US20210279570A1

US20210279570A1 - Method for proving or identifying counter-examples in neural network systems that process point cloud data

Info

Publication number: US20210279570A1
Application number: US17/078,079
Authority: US
Inventors: Michael A. Warren; Christopher Serrano; Aleksey Nogin
Original assignee: HRL Laboratories LLC
Current assignee: HRL Laboratories LLC
Priority date: 2020-03-03
Filing date: 2020-10-22
Publication date: 2021-09-09
Also published as: US20210278854A1; US11928585B2; EP4115348A1; EP4115349A1; WO2021178009A1; WO2021178011A1; CN115210721A; CN114981819A

Abstract

Described is a system for proving correctness properties of a neural network for providing estimates for point cloud data. The system receives as input a description of a neural network for generating estimates from a set of point cloud data. The description of the neural network is parsed to obtain a symbolic representation. Based on a combination of the symbolic representation and a set of analysis parameters, the system generates an analysis output indicating whether the neural network satisfies a correctness property in generating the estimates from the set of point cloud data. The analysis output is a mathematical proof artifact proving that the set of analysis parameters is satisfied, a list of one or more point clouds for which the set of analysis parameters is violated, or a report that progress could not be made by the analysis.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

This is a Non-Provisional Application of U.S. Provisional Patent Application No. 62/984,693, filed Mar. 3, 2020, entitled, “An Automatic Procedure for Either Proving or Identifying Counter-Examples to Correctness Properties for Neural Network Based Slope Estimation and Ground Plane Segmentation Systems that Process Point Cloud Data”, the entirety of which is incorporated herein by reference.
This is also a Non-Provisional Application of U.S. Provisional Patent Application No. 62/984,676, filed Mar. 3, 2020, entitled, “A Neural Network Architecture for Small Lidar Processing Networks for Slope Estimation and Ground Plane Segmentation”, the entirety of which is incorporated herein by reference.

GOVERNMENT LICENSE RIGHTS

This invention was made with government support under U.S. Government Contract Number FA8750-18-C-0092. The government may have certain rights in the invention.

BACKGROUND OF INVENTION

(1) Field of Invention

The present invention relates to a system for proving correctness properties of a neural network for providing estimates and, more particularly, to a system for proving correctness properties of a neural network for providing estimates based on point cloud data.

(2) Description of Related Art

Neural network based systems for providing estimates from point cloud data (e.g., of the kind generated by lidar or stereo camera systems) have become more prominent in the past several years (see, for instance, Literature References No. 7 and 9 of the List of Incorporated Literature References). At the same time, many questions have been raised regarding the reliability and robustness of neural network based systems (see Literature Reference Nos. 2 and 5), which makes it more pressing to be able to provide strong guarantees on the behavior of these kinds of systems. Recent work in analysis and verification of neural networks has focused almost exclusively on attempting to provide guarantees regarding what is called adversarial robustness, which is robustness in the face of small perturbations of inputs. In this context, the specification (i.e., definition of what it means for the system to be correct) considered is connected to so-called local adversarial robustness.
Katz et al. (see Literature Reference No. 8) and Wong and Kolter (see Literature Reference No. 11) describe work on verification of neural networks. Their work is only concerned with the perception domain with pixel image input and not with point cloud data. Additionally, local adversarial robustness is a weak correctness property for several reasons (e.g., dependence on a finite set of fixed inputs and a fixed norm). The notion of correctness embodied in local adversarial robustness is that the neural network does not change its output when the input is perturbed by a perturbation with bounded norm, which says nothing about the general correctness of the neural network.
The prior art is concerned exclusively with image (pixel) data and with the typical machine learning benchmark datasets for image based classification tasks, such as the MNIST handwritten digit dataset, where the goal is to classify images of handwritten digits. In the context of classification, the problem of defining correctness in a mathematically precise way is very different from the case of geometric tasks of the kind considered when working with point cloud data. For this reason, looking at stronger geometric correctness properties does not make sense for the applications being considered by the prior art in this area. Moreover, the satisfaction modulo theories (SMT) or mixed integer linear programing (MILP) based tools that have been employed in the prior art are not capable of reasoning about the trigonometric functions that occur prominently in analysis for the case of estimating surface normals.
Thus, a continuing need exists for a system for analysis of correctness properties of a neural network for providing estimates of surface normals for point cloud data.

SUMMARY OF INVENTION

The present invention relates to a system for proving correctness properties of a neural network for providing estimates for data and, more particularly, to a system for proving correctness properties of a neural network for providing estimates for point cloud data. The system comprises one or more processors and a non-transitory computer-readable medium having executable instructions encoded thereon such that when executed, the one or more processors perform multiple operations. The system receives, as input, a concrete description of a neural network for generating estimates from a set of point cloud data, wherein the concrete description of the neural network is in one of a source code format or a serialized format. The description of the neural network is parsed to obtain a symbolic representation. Based on a combination of the symbolic representation and a set of analysis parameters, an analysis output indicating whether the neural network satisfies a correctness property in generating the estimates from the set of point cloud data is generated. The analysis output is one of a mathematical proof artifact proving that the set of analysis parameters is satisfied, a list of one or more point clouds for which the set of analysis parameters is violated, and a report that progress could not be made by the analysis.
In another aspect, the analysis output acts as a stopping condition for preventing deployment of the neural network.
In another aspect, the estimates comprise an estimate of a slope of a patch of the set of point cloud data.
In another aspect, the estimates comprise an estimate of a surface normal of a patch of the set of point cloud data.
In another aspect, the estimates comprise an estimate regarding whether a patch of the set of point cloud data is ground or non-ground.
In another aspect, the set of analysis parameters describe geometric properties of the set of point cloud data.
Finally, the present invention also includes a computer readable program and a computer implemented method. The computer readable program includes computer-readable instructions stored on a non-transitory computer-readable medium that are executable by a computer having one or more processors, such that upon execution of the instructions, the one or more processors perform the operations listed herein. Alternatively, the computer implemented method includes an act of causing a computer to execute such instructions and perform the resulting operations.

BRIEF DESCRIPTION OF THE DRAWINGS

The objects, features and advantages of the present invention will be apparent from the following detailed descriptions of the various aspects of the invention in conjunction with reference to the following drawings, where:

FIG. 1 is a block diagram depicting the components of a system for proving correctness properties of a neural network for providing estimates for point cloud data according to some embodiments of the present disclosure;

FIG. 2 is an illustration of a computer readable program according to some embodiments of the present disclosure;

FIG. 3 is an illustration of a point cloud from an environment with box-shaped obstacles according to some embodiments of the present disclosure;

FIG. 4 is an illustration of neural network based lidar processing according to some embodiments of the present disclosure;

FIG. 5 is an illustration of modules in the system for analysis of correctness properties of a neural network for providing estimates for point cloud data. according to some embodiments of the present disclosure;

FIG. 6A is an illustration of a first neural network architecture for estimating surface normals of point cloud patches according to some embodiments of the present disclosure;

FIG. 6B is an illustration of a second neural network architecture for estimating surface normals of point cloud patches according to some embodiments of the present disclosure;

FIG. 7 is an illustration of a counter-example obtained for the case of surface normal according to some embodiments of the present disclosure;

FIG. 8 is an illustration of test points and disjoint sets F₁as y-balls according to some embodiments of the present disclosure; and

FIG. 9 is a flow diagram illustrating a system for analysis of correctness properties of a neural network for providing estimates for point cloud data. according to some embodiments of the present disclosure.

DETAILED DESCRIPTION

The present invention relates to a system for proving correctness properties of a neural network for providing estimates for data and, more particularly, to a system for proving correctness properties of a neural network for providing estimates from point cloud data. The following description is presented to enable one of ordinary skill in the art to make and use the invention and to incorporate it in the context of particular applications. Various modifications, as well as a variety of uses in different applications will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to a wide range of aspects. Thus, the present invention is not intended to be limited to the aspects presented, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
In the following detailed description, numerous specific details are set forth in order to provide a more thorough understanding of the present invention. However, it will be apparent to one skilled in the art that the present invention may be practiced without necessarily being limited to these specific details. In other instances, well-known structures and devices are shown in block diagram form, rather than in detail, in order to avoid obscuring the present invention.
The reader's attention is directed to all papers and documents which are filed concurrently with this specification and which are open to public inspection with this specification, and the contents of all such papers and documents are incorporated herein by reference. All the features disclosed in this specification, (including any accompanying claims, abstract, and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise. Thus, unless expressly stated otherwise, each feature disclosed is one example only of a generic series of equivalent or similar features.
Furthermore, any element in a claim that does not explicitly state “means for” performing a specified function, or “step for” performing a specific function, is not to be interpreted as a “means” or “step” clause as specified in 35 U.S.C. Section 112, Paragraph 6. In particular, the use of “step of” or “act of” in the claims herein is not intended to invoke the provisions of 35 U.S.C. 112, Paragraph 6.
Before describing the invention in detail, first a list of cited references is provided. Next, a description of the various principal aspects of the present invention is provided. Finally, specific details of various embodiment of the present invention are provided to give an understanding of the specific aspects.

(1) LIST OF INCORPORATED LITERATURE REFERENCES

The following references are cited and incorporated throughout this application. For clarity and convenience, the references are listed herein as a central resource for the reader. The following references are hereby incorporated by reference as though fully set forth herein. The references are cited in the application by referring to the corresponding literature reference number, as follows:

1. Barrett, C., Stump, A., & Tinelli, C. (2010). The SMT-LIB Standard—Version 2.0. SMT. Edinburgh.
2. Cao, Y., Xiao, C., Cyr, B., Zhou, Y., Park, W., Rampazzi, S., . . . Mao, Z. M. (2019). Adversarial Sensor Attack on LiDAR-based Perception in Autonomous Driving. ACM Conference on Computer and Communications Security.
3. Facebook. (2020, Feb. 19). PyTorch. Retrieved from https://pytorch.org.
4. Gao, S., Kong, S., & Clarke, E. M. (2013). dReal: An SMT solver for nonlinear theories over the reals. CADE.
5. Goodfellow, I., Shlens, J., & Szegedy, C. (2015). Explaining and harnessing adversarial examples. ICLR.
6. Google. (2020, Feb. 19). TensorFlow. Retrieved from https://tensorflow.org.
7. Guerrero, P., Kleiman, Y., Ovsjanikov, M., & Mitra, N. J. (2018). PCPNet: Learning local shape properties from raw point clouds. Computer Graphics Forum, 75-85.
8. Katz, G., Barrett, C. W., Dill, D. L., Julian, K., & Kochenderfer, M. J. (2017). Reluplex: An efficient SMT solver for verifying deep neural networks. CAV, (pp. 97-117).
9. Qi, C. R., Su, H., Mo, K., & Guibas, L. J. (2017). PointNet: Deep learning on point sets for 3D classification and segmentation. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
10. The Linux Foundation. (2020, Feb. 19). Open Neural Network Exchange (ONNX). Retrieved from https://onnx.ai.
11. Wong, E., & Kolter, J. Z. (2018). Provable defenses against adversarial examples via the convex outer adversarial polytope. ICML.

(2) PRINCIPAL ASPECTS

Various embodiments of the invention include three “principal” aspects. The first is a system for proving correctness properties of a neural network for providing estimates for point cloud data. The system is typically in the form of a computer system operating software or in the form of a “hard-coded” instruction set. This system may be incorporated into a wide variety of devices that provide different functionalities. The second principal aspect is a method, typically in the form of software, operated using a data processing system (computer). The third principal aspect is a computer readable program. The computer readable program generally represents computer-readable instructions stored on a non-transitory computer-readable medium such as an optical storage device, e.g., a compact disc (CD) or digital versatile disc (DVD), or a magnetic storage device such as a floppy disk or magnetic tape. Other, non-limiting examples of computer-readable media include hard disks, read-only memory (ROM), and flash-type memories. These aspects will be described in more detail below.
A block diagram depicting an example of a system (i.e., computer system 100) of the present invention is provided in FIG. 1. The computer system 100 is configured to perform calculations, processes, operations, and/or functions associated with a program or algorithm. In one aspect, certain processes and steps discussed herein are realized as a series of instructions (e.g., software program) that reside within one or more computer readable memory units and are executed by one or more processors of the computer system 100. When executed, the instructions cause the computer system 100 to perform specific actions and exhibit specific behavior, such as described herein.
The computer system 100 may include an address/data bus 102 that is configured to communicate information. Additionally, one or more data processing units, such as a processor 104 (or processors), are coupled with the address/data bus 102. The processor 104 is configured to process information and instructions. In an aspect, the processor 104 is a microprocessor. Alternatively, the processor 104 may be a different type of processor such as a parallel processor, application-specific integrated circuit (ASIC), programmable logic array (PLA), complex programmable logic device (CPLD), or a field programmable gate array (FPGA).
The computer system 100 is configured to utilize one or more data storage units. The computer system 100 may include a volatile memory unit 106 (e.g., random access memory (“RAM”), static RAM, dynamic RAM, etc.) coupled with the address/data bus 102, wherein a volatile memory unit 106 is configured to store information and instructions for the processor 104. The computer system 100 further may include a non-volatile memory unit 108 (e.g., read-only memory (“ROM”), programmable ROM (“PROM”), erasable programmable ROM (“EPROM”), electrically erasable programmable ROM “EEPROM”), flash memory, etc.) coupled with the address/data bus 102, wherein the non-volatile memory unit 108 is configured to store static information and instructions for the processor 104. Alternatively, the computer system 100 may execute instructions retrieved from an online data storage unit such as in “Cloud” computing. In an aspect, the computer system 100 also may include one or more interfaces, such as an interface 110, coupled with the address/data bus 102. The one or more interfaces are configured to enable the computer system 100 to interface with other electronic devices and computer systems. The communication interfaces implemented by the one or more interfaces may include wireline (e.g., serial cables, modems, network adaptors, etc.) and/or wireless (e.g., wireless modems, wireless network adaptors, etc.) communication technology. Further, one or more processors 104 (or devices, such as autonomous platforms) can be associated with one or more associated memories, where each associated memory is a non-transitory computer-readable medium. Each associated memory can be associated with a single processor 104 (or device), or a network of interacting processors 104 (or devices), such as a network of autonomous platforms (e.g., autonomous vehicles, robots).
In one aspect, the computer system 100 may include an input device 112 coupled with the address/data bus 102, wherein the input device 112 is configured to communicate information and command selections to the processor 104. In accordance with one aspect, the input device 112 is an alphanumeric input device, such as a keyboard, that may include alphanumeric and/or function keys. Alternatively, the input device 112 may be an input device other than an alphanumeric input device. In an aspect, the computer system 100 may include a cursor control device 114 coupled with the address/data bus 102, wherein the cursor control device 114 is configured to communicate user input information and/or command selections to the processor 104. In an aspect, the cursor control device 114 is implemented using a device such as a mouse, a track-ball, a track-pad, an optical tracking device, or a touch screen. The foregoing notwithstanding, in an aspect, the cursor control device 114 is directed and/or activated via input from the input device 112, such as in response to the use of special keys and key sequence commands associated with the input device 112. In an alternative aspect, the cursor control device 114 is configured to be directed or guided by voice commands.
In an aspect, the computer system 100 further may include one or more optional computer usable data storage devices, such as a storage device 116, coupled with the address/data bus 102. The storage device 116 is configured to store information and/or computer executable instructions. In one aspect, the storage device 116 is a storage device such as a magnetic or optical disk drive (e.g., hard disk drive (“HDD”), floppy diskette, compact disk read only memory (“CD-ROM”), digital versatile disk (“DVD”)). Pursuant to one aspect, a display device 118 is coupled with the address/data bus 102, wherein the display device 118 is configured to display video and/or graphics. In an aspect, the display device 118 may include a cathode ray tube (“CRT”), liquid crystal display (“LCD”), field emission display (“FED”), plasma display, or any other display device suitable for displaying video and/or graphic images and alphanumeric characters recognizable to a user.
The computer system 100 presented herein is an example computing environment in accordance with an aspect. However, the non-limiting example of the computer system 100 is not strictly limited to being a computer system. For example, an aspect provides that the computer system 100 represents a type of data processing analysis that may be used in accordance with various aspects described herein. Moreover, other computing systems may also be implemented. Indeed, the spirit and scope of the present technology is not limited to any single data processing environment. Thus, in an aspect, one or more operations of various aspects of the present technology are controlled or implemented using computer-executable instructions, such as program modules, being executed by a computer. In one implementation, such program modules include routines, programs, objects, components and/or data structures that are configured to perform particular tasks or implement particular abstract data types. In addition, an aspect provides that one or more aspects of the present technology are implemented by utilizing one or more distributed computing environments, such as where tasks are performed by remote processing devices that are linked through a communications network, or such as where various program modules are located in both local and remote computer-storage media including memory-storage devices.
An illustrative diagram of a computer readable program (i.e., storage device) embodying the present invention is depicted in FIG. 2. The computer readable program is depicted as floppy disk 200 or an optical disk 202 such as a CD or DVD. However, as mentioned previously, the computer readable program generally represents computer-readable instructions stored on any compatible non-transitory computer-readable medium. The term “instructions” as used with respect to this invention generally indicates a set of operations to be performed on a computer, and may represent pieces of a whole program or individual, separable, software modules. Non-limiting examples of “instruction” include computer program code (source or object code) and “hard-coded” electronics (i.e. computer operations coded into a computer chip). The “instruction” is stored on any non-transitory computer-readable medium, such as in the memory of a computer or on a floppy disk, a CD-ROM, and a flash drive. In either event, the instructions are encoded on a non-transitory computer-readable medium.

(3) SPECIFIC DETAILS OF VARIOUS EMBODIMENTS

Neural network based systems for providing estimates from point cloud data (e.g., of the kind generated by lidar or stereo camera systems) have become more prominent in the past several years (see Literature Reference Nos. 7 and 9). At the same time, many questions have been raised regarding the reliability and robustness of neural network based systems (see Literature Reference Nos. 2 and 5), which makes it more pressing to be able to provide strong guarantees on the behavior of these kinds of systems. Recent work in analysis and verification of neural networks has focused almost exclusively on the attempting to provide guarantees regarding what is called adversarial robustness: robustness in the face of small perturbations of inputs. FIG. 3 depicts an example of a point cloud from an environment with box-shaped obstacles. Light grey points (e.g., 300) represent ground plane, and dark grey points (e.g., 302) represent non-ground plane.
One specification, or definition of what it means for the system to be correct, considered in the literature is so-called local adversarial robustness. Explicitly, where X is the input space (e.g., the space of pixel images) and Y is the output space (e.g., the space of classifications such as “cat”, “dog”, etc.), the neural network estimator can itself be thought of as a map f: X→Y. In this context, local adversarial robustness states that, for a fixed finite batch of inputs B⊆X, there is image b E B and no perturbation π with norm ∥π∥<δ such that f(b)≠f(b+π). Several remarks regarding this definition are in order. First, the “local” in this definition refers to the fact that although this property quantifies over all possible perturbations bounded in norm by δ, the set of images over which the guarantee holds is merely finite. Second, to state this property one must first fix a norm ∥−∥, but the question of which norm is well suited to the actual data distribution is non-trivial. Thus, local adversarial robustness provides somewhat weaker guarantees than might be hoped.
The invention described herein is concerned with attempting to provide more meaningful guarantees than those of local adversarial robustness in the case where the input space X is a space of point cloud data (e.g., of the kind coming from lidar or stereo camera) and where the task of the neural network f is to make one of the following kinds of estimates: (1) to estimate the surface normal vector in three-dimensional Euclidean space; (2) to estimate whether or not the input points constitute a ground (roughly flat) or non-ground plane; and (3) to estimate other geometric properties (e.g., curvature) from point cloud data where a ground truth on certain representative data can be determined analytically. Task (1) itself can be used as part of the solution to (2), but these can also be separate tasks. An example of where these tasks are relevant is given by providing an estimate of where an autonomous ground vehicle can safely drive. For example, obstacles (such as pedestrians or buildings) would be tagged as non-ground plane. An observation is that, in these cases, there is a natural norm (the standard norm of three-dimensional Euclidean space) on the space X that is relevant to the properties being estimated. The present invention leverages this crucial observation.
Given the source code, mathematical description, or serialized format (e.g., TensorFlow or PyTorch saved model) of a neural network f that aims to address one of the tasks (1) or (2), the present invention is a method for automatically generating either (a) a mathematical proof that, for any point cloud x∈X of a certain kind, the neural network f will produce a correct estimate f(x); or (b) a potential counter-example x∈X such that f(x) is an incorrect estimate. A detailed description of the input parameters to the process described herein, the conditions/limitations on the point clouds to which the process applies, and the output of the process, a mathematical proof (which can be regarded as text data) or a concrete point-cloud, is provided below. The invention described herein provides a more thorough analysis of the correctness of a neural network for carrying out one of the tasks (1) or (2) above. Current state-of-the art approaches to this problem are based purely on extensive testing or else using adversarial machine learning. The disadvantages of these two approaches is that they only provide a means to obtain counter-examples ((b) above). The prior testing approach, in particular, by its very nature can only provide guarantees over a finite range of inputs.
The invention described herein is directed to the analysis of the source code or serialized representations of neural networks for analyzing point cloud data. The input is a concrete description (i.e., source code or serialized representation) of a neural network which could be a mathematical (symbolic) description or source code including saved values for all parameters (matrix weights, bias vectors, convolutional kernels, etc.) of the network. The input space of the neural network to be analyzed must be of the form
³ⁿ(i.e., inputs must be n-tuples of three floating point values where these values correspond to x-, y-, and z-coordinates of a point in Euclidean space). The output of the neural network to be analyzed must be either: the estimate of a surface normal (and therefore a three-tuple of floating point values); the estimate of a slope; the estimate of ground or non-ground plane; or the estimate of another geometric property for which the ground-truth value can be calculated analytically for a mathematically well-defined class of representative examples.
One example, where the sensor that produces the data to be consumed by the neural network is a lidar sensor, is shown in FIG. 4. Given a non-ground plane object 400, the lidar sensor 402 (and its firmware) generates point cloud data 404 that is then (possibly after some pre-processing) consumed by the neural network f 406. Pre-processing could include clustering points into batches for multiple smaller collections of points that are closer in space rather than a gigantic point cloud that is dispersed over the entire field-of-view. Pre-processing can also include centering or scaling the point cloud, and similar geometric operations. Such pre-processing will be familiar to one skilled in the art of machine learning. Some kind of pre-processing is often needed to improve performance of machine learning systems and admit scalable computation (e.g., feeding a gigantic point cloud to a neural network might be a bottleneck to real-time operation).
It is assumed that a description artifact that completely describes the neural network f 406 is given. Crucially, this artifact should contain sufficient information to be able to completely recreate the network. In particular, the description artifact should consist either of the source code (including saved weights) or a serialized representation of the neural network. Non-limiting examples of description artifacts include networks stored in ONNX (see Literature Reference No. 10), TensorFlow (see Literature Reference No. 6), or PyTorch (see Literature Reference No. 3) serialization formats.
It is assumed that the neural network f 406 has, as its input, point clouds 404 of a fixed size n. Thus, f 406 can be described mathematically as a function
³ⁿ→Y where the output space Y, which consists of estimates 408, is determined by the particular task of f 406. Concretely this means that f 406 is realized in source code by a function that takes as input n-many triples of floating point values (which could be captured in code as a list of such values) and returns a corresponding value of the appropriate type (typically also a list of floating point values). There are several additional parameters that the invention requires. The following are parameters for the ground plane segmentation task: the description artifact; a coordinate system for Euclidean space fixed (e.g., cartesian x-, y-, z-coordinates, spherical coordinates (r, φ, ν) with φ the azimuthal angle and ν the polar angle); n many (compact and convex) subsets F₁, . . . , F_nof
²that are definable by formulas in the language of first-order real arithmetic with common non-linear function symbols (see below for examples); a tolerance constant ε>0; and any additional implementation specific constant values (see below for examples). Analogous parameters are required for estimation of surface normals or slopes, and these would be evident to one skilled in the art.
A high-level overview of the invention is depicted in FIG. 5. Here, the Parsing module 500 transforms the neural network description artifact 502 into a symbolic representation 504 that mathematically corresponds to an expression in first-order real arithmetic with common non-linear function symbols. The nature of the Parsing module 500 and symbolic representation 504 will vary in different reductions to practice (e.g., depending on the agreed upon description artifact 502 format). As an example, suppose the neural network f in question is the scalar map
→
consisting of an affine transformation x
0.5 x+1.0 followed by a sigmoid activation (i.e., f (x)=σ(0.5x+1.0)). Then, the symbolic representation 504 in SMT-LIB format (see Literature Reference No. 1) would be:
(define-funf((x Real))Real(sigmoid(+(*0.5x)1.0))
assuming the sigmoid function has already been declared/defined. The line above (and those below) are examples of code written in a SMT-LIB programming language corresponding to the definition of the mathematical function described above, as known to one skilled in the art. Once the symbolic representation 504 has been obtained by parsing the description artifact 502, it is fed together with the analysis parameters 506 to the Analyzer module 508. For each i=1, . . . , n, the Analyzer module 508 creates pairs (r_i, φ_i) of symbolic variables ranging over real numbers and corresponding to two-dimensional planar data (here represented for expositional purposes in spherical coordinates, although other coordinate systems can be used). As above, the concrete representation of the symbolic variables will vary depending on the specific features of the experimental studies (e.g., they could correspond to SMT-LIB variable declarations of the form:

- (declare-fun r1 ( ) Real)
- (declare-fun r2( ) Real).

The ensemble of symbolic variables (r_i, φ_i) represents a planar patch in the two-dimensional Euclidean plane in spherical coordinates (i.e., with azimuthal angle φ_iand polar angle 0). The Analyzer module 508 also creates variables ν corresponding to rotations of the planar patch (r_i, φ_i)_iin three-dimensional space. Depending on the precise task, it may be necessary to create either a single variable or multiple variables. Thus, ν might represent a tuple. Denote by p the planar patch of variables and by p*ν the result of rotating this planar patch by the angle(s)ν. The precise details of the mathematical formula required to express this rotation will depend on the choice of coordinate system and the nature of the constraint (e.g., ground plane segmentation or estimate of surface normal) being considered. The precise details in each of these cases will be straightforward for one skilled in the art. The Analyzer module 508 then creates the (symbolic) formula Φ_idefined mathematically as:
(r _i,φ_i)∈F _i.
Denote by Φ the conjunction:
(0<ν<ε)∧∧_iΦ_i.
This symbolic formula is constructed symbolically in the Analyzer module 508, and it represents initial conditions on the variables employed in the analysis of the neural network. As above, such symbolic formulae will be captured differently depending on the details of the embodiment of the invention (in particular, the choice of SMT solver). In SMT-LIB format, the foregoing symbolic formula can be realized as:

- (and (and (<0 theta) (<theta epsilon)) big_and_phi_i),
  where big_and_phi_i is the large conjunction that has already been declared.
  Note that in order for the verification to be sufficiently meaningful, it is important to carefully select the input F_i. For example, these should be such that any collection of points satisfying condition Φ must necessarily contain some non-colinear triple and, for practical reasons, they should contain at least one non-colinear triple that are not too close (what counts as not too close is itself subject to the particular reduction to practice and the additional constants required therein, but should be easily determined by one skilled in the art).

The Analyzer module 508 then constructs the (symbolic) formulas which state that the neural network estimator satisfies the appropriate correctness property. Start by considering the case of ground/non-ground classification. Where f (x) denotes the output ground/non-ground classification of the neural network when provided as input the patch x, the solving task is to show that f(p*ξ) yields the classification:

- Ground, when the initial angle is ν=0 and d(ν,ξ)<ε; or
- Non-ground, when the initial angle is

$ϑ = \frac{π}{2} and d (ϑ, ξ) < ɛ .$
The cases of estimating surface normals and slopes are similar.
Once the Analyzer module 508 has constructed this symbolic formula, it is passed as a query 510 (i.e., the symbolic formula), together with any additional parameters (e.g., timeouts, precision values, etc.), to the SMT solver 512, which carries out its analysis and returns a response 514. The details of how this is handled will be familiar to anyone skilled in the art of using an SMT solver. The Analyzer module 508 then processes this response 514 (e.g., to confirm whether or not a potential counter-example is correct) and returns analysis outputs 516 indicating whether or not the neural network satisfies the correctness property or if a solution could not be found. In the case where the neural network has been shown to fail to satisfy the correctness property, a counter-example will be returned consisting of instantiations of the variables described above as concrete floating point values.
There are three possible kinds of analysis outputs 516 (or indicators): a proof artifact that captures the mathematical reasoning required to prove that the rule is satisfied, a concrete point cloud (i.e., a data structure) or list of such point clouds for which the rule is violated, or a report that sufficient progress could not be made by the analysis (in which case the user would likely modify analysis parameters and try again). These outputs 516 may then be used in a variety of ways. For instance, the analysis outputs 516 can be used to determine what further data should be collected and trained on (in cases of failure) before the neural network is deployed. In this way, the outputs 516 (or indicators) act as a stopping condition preventing the neural network from being deployed. The analysis outputs 516 can also be provided to regulators as part of a certification process. In addition, the analysis outputs 516 can be used by security experts to analyze potential system vulnerabilities arising due to failures identified by this analysis (in the case of counter-examples). Further, the outputs can be fed-back into the training process directly to improve the training of the neural network. In summary, the analysis outputs 516 can be utilized to inform future training, provided as evidence to regulators for certification purposes, fed back directly into the training system in order to augment the original training, and/or utilized by security teams in their inspection of the system for certification or other purposes.
In a first experimental study, the invention was applied to the case of neural networks for estimating surface normals of point cloud patches. In this case, the coordinate system was cartesian and the neural networks considered had architectures of one of the forms illustrated in FIGS. 6A and 6B. The studies were carried out using a combination of Python and C++ code and the SMT solver 512 that used was the dReal SMT solver (see Literature Reference No. 4). The neural network description artifacts 502 were in the TensorFlow serialized model format. A unitless example of a counter-example analysis output 516 for the case of surface normals obtained as part of this experimental study is shown in FIG. 7. In this experimental study, the sets F₁were defined as regularly spaced disjoint rectangular subsets of the unit box [−1,1]²arranged in a grid with two rows and >1 rows to ensure non-degenerate counter-examples would be found. In this case, the rotation was specified by two separate angles of rotation. The dots 700 in FIG. 7 represent the points in the point cloud.
In a second experimental study, neural networks were considered for ground plane segmentation (i.e., classifying whether or not a patch consists of ground plane). As above, the dReal SMT solver 512 was used ibid. The neural network description artifacts 502 were in the PyTorch neural network serialization format. The neural networks considered used an architecture that is not described here. The coordinate system was spherical and the rotation was specified in terms solely of the polar angle D. The sets F₁were defined by first selecting fixed “test points” u_i(e.g., at (0,0), (−1,0), (1,0), (0,−1), (−1,−1) and (1,1)) and then adding as a parameter a value γ>0 such that F₁is defined to be the y-ball (element 800) centered at u_i, which is illustrated in FIG. 8. Specifically, FIG. 8 depicts test points and disjoint sets F_ias y-balls (i.e., sets of the form {x|d(x,u_i)<γ}, where d(−,−) denotes the standard Euclidean metric (element 800). In this case, the code was a mixture of Python and C++ code. Not only were counter-examples found, but also proofs of correctness of ground plane segmentation.
FIG. 9 is a flow diagram illustrating a system for analysis of correctness properties of a neural network for providing estimates for point cloud data. according to some embodiments of the present disclosure. As described in detail above, a neural network based estimate related to a set of point cloud data is received (element 900). A description of the neural network is transformed into a symbolic representation (element 902). Based on a combination of the symbolic representation and a set of parameters, an analysis output indicating whether the neural network satisfies a correctness property in generating the neural network based estimate is generated (element 904).
Current state-of-the-art approaches to the problem of defining correctness are based purely on extensive testing or else using adversarial machine learning. The principal disadvantage of these two approaches is that they only provide a means to obtain counter-examples, whereas the invention described herein enables the user to obtain a proof artifact that imparts mathematical correctness guarantees on the point cloud processor. Such proof artifacts can be used by regulators, certifiers, or quality assurance departments to augment their standard testing approