ) βƒ(z|y,θ')dz=Uβ(θ,θ')−1βSβ(θ') where the term Uβ (θ, θ') is the negative Q-function where the posterior distribution p(y,z|θ) has been replaced by the function f(z|y,θ'). The term S β (θ') is the entropy of the function f.

[0015] Accordingly, the following applies: Uβ(θ,θ')=−∫ƒ(z|y,θ')logp(y,z|θ)dz Sβ(θ')=−∫ƒ(z|y,θ')logƒ(z|y,θ)dz

[0016] The upper limit of the equation Fβ(θ) The free energy of the Lagrangian corresponds in principle to the maximum entropy, and the parameter β is a Lagrange multiplier. In the DA method, the new posterior distribution f is derived to minimize the Lagrangian under the condition f fdz = 1. The optimal distribution can be derived as follows: ƒ(z|y,θ)=p(y,z|θ)β∑z'(y,z'|θ)β)

[0017] By inserting the derived posterior into the equation Fβ(θ) of the free energy, the upper limit coincides with the free energy, ie, Fβ(θ)=Uβ(θ,θ)−1βSβ(θ)

[0018] The free energy equation Fβ(θ) has the same form as the free energy in statistical physics. Therefore, the minimization of the free energy equation Fβ(θ) at a fixed temperature can be interpreted as an approach to thermodynamic equilibrium.

[0019] According to the invention, the temperature is gradually lowered in the EM algorithm, and the posterior distribution is deterministically optimized at each temperature. The inventive method is as follows: 1. The EM algorithm starts with an initial model parameter θ (0) , and an initial value β (0) 2. The EM algorithm is executed iteratively with a fixed value for β until the free energy equation Fβ(θ) converges: E-step: determine Uβ(θ,θ(k)) M-step: θ(k+1)=argmaxθU(θ,θ(k)) 3. Increasing the value for β 4. Steps 2 and 3 can be repeated as long as β < 1.

[0020] 1 / β (0) corresponds to an initial temperature of the DA process. 1 / β (0) is advantageously chosen so high that the EM steps of the EM algorithm form a single global minimum of the free energy equation Fβ(θ) At the initial temperature, the entropy S β (θ') and not the negative Q-function U β(θ, θ') are minimized, therefore the posterior f assumes a nearly uniform distribution. As the temperature decreases, the shape of f changes from the uniform distribution to the original posterior, and at the final temperature 1 / β = 1, the EM algorithm with DA methods is consistent with the original EM algorithm without DA methods. Similar to the EM algorithm, the EM algorithm with DA methods also converges at a fixed temperature due to the decrease in the free energy Fβ(θ) guaranteed. In the DA method, the negative free energy function slowly converges from a smooth, unimodal function to the original log-likelihood as the temperature decreases.

[0021] According to one embodiment of the method, the temperature parameter 1 / β is applied only to the predictive model P(y|z), in particular the noisy channel model. Typically, in the DA method, the temperature parameter 1 / β is applied to the combined probability P(y,z|x) = P(y|z)*P(z|x), i.e., P(y|z) β * P(z|x) β In this case, it is proposed that the temperature parameter 1 / β only applies to the predictive model P(y|z), i.e., the noise channel model, so that P(y|z) β * P(z|x) is obtained as a combined probability after normalization. This results in a tighter bound closer to the log-likelihood. This, in turn, enables effective pretraining of a neural network when applying the method to neural networks.

[0022] According to one embodiment of the method, in the posterior distribution of the hidden classification labels z, a maximum value of the posterior is set to 1 and the remaining values ​​of the posterior are set to zero for further calculation. It is therefore proposed to use not the exact posterior, but rather the so-called maximum posterior estimator, in which the probability for the maximum of the posterior is set to 1 and the remaining values ​​are set to zero. This can advantageously lead to better convergence. This is also described, for example, in "IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 53, NO. 10, OCTOBER 2005 1671 "The Expectation-Maximization Viterbi Algorithm for Blind Adaptive Channel Equalization."

