Classification with Chained Components for Interpretable Machine Learning

Abstract

The present invention provides a computer-implemented method for creating an interpretable machine learning classification model, the method being implemented on one or more processors connected to a memory. According to an embodiment, the method includes the steps of defining the model by setting up a chained component classification (CBC) architecture (300) including several multiple chained component classification (CBC) blocks (310), training the model end-to-end with available input data (402) including forward propagating samples (401) of the input data (402) through the chained CBC architecture (300), and interpreting the output of each CBC block (310) of the chained CBC architecture (300) except for the last CBC block (310) as a likelihood vector of the concept detected in the respective CBC block (310). The present invention can be used for several anticipated medical / healthcare use cases, such as transparent disease classification by analyzing markers in blood tests and knowledge discovery for disease classification.

Description

[Technical Field]

[0001] The present invention relates to a computer-implemented method and system for creating interpretable machine learning classification models. [Background technology]

[0002] Embodiments of the present disclosure address the problem of learning deep, fully interpretable probabilistic classification models. Currently existing interpretable models, such as k-nearest neighbor classifiers or linear regression models, are shallow and therefore limited in their capabilities, leading to a limited complexity of problems they can efficiently solve.

[0003] To overcome this limitation, classification models deployed are often deep neural networks (DNNs). However, this comes at the cost of deploying models that are uninterpretable, as their design, considered a black box, makes it generally difficult to gain insight into their decision-making process. This becomes problematic in situations where there is interest in understanding classification decisions, for example, in high-stakes domains such as the medical domain. Thus, in short, the problem is that current machine learning models are either interpretable but with limited capabilities (shallow models) or uninterpretable but with more than limited capabilities (deep models). [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Sararajew et al., “Classification-by-Components: Probabilistic Modeling of Reasoning over a Set of Components,” NeurIPS 2019 [Non-patent document 2] Niepert et al., “Implicit MLE: Backpropagating Through Discrete Exponential Family Distributions,” NeurIPS 2021 Summary of the Invention [Problem to be solved by the invention]

[0005] It is therefore an object of the present invention to improve and further develop methods and systems of the type described above such that a deep neural network architecture is provided that is fully interpretable. [Means for solving the problem]

[0006] According to the present invention, the aforementioned objects are achieved by a computer-implemented method for creating an interpretable machine learning classification model, the method being implemented in one or more processors connected to a memory and comprising the steps of: defining the model by setting up a chained CBC architecture comprising classification CBC blocks with several chained components; training the model end-to-end on available input data comprising forward propagating examples of the input data through the chained CBC architecture; and interpreting the output of each CBC block of the chained CBC architecture, except for the last CBC block, as a possibility vector of the concepts detected in the respective CBC block.

[0007] Furthermore, the above-mentioned object is achieved by a system and a tangible, non-transitory computer-readable medium as defined in the independent claims.

[0008] The present invention first recognizes that the aforementioned problem can be solved by chaining component classification networks (chained CBC) to present a deep yet potentially interpretable architecture. According to an embodiment, at the heart of this architecture is a CBC block, which may be directly associated with a probability tree diagram, such that each parameter within the block has a probabilistic interpretation. According to a further embodiment, by interpreting the output of the CBC block as a probability value for the detected primitive and applying a binary gate, these CBC blocks can be chained together such that intermediate weights within the network are associated with probabilistic interpretations, making the network's inference process understandable. As a result, the model's capabilities are enhanced while simultaneously preserving interpretability.

[0009] Embodiments of the present disclosure provide a deep machine learning architecture for classification that is constrained to be fully interpretable. In particular, each parameter of the architecture is associated with a probabilistic meaning so that the model's internal reasoning process can be understood by experts. This makes the proposed architecture superior when machine learning methods must be applied to high-stakes decisions.

[0010] According to an embodiment, the CBC block has a two-stage approach: in the first stage, the input is analyzed and matching is performed on a probability measure; in the second stage, the probability of detected components is analyzed to find the reason for class assignment. Inference is made by calculating the probability of detection for each class.

[0011] According to an embodiment, each CBC block may include two stages: a first stage may be configured to determine a detection probability measure for each input based on an analysis of the respective input for matches with a given set of primitives, and a second stage may be configured to analyze the probability of a detected primitive from the set of primitives to infer about the current concept in the given input.

[0012] According to an embodiment, the probabilistic nature of the chained CBC architecture may be ensured by applying a binarization technique configured to binarize the likelihood vector between subsequent CBC blocks. Alternatively or additionally, the probabilistic nature of the chained CBC architecture may be ensured by applying a binarization technique to all detected primitives after the first CBC block of the chained CBC network. For example, the binarization technique may apply the Gumbel softmax concept or the Implicit Maximum Likelihood Estimation (I-MLE) framework.

[0013] According to an embodiment, it may be provided that the model is updated based on backpropagation of classification errors.

[0014] According to an embodiment, it may be provided that the trained model is used to predict the class of new data samples and to interpret the classification process by means of sample-based local explanations and / or model-based global explanations.

[0015] For example, according to an embodiment, a model may be trained with health-related patient data including at least one of blood pressure, temperature, and EEG and / or image data from a patient, and the trained model may be used to predict one or more classified diseases of the patient and obtain descriptions of the model's intermediate decision process for validation by a medical professional.

[0016] According to another embodiment, the model may be trained with blood markers from blood test results of patients with a diagnosed disease. After training, the intermediate representation of the model may be tested to generate knowledge about as yet unknown correlations between the blood markers and the disease.

[0017] According to yet another embodiment, the model may be trained on satellite imagery data of the Earth. After training, the trained model may be used to predict unknown indicators of a particular nature loss and the relationships between those indicators, and to obtain knowledge from the intermediate representation of the model regarding which past policies and / or actions were responsible for a given nature loss.

[0018] The present invention can be used for several anticipated medical / healthcare use cases, for example, transparent disease classification by analyzing markers in blood tests and knowledge discovery for disease classification.

[0019] There are several ways in which the teaching of the present invention can be advantageously designed and further developed. For this purpose, it is relied on, on the one hand, to the dependent claims and, on the other hand, to the following description of preferred embodiments of the invention, which are illustrated by way of example in the figures. In connection with the description of preferred embodiments of the invention with the aid of the figures, generally preferred embodiments and further developments of the teaching are described. In the figures: [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 1 is a schematic diagram showing a typical CBC structure according to the prior art. [Figure 2] FIG. 1 is a schematic diagram illustrating an exemplary implementation of the classification process of a CBC architecture in a digit classification task. [Figure 3] FIG. 1 is a schematic diagram illustrating a system of chained CBC networks according to an embodiment of the present invention. [Figure 4]FIG. 1 is a schematic diagram illustrating a method for producing fully interpretable machine learning methods according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0021] Currently, there are two main themes for addressing the interpretability of machine learning methods. The first theme is constructing post-hoc explanations. These explanations are often based on approximations or relaxations of the network after it has been trained. The problem with these explanations is that it is unclear how faithful they are (i.e., do they reflect the true inference process of the model). Methods that derive explanations as part of training can also be considered to fall into this category.

[0022] The second approach is models with built-in interpretability, such as k-nearest neighbors. Built-in interpretability means that by design / through constraints (rather than by adding regularization terms to the training loss), each parameter or subset of parameters has a clear meaning in the classification process that can be understood by a human expert. Furthermore, this also means that the interpretation of these parameters reflects the model's true inference process. Such models that are interpretable by design are preferable for high-stakes decisions. However, there are currently no available methods that can build deep, fully interpretable architectures so that complex classification problems can be addressed.

[0023] Embodiments of the present disclosure address this problem by presenting approaches for creating deep, fully interpretable architectures that are end-to-end trainable.

[0024] For example, a deep machine learning model, e.g., for classification learning, is fully interpretable if each parameter (or set of parameters) in the model is associated with a probability value of an event, such as, for example, "the probability that high blood pressure and fatigue should be observed due to diabetes." Embodiments of the present disclosure provide such models for supervised learning with respect to different types of data inputs based on an explicit probabilistic structure. In embodiments, the model uses a classification-by-components (CBC) network probabilistic structure, as disclosed in Saralajew et al., "Classification-by-Components: Probabilistic Modeling of Reasoning over a Set of Components," NeurIPS 2019, incorporated herein by reference in its entirety. CBC was introduced as a standalone network for classification consisting of two layers and predefined / trained components.

[0025] FIG. 1 provides an overview of a typical CBC structure 100 according to the prior art, while FIG. 2 shows an exemplary implementation of the classification process of the CBC architecture in a digit classification task.

[0026] The idea of ​​CBC is that the classification of an input 102 follows a two-stage approach. In the first stage, the input 102 is analyzed by using a set of (learned or predefined) components 112 from the storage 104. This analysis is performed by the detection element 106 searching for a match with the components 112 with respect to an appropriate probability measure (see the first part of Figure 1 and the left part of Figure 2). The corresponding detection probability measure of a given component 112 is high if the component 112 matches the input 102 (or part of it) and low otherwise.

[0027] After this first stage, in a second stage, which includes an inference probability storage 108, the probabilities of detected components 112 are analyzed to make inferences about the class of a given input 102 (see the second part of FIG. 1 and the right and center parts of FIG. 2). This inference is made by a probabilistic inference element 110 by calculating, for each class, the probability that each component 112 should have been detected and was detected in the input (indicated by matching a "1" in FIG. 2), or should not have been detected and was not detected in the input (indicated by matching a "0" in FIG. 2).

[0028] Furthermore, to avoid that each component 112 has to be considered for the calculation of the class probability, the model may also learn that components 112 are ignored (denoted by "x" in FIG. 2) for the decision process. Overall, the probability of a given class c is given by

[0029]

number

[0030] (see equation (1) on page 3 of the above-referenced paper by Saralajew et al.), where d(x) models the detection probability of component 112 and

[0031]

number

[0032] models the inference process.

[0033] A special feature of this model is that it is associated with a probability tree, and therefore the entire model follows a probabilistic model and is fully interpretable. Furthermore, the model is fully differentiable. As a result, all model parameters (i.e., component 112 and inference probabilities) can be learned through end-to-end training by maximizing the probability of the correct class. It should be emphasized that, thanks to the association with the probability tree, each parameter of the model can be interpreted and therefore has a clear meaning. This property can be used, for example, to interpret why the model is fooled by adversarial examples or what caused the model to predict a particular class.

[0034] For example, component κ k Detection probability d k For the calculation of (x), we use the rbf (radial basis function) kernel, i.e.,

[0035]

number

[0036] It may be given to consider d E is the Euclidean distance and σ is the temperature parameter.

[0037] Even though the model has better interpretability compared to standard neural networks, it is limited to the two-stage (two-layer) architecture described above in connection with Figures 1 and 2. It is not possible to simply chain these CBC blocks, i.e., to perform successive applications of these two layers, because after chaining, the output of each "probabilistic inference" block is the vector {0,1} C(also shown in Figure 1), since the stochastic nature is no longer given. Finally, note that all these operations can be extended to sliding operations so that convolution-like architectures can be built.

[0038] Embodiments of the present disclosure provide systems and methods that overcome this limitation and present an approach to chaining CBC networks so that higher-level components and inference concepts can be learned. In a sense, this can be considered equivalent to a neural network that learns more complex, higher-level features in deeper layers. The overall concept of a chained CBC network 300 according to embodiments of the present disclosure is shown in FIG. 3, where the network 300 includes n CBC blocks 310 (three of which are explicitly shown).

[0039] First, in contrast to the regular CBC described above, embodiments of the present disclosure consider CBC components as "primitives" to clearly distinguish them from the "components" used in regular CBC. According to embodiments, these primitives are learned during training and, together with the inference process, may encode increasingly complex concepts in the respective domains.

[0040] With regard to the terminology used herein, it should be noted that, essentially, in traditional CBC, components are combined to infer about classes, such as "we detected component A but not component B, and therefore the input sample belongs to class 1." In the context of the present disclosure, the situation may be (with respect to the first CBC block 3101 of the CBC network 300) "we detected primitives A and B, but not primitive C, and therefore the input sample constitutes concept 1." In this regard, the term "primitive" should be understood in the sense of a basic element (regardless of type) that has a level of complexity lower than that of a "concept." Thus, the term "concept" represents a higher-level complexity structure based on the combination of detected primitives.

[0041] Then, in the next CBC block of the CBC network 300, i.e., CBC block 3102, the detected concept output from the previous CBC block, i.e., "Concept 1," "Concept 2," etc., from CBC block 3101 in the above example, may be taken as input to CBC block 3102. This input is then compared with the primitives of CBC block 3102, thus performing template matching for concepts given the primitives. The primitives of CBC block 3102 encode which concepts should or should not be detected from the input. Based on the detection probabilities of CBC block 3102, a new higher-level concept is formed by an inference process, such as "Concept 1 is detected and Concept 2 is not detected, thus building the higher-level concept ALPHA." In the next block, in the simplest cast, the primitive could be the concept ALPHA (and the remaining notation, e.g., BETA, GAMMA, etc.). This technique is carried out iteratively, block by block, resulting in "concepts" that are potentially more and more complex.

[0042] As an image-level example, considering digit recognition, the above mechanism can be thought of as follows: First, primitives may be horizontal, vertical, and diagonal lines (multiples of 45 degrees). Through first inference, these primitives may be combined into concepts in the first block, such as a cross, a corner, a V, etc. Then, in the next CBC block, these concepts are considered as primitives, and matches are again attempted to be found, such that the first primitive is a cross with a V overlaid, etc. These detected higher-level primitives are then recombined by the next inference block.

[0043] Thus, the method may include detecting, by a CBC block of the CBC network, a particular combination of predefined primitives and deriving the existence of one or more concepts based on the detected combination of primitives. Further, the method may include using, by a subsequent CBC block of the CBC network, the derived one or more concepts as primitives for further recombination.

[0044] Second, in contrast to regular CBC, according to embodiments of the present disclosure, the intermediate CBC blocks 310 of the chained CBC network 300 iThe output of (i = 2, ..., n-1) is interpreted as a likelihood vector of the detected concept. According to an embodiment, to ensure the probabilistic nature of the CBC block 310, when the likelihood vector is fed to another CBC block 310, the likelihood vector or the primitive or both are binarized. For the binarization of the likelihood vector, a binary gate 320 may be implemented between subsequent CBC blocks 310 in the chain 300, as shown in FIG. 3. Alternatively or additionally, for the binarization of the primitive, a binary gate 330 may be implemented within each CBC block 310 (except the first CBC block 3101) between the storage 304 of the primitive and the detection element 306 configured to search the input 320 for a match with the component primitive. That is, according to an embodiment of the present disclosure, at least one of the two vectors is binarized before the calculation of the detection probability of the primitive is performed.

[0045] Without such a binary gate 320, x = κ k If , high detection probability d k Since (x) = 1 can be achieved for any vector, (the last block 310 of the chain 300) n It is recognized that the probabilistic properties are not preserved (except for p(x) = κ). That is, if the probability vector of the detected concept is equal to the primitive vector, e.g., p(x) = κ k = (0.5, 0.5, ..., 0.5), a high probability can be achieved. Consequently, this results in a high detection probability of the primitive, even if the probability of the detected concept reflects uncertainty. Binary gate 320 solves this problem, so that the force that generates the {0,1} probability vector is backpropagated through the network when the output probability is maximized, preserving the probabilistic nature of CBC block 310 in chained version 300. Precisely, with binary gate 320, a high output probability can be achieved only if all probabilities tend to be {0,1} vectors.

[0046] As shown in FIG. 3, the concatenated CBC network 300 may generate an output based on the class label vector, which may be interpreted as a decision signal (e.g., to apply either treatment A or treatment B).

[0047] According to embodiments of the present disclosure, binarization can be performed by applying Gumbel softmax or the concept of Implicit-MLE disclosed in Niepert et al., "Implicit MLE: Backpropagating Through Discrete Exponential Family Distributions," NeurIPS 2021, which is incorporated herein by reference in its entirety. The concept of Implicit-MLE is beneficial because it allows for the incorporation of additional constraints, such as the sparsity of {0,1} vectors. In principle, applying binarization to the probabilities of detected concepts or primitives is sufficient. However, the latter is beneficial because it prevents estimated gradients on binary gates from being backpropagated throughout the network, thus avoiding learning instability. Furthermore, binarizing primitives increases the comprehensibility of learned primitives, since only two states are possible per feature.

[0048] A key advantage of the chained CBC architecture 300 disclosed herein compared to traditional deep neural networks is that the chained CBC architecture 300 performs efficient feature engineering that is end-to-end trainable and fully interpretable via the chained CBC blocks 310. Thus, each parameter in the network is associated with a clear probabilistic meaning, making the entire feature engineering and classification process interpretable.

[0049] 4 is a schematic diagram illustrating a method for generating a fully interpretable machine learning method according to an embodiment of the present invention. As a requirement, as shown in S410, the method includes defining the architecture of the model (e.g., an architecture such as the CBC chain 300 shown in FIG. 3) by specifying how many CBC blocks and primitives / concepts per CBC block the model should consist of and chaining the respective number of CBC blocks accordingly. This step further includes specifying binary gates between the CBC blocks and selecting an appropriate discretization method to ensure the probabilistic nature of the model (e.g., by either binarizing the primitives or binarizing the possibility vectors).

[0050] Thereafter, as shown in S420, the method may include performing end-to-end training of the model on available data, for example data samples 401 drawn from a pool of training data 402.

[0051] End-to-end training of the model may involve forward propagation of data samples 401 through the chained CBC models, as shown in step S420a. During this forward propagation, the output of each CBC block of the chained CBC models, except for the last CBC block, is interpreted as a likelihood vector for the detected concept.

[0052] As shown in step S420b, in the context of end-to-end training of a model, it may be specified that a binarization technique is applied during forward calculation. This may be performed by binarizing all primitives after the first CBC block of a chained CBC model. Alternatively, a binarization technique may be applied between all CBC blocks except the last CBC block, binarizing each likelihood vector.

[0053] As shown in step S420c, in the context of end-to-end training of a model, it may be further defined that the error is back-propagated through the chained CBC models to update the model.

[0054] As shown in step S430, the method may include examining the trained model and the classification results. According to an embodiment, this step may include extracting explanations from the trained model by interpreting the intermediate representations (i.e., the output of the CBC blocks except for the last CBC block in the chain) as concept probabilities. According to a further embodiment, this step may also include using the trained model to predict the class of new data samples and to interpret the classification process by local explanations (sample-based) and / or global explanations (model-based).

[0055] As described above, the present disclosure provides a generalized method for learning higher-order inference through concepts that enables full traceability of deep machine learning methods. Embodiments of the present disclosure achieve this through one or more of the following aspects (see also Figures 3 and 4): 1. Chaining CBC blocks together by interpreting the output of each CBC block except the last one as a likelihood vector for the detected concept. 2. Ensuring the probabilistic nature of the model by applying binarization techniques between all CBC blocks except the last one, or for all primitives after the first one. 3. Stable end-to-end training with improved understandability of the provided explanations by only binarizing primitives.

[0056] The embodiments of the invention disclosed herein provide a significant step toward high-performance yet interpretable AI models, which is crucial whenever it is necessary to be able to assess whether an AI system complies with a particular regulation or rule. In this context, the embodiments of the invention are broadly applicable. In a broad sense, the embodiments of the present disclosure provide an important class of models that can be used in any application where building user confidence in the model's predictions is important, since the models are by design fully interpretable and therefore explainable. This may find important use, for example, as AI-augmented human decision support systems in areas such as public safety or public services, to name just a few.

[0057] The architecture proposed herein is generally designed for classification problems. However, it should be noted that although the model is interpretable by design with an associated probability tree, this does not guarantee that the explanation that can be created by visualizing the learned parameters is easily understandable to humans. It only means that the parameters and the associated explanation reflect the true inference process of the model. To ensure an explanation that is easily understandable to humans, the model may be further constrained or regularized.

[0058] Hereinafter, some exemplary use cases of application of the methods according to the embodiments disclosed herein are described in detail.

[0059] 1. Fully transparent disease classification A first exemplary use case is selected from the health sector and concerns patient monitoring and automatic detection of diseases in patients.

[0060] According to this use case, a model of concatenated CBC blocks may be established and trained with available data, as described herein. The available data may include health data related to a patient obtained from multiple data sources. For example, the training data may include patient data such as blood pressure, body temperature, EEG, etc. Alternatively or additionally, the training data may include image data, such as X-ray or ultrasound images.

[0061] The model may be trained to output a classified disease (of any kind), and there may be multiple hot classes. After the model is trained, each parameter in the network and the intermediate representation is fully interpretable and can be validated by experts before and optionally during deployment of the model.

[0062] According to an embodiment, the input of the chained model of the CBC block may come from a patient monitored in real time, and the output (classified disease) may include a properly prepared visualization of the network's intermediate decision process, which may be given to a doctor / physician, i.e., a (medical) expert. Instead of just considering the predicted disease, the expert may interact with the system by examining the visualization to understand which input features support the predicted disease and which features refute it. In particular, the expert may also understand which features support the prediction of another disease and how likely this disease is. This allows the expert to verify the prediction before a prescription is issued and to notice side effects if the prediction of another disease is also conclusive (i.e., when some features or a combination of features supports the prediction of another disease).

[0063] 2. Knowledge discovery for disease classification by analyzing blood test markers A second exemplary use case is also chosen from the health sector and concerns the identification of unknown indicators for specific diseases and their relationships.

[0064] According to this use case, a model of linked CBC blocks may be established and trained with available data, which may include blood test results (markers) of patients with diagnosed diseases, as described herein.

[0065] The model may be trained to output a disease. After the model is trained, each parameter in the network and the intermediate representation are fully interpretable and can be analyzed by experts. This allows for the identification of "hidden" patterns that indicate a particular disease. Given a particular disease, correlations between different blood markers can be identified.

[0066] According to embodiments, once the network has been trained and validated by experts (by inspecting the intermediate representation), it may be used to generate knowledge about unknown correlations between blood markers and diseases. This knowledge may be used to design more efficient blood testing machines by deploying the validated network to a blood testing machine, for example, so that when a blood sample is analyzed for a particular disease (input may come from an expert, e.g., suspecting a particular illness), the machine may automatically select blood markers to be analyzed for the sample to confirm or negate the disease based on the knowledge stored in the network. According to embodiments, this may be achieved in a continuous / active learning context. This automatic (minimal) blood marker selection allows the machine to perform disease verification faster than traditional methods that generate more comprehensive blood tests. Furthermore, by utilizing the generated knowledge, it is also possible to directly select markers that mask diseases indicated by similar markers (to prevent erroneous diagnoses). Instead of simply returning blood test results, the machine may also be configured to return the probability of the suspected disease, the probability of other (similar) diseases, and an explanation of each disease (i.e., why the analyzed features caused the prediction).

[0067] 3. Policy recommendations to prevent the loss of nature The third exemplary use case is selected from the technical field of achieving carbon neutrality and concerns the identification of unknown indicators of specific nature loss and their relationships.

[0068] According to this use case, a model of chained CBC blocks may be established and trained with available data, which may include, for example, satellite image data of the Earth or a particular region, as described herein.

[0069] The model may be trained to output correlations between specific indicators and their relationships with respect to specific nature loss. After the model is trained, each parameter in the network and the intermediate representations are fully interpretable and can be analyzed by experts. The intermediate representations may indicate which past policies or actions were responsible for nature loss (e.g., the extinction or decline of specific animal and / or plant species). For example, the intermediate representation may highlight that growing suburban areas led to the loss of healthy trees in nearby forests. Thus, the model may be trained to output recommendations regarding what policies / actions should be implemented for the area of ​​interest, such as how to ensure that forests do not decline.

[0070] According to an embodiment, a chained CBC network may be configured to output actionable information to either a computer system or a human operator. For example, a chained CBC network may recommend that a particular water system divert water from a different region.

[0071] As will be appreciated by those skilled in the art, further application scenarios can be envisioned in a variety of different technical fields.

[0072] Many modifications and other embodiments of the inventions described herein will come to mind to one skilled in the art to which these inventions pertain having the benefit of the teachings presented in the foregoing descriptions and the associated drawings. It is to be understood, therefore, that the inventions are not to be limited to the particular embodiments disclosed, and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terms have been employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation. [Explanation of symbols]

[0073] 100 Normal CBC structure 102 Input 104 Storage 106 Detection Elements 108 Storage 110 Probabilistic Reasoning Elements 112 Components 300 Chained CBC Network, Chain 304 Storage 306 Detection Elements 310 CBC blocks 3101 First CBC Block 310 i Intermediate CBC Block 310 n Last Block 320 Binary Gate 330 Binary Gate 401 Data Sample 402 training data

Claims

1. 1. A computer-implemented method for creating an interpretable machine learning classification model, the method being implemented in one or more processors coupled to a memory, the method comprising: defining the model by setting up a chained component classification (CBC) architecture (300) comprising several CBC blocks (310); training the model end-to-end with available input data (402), including forward propagating samples (401) of the input data (402) through the chained CBC architecture (300); interpreting the output of each CBC block (310) of said chained CBC architecture (300) except for the last CBC block (310) as a likelihood vector of the concept detected in each of said CBC blocks (310); A method comprising:

2. Each of the CBC blocks (310) a first stage in which a detection probability measure is determined for each input (102) based on an analysis of each input (102) for matches with a given set of primitives (112); a second stage in which the probabilities of the detected primitives of said set of primitives (112) are analyzed to make inferences about the current concept in the given input (102); 2. The method of claim 1, comprising:

3. 3. The method according to claim 1 or 2, wherein the probabilistic nature of the chained CBC architecture (300) is ensured by applying a binarization technique between subsequent CBC blocks (310).

4. The probabilistic nature of the chained CBC architecture (300) is such that the first CBC block (310) of the chained CBC network (300) 1 4. The method according to claim 1, wherein the thresholding is ensured by applying a binarization technique to all detected primitives after

5. The method according to claim 3 or 4, wherein the binarization technique applies the Gumbel softmax concept or the Implicit Maximum Likelihood Estimation (I-MLE) framework.

6. 6. The method of claim 1, further comprising updating the model based on backpropagation of classification error.

7. 7. The method of claim 1, further comprising using the trained model to predict classes of new data samples and to interpret the classification process by sample-based local explanations and / or model-based global explanations.

8. training the model with health-related patient data, including at least one of blood pressure, temperature, and EEG, and / or image data from a patient; using the trained model to predict one or more classified diseases of a patient and obtain a description of the intermediate decision process of said model for validation by a medical professional; 8. The method of claim 1, further comprising:

9. training the model with blood markers from blood test results of patients with a diagnosed disease; examining the intermediate representation of the model to generate knowledge about as yet unknown correlations between blood markers and diseases; 8. The method of claim 1, further comprising:

10. training said model with satellite imagery data of the Earth; using the trained model to predict unknown indicators and relationships of said indicators for a particular nature loss and to obtain knowledge from the intermediate representation of said model regarding which past policies and / or actions were responsible for a given nature loss; 8. The method of claim 1, further comprising:

11. In particular for the implementation of the method according to any one of claims 1 to 10, a system for creating an interpretable machine learning classification model, comprising, alone or in combination, the following steps: defining the model by setting up a chained component classification (CBC) architecture (300) comprising several CBC blocks (310); training the model end-to-end with available input data (402), including forward propagating samples (401) of input data (402) through the chained CBC architecture (300); and interpreting the output of each CBC block (310) of said chained CBC architecture (300) except for the last CBC block (310) as a likelihood vector of the concept detected in each of said CBC blocks (310); 1. A system including one or more processors configured to provide execution of:

12. Each of the CBC blocks (310) a first stage configured to determine a detection probability measure for each input (102) based on an analysis of each input (102) for matches with a given set of primitives (112); a second stage configured to analyze the probabilities of detected primitives of said set of primitives (112) in order to infer about the class of a given input (102); The system of claim 11 , comprising:

13. 13. The system of claim 11 or 12, further comprising a binary gate (320) implemented between subsequent CBC blocks (310) of the chained CBC architecture (300) and configured for binarizing the likelihood vector.

14. The first CBC block (310) of the chained CBC architecture (300) 1 14. The system of claim 11, further comprising a binary gate (330) implemented within each CBC block (310) except for the first CBC block (310) and configured for binarization of primitives.

15. A tangible, non-transitory computer-readable medium having instructions that, when executed by one or more processors, alone or in combination, provide for the performance of a method for creating an interpretable machine learning classification model, the method comprising: defining the model by setting up a chained component classification (CBC) architecture (300) including several CBC blocks (310); training the model end-to-end with available input data (402), including forward propagating samples (401) of input data (402) through the chained CBC architecture (300); and interpreting the output of each CBC block (310) of said chained CBC architecture (300) except for the last CBC block (310) as a likelihood vector of the concept detected in each of said CBC blocks (310); 1. A tangible, non-transitory computer-readable medium, including:

Images