[0023] According to one embodiment of the method, it is provided that in the M-step, the model parameters w of the predictive model P(z|x) and / or the model parameters θ of the probabilistic dependency P(y|z) are partially adapted. It is proposed to adapt not all parameters, but only a portion of the model parameters w of the predictive model P(z|x) and / or only a portion of the model parameters θ of the probabilistic dependency P(y|z). In this way, an improvement in the convergence properties of the EM algorithm can be achieved.

[0024] According to one embodiment of the method, it is provided that a further model is trained on the noisy classification labels, and this further model is used to adapt the classification labels.

[0025] Further embodiments of the method relate to the use of the method according to the described embodiments, wherein the predictive model is a classification algorithm, in particular a neural network.

[0026] According to one embodiment of the use, it can be provided that an initial neural network is trained on the observed noisy classification label y.

[0027] According to one embodiment of the use, it can be provided that in the M-step of the EM algorithm either the initial neural network is used as a basis, or wherein in the M-step of the EM algorithm the network trained in the previous iteration is used as a basis.

[0028] When used with a neural network, the method can be applied once the neural network has been trained on the noisy classification labels. In the case of other classification algorithms, such as a boosted tree algorithm, the method can be applied directly, i.e., on the noisy data.

[0029] In principle, the method can be applied to a wide variety of classification problems. It is particularly suitable for use with a neural network in the field of automatic image classification, particularly in the field of at least partially automated manufacturing processes or in the field of autonomous or semi-autonomous driving, and / or in the field of time series classification or regression problems.

[0030] Further advantages will become apparent from the description and the accompanying drawings. Exemplary embodiments of the invention are illustrated in the drawings and explained in more detail in the following description. Identical reference numerals in different figures designate identical or at least functionally comparable elements. When describing individual figures, reference may also be made to elements from other figures. These show, in schematic form: Fig. 1 a schematic representation of models for determining a respective hidden classification label z in the space of hidden classification labels Z by means of a predictive model P(z|x) for a respective input x in the space of input data X based on a respective observed noisy classification label y in the space of observed noisy classification labels Y by means of a probabilistic dependency P(y|z); Fig. 2 a schematic representation of steps of a method for determining a respective hidden classification label z in the space of hidden classification labels Z by means of a predictive model P(z|x) for a respective input x in the space of input data X based on a respective observed noisy classification label y in the space of observed noisy classification labels Y by means of a probabilistic dependency P(y|z) according to a first embodiment, and Fig.3 a schematic representation of steps of a method for determining a respective hidden classification label z in the space of hidden classification labels Z by means of a predictive model P(z|x) for a respective input x in the space of input data X based on a respective observed noisy classification label y in the space of observed noisy classification labels Y by means of a probabilistic dependency P(y|z) according to a further embodiment.

[0031] The method described below assumes that each input data x is assigned a true but hidden classification label z. For each input data x, a noisy classification label y is observed. Each noisy classification label y is a noisy version of each hidden classification label z.

[0032] Fig.1 shows a schematic view of a predictive model P(z|| x), wherein, by means of the predictive model P(z|x) and by means of a probabilistic dependency P(y|z), a classification label z in the space of hidden classification labels Z is determined for a respective input x in the space of input data X based on a respective observed noisy classification label y in the space of observed noisy classification labels Y. The predictive model is, for example, a classification algorithm, in particular a neural network.

[0033] The probabilistic dependency P(y|z) is modeled in the example as a communication channel, noisy channel, with model parameters θ(i,j) = p(y = j|z = i), and the probability of observing a noisy classification label y for a respective input x is given by p(y=j|x;w,θ)=∑i=1np(y=j|z=i;θ)p(z=i|x;w), where n is the number of classes.

[0034] The procedure involves applying an Expectation-Maximization (EM) algorithm to iteratively improve the predictive model P(z|x) and the noisy channel model P(y|z), cf. Fig. 2 and Fig. 3, where in the Fig. 2 and Fig. 3 different embodiments are shown.

[0035] Furthermore, a combination of an EM algorithm and a deterministic annealing (DA) method is proposed. Deterministic annealing leads to a more "gentle" sequence of EM updates by varying a temperature parameter in an outer loop of the EM algorithm iterations, in the sense of a slower change in the loss function of the M-step. The DA method involves replacing a non-smooth, jagged cost function, whose minimization most likely results in a local minimum, with a cost function smoothed by high temperature, whose global minimum can be found comparatively easily. To find the global minimum, the temperature is slowly lowered, and the respective global minimum can be easily "tracked" because, with a slow temperature decrease, it is located not far from the global minimum of the previous temperature.

[0036] By combining DA methods with EM algorithms, the problem of maximizing the logarithmic likelihood is reformulated by minimizing the thermodynamic free energy, which is determined by the principle of maximum entropy and an analogy from statistical mechanics. The influence of model parameters is controlled by the temperature parameter, thus reducing the dependence on initial model parameters.

[0037] A temperature parameter in an outer loop of the iterations of the EM algorithm is initially set accordingly high and then slowly cooled down over the course of the iterations to find a global optimum.

[0038] The goal of the EM algorithm is to determine a set of model parameters such that the incomplete log-likelihood function is maximized: L(θ)=log∫p(y,z|θ)dz

[0039] The procedure of the EM algorithm consists in calculating the auxiliary function, the so-called Q-function, over θ at each iteration, cf. Fig. 1, to maximize: Q(θ,θ')=∫p(z|y,θ')logp(y,z|θ) dz

[0040] θ' is the parameter from the previous iteration, and θ is the parameter from the current iteration. p(z|y, θ) is the posterior probability distribution. The posterior probability distribution can be determined using Bayes' theorem as: p(z|y,θ)=p(y,z|θ)∑z'p(y,z'|θ)

[0041] The logarithmic probability for the training data is guaranteed to increase as the value of the Q-function is increased, namely Q(θ,θ')≥Q(θ,θ)⇒L(y|θ')≥L(y|θ).

[0042] Therefore, maximizing the value of the Q-function at each iteration maximizes the likelihood for the training data. The EM algorithm starts with an initial model parameter θ (0) and iterates between the following two steps: E-step Determining Q(θ,θ(k)) and determining the posterior probability distribution p(z|y, θ). M-step θ(k+1)=argmaxθQ(θ,θ(k)) where k denotes the number of iterations. According to the method, the log-likelihood is increased in each iteration or at least remains the same, whereby after inference in the E-step, the Q-function is equal to the (new) log-likelihood.

[0043] To apply the DA method to the EM algorithm, the problem of maximizing the log-likelihood function is reformulated as the problem of minimizing a free energy function: Fβ(θ)=−1βlog∫p(y,z|θ)βdz where 1 / β is the temperature. For β=1, the negative free energy −Fβ(θ) the log-likelihood function L(θ). To solve this minimization problem, a new posterior distribution is introduced using Jensen's inequality: Fβ(θ)=−1βlog∫ƒ(z|y,θ')p(y,z|θ)βƒ(z|y,θ')dz≤−1β∫ƒ(z|y,θ')logp(y,z|θ)βƒ(z|y,θ')dz=Uβ(θ,θ')−1βSβ(θ') where the term U β (θ, θ') is the negative Q-function where the posterior distribution p(y,z|θ) has been replaced by the function f(z|y,θ'). The term S β (θ') is the entropy of the function f.

[0044] Accordingly, the following applies: Uβ(θ,θ')=−∫ƒ(z|y,θ')logp(y,z|θ) dz Sβ(θ')=−∫ƒ(z|y,θ')logƒ(z|y,θ) dz

[0045] The upper limit of the equation Fβ(θ) The free energy of the Lagrangian corresponds in principle to the maximum entropy, and the parameter β is a Lagrange multiplier. In the DA method, the new posterior distribution f is derived to minimize the Lagrangian under the condition f fdz = 1. The optimal distribution can be derived as follows: ƒ(z|y,θ)=p(y,z|θ)β∑z'(y,z'|θ)β)

[0046] By inserting the derived posterior into the equation Fβ(θ) of the free energy, the upper limit coincides with the free energy, ie, Fβ(θ)=Uβ(θ,θ)−1βSβ(θ')

[0047] The free energy equation Fβ(θ) has the same form as the free energy in statistical physics. Therefore, the minimization of the free energy equation Fβ(θ) at a fixed temperature can be interpreted as an approach to thermodynamic equilibrium.

[0048] According to the invention, the temperature is gradually lowered in the EM algorithm, and the posterior distribution is deterministically optimized at each temperature. The inventive method is as follows: 1. The EM algorithm starts with an initial model parameter θ (0) , and an initial value β (0) 2. The EM algorithm is executed iteratively with a fixed value for β until the free energy equation Fβ(θ) converges: E-step: determine U β (θ, θ (k) ) M-step: θ(k+1)=argmaxθU(θ,θ(k)) 3. Increasing the value for β 4. Steps 2 and 3 can be repeated as long as β < 1.

[0049] 1 / β (0) corresponds to an initial temperature of the DA process. 1 / β (0)is advantageously chosen so high that the EM steps of the EM algorithm form a single global minimum of the free energy equation Fβ(θ) At the initial temperature, the entropy S β (θ') and not the negative Q-function U β (θ, θ') are minimized, therefore the posterior f assumes a nearly uniform distribution. As the temperature decreases, the shape of f changes from the uniform distribution to the original posterior, and at the final temperature 1 / β = 1, the EM algorithm with DA methods is consistent with the original EM algorithm without DA methods. Similar to the EM algorithm, the EM algorithm with DA methods also converges at a fixed temperature due to the decrease in the free energy Fβ(θ) guaranteed. In the DA method, the negative free energy function slowly converges from a smooth, unimodal function to the original log-likelihood as the temperature decreases.

[0050] According to one embodiment of the method, the temperature parameter 1 / β is applied only to the predictive model P(y|z), in particular the noisy channel model. Typically, in the DA method, the temperature parameter 1 / β is applied to the combined probability P(y,z|x) = P(y|z)*P(z|x), i.e., P(y|z) β * P(z|x) β In this case, it is proposed that the temperature parameter 1 / β only applies to the predictive model P(y|z), i.e., the noise channel model, so that P(y|z) β* P(z|x) is obtained as a combined probability after normalization. This results in a tighter bound closer to the log-likelihood. This, in turn, enables effective pretraining of a neural network when applying the method to neural networks.

[0051] According to one embodiment of the method, in the posterior distribution of the hidden classification labels z, a maximum value of the posterior is set to 1 and the remaining values ​​of the posterior are set to zero for further calculation. It is therefore proposed to use not the exact posterior, but rather the so-called maximum posterior estimator, in which the probability for the maximum of the posterior is set to 1 and the remaining values ​​are set to zero. This can advantageously lead to better convergence.

[0052] According to one embodiment of the method, it is provided that in the M-step, the model parameters w of the predictive model P(z|x) and / or the model parameters θ of the probabilistic dependency P(y|z) are partially adapted. It is proposed to adapt not all parameters, but only a portion of the model parameters w of the predictive model P(z|x) and / or only a portion of the model parameters θ of the probabilistic dependency P(y|z). In this way, an improvement in the convergence properties of the EM algorithm can be achieved.

[0053] According to one embodiment of the method, it is provided that a further model is trained on the noisy classification labels, and this further model is used to adapt the classification labels.

[0054] According to one embodiment of the use, it can be provided that an initial neural network N_init is trained on the observed noisy classification label y, cf. Fig. 3.

[0055] According to one embodiment of the use, it can be provided that in the M-step of the EM algorithm either the initial neural network is used as a basis, cf. Fig. 3, or where in the M-step of the EM algorithm the network N_iter trained in the previous iteration is used as a basis, cf. Fig. 2.

[0056] When used with a neural network, the method can be applied once the neural network has been trained on the noisy classification labels. In the case of other classification algorithms, such as a boosted tree algorithm, the method can be applied directly, i.e., on the noisy data.

[0057] In principle, the method can be applied to a variety of classification problems. It is particularly suitable for application of the method, especially with a neural network, in the field of automatic image classification, especially in the field of at least partially automated manufacturing processes or in the field of autonomous or semi-autonomous driving, and / or in the field of time series classification problems or regression problems. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited non-patent literature

[0000] J. Goldberger & E. Ben-Reuven, TRAINING DEEP NEURAL-NETWORKS USING A NOISE ADAPTATION LAYER, ICLR 2017 conference Paper

[0003] IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 53, NO. 10, OCTOBER 2005 1671 „The Expectation-Maximization Viterbi Algorithm for Blind Adaptive Channel Equalization

[0022]

Claims

[1] Method for determining a respective hidden classification label z in the space of hidden classification labels Z by means of a predictive model P(z|x) for a respective input x in the space of input data X based on a respective observed noisy classification label y in the space of observed noisy classification labels Y by means of a probabilistic dependency P(y|z), wherein an Expectation-Maximization (EM) algorithm is applied for iteratively improving the predictive models P(z|x), P(y|z), wherein in an Expectation (E) step a posterior distribution of the hidden classification labels z and the probabilistic dependency P(y|z) of the hidden classification labels z and the noisy classification labels y are determined and in a Maximization (M) step model parameter w of the predictive model P(z|x) and model parameter θ of the probabilistic dependency P(y|z) can be adjusted,wherein a deterministic annealing (DA) method is applied to specifically influence, in particular to slow down, a rate of change of a loss function of the M-step by varying a temperature parameter 1 / β in an outer loop of the iterations of the E-step and the M-step. [2] Method according to claim 1, wherein the temperature parameter 1 / β is gradually decreased from β > 0 to β = 1 in each iteration, and the posterior distribution is determined deterministically in each iteration. [3] Method according to one of claims 1 or 2, wherein the probabilistic dependency P(y|z) is modeled as a communication channel, noisy channel, with model parameters θ(i,j) = p(y = j|z = i), and the probability of observing a noisy classification label y for a respective input x is given by p(y=j|x;w,θ)=∑i=1np(y=j|z=i;θ)p(z=i|x;w), where n is the number of classes. [4] Method according to one of the preceding claims, wherein the temperature parameter 1 / β is applied only to the predictive model P(z|x) or only to the predictive model P(y|z), in particular noisy channel. [5] Method according to one of the preceding claims, wherein in the posterior distribution of the hidden classification labels z a maximum value of the posterior is set to 1 and the remaining values ​​of the posterior are set to zero. [6] A method according to any one of the preceding claims, wherein a further model is trained on the noisy classification labels, and this further model is used to adapt the classification labels. [7] Method according to one of the preceding claims, wherein in the M-step the model parameters w of the predictive model P(z|x) and / or the model parameters θ of the probabilistic dependency P(y|z) are partially adapted. [8] Use of a method according to one of the preceding claims, wherein the predictive model is a classification algorithm, in particular a neural network or a boosted tree algorithm. [9] Application of the method according to one of the preceding claims, in particular with a neural network, in the field of automatic image classification, in particular in the field of at least partially automated manufacturing processes or in the field of autonomous or semi-autonomous driving, and / or in the field of time series classification problems or regression problems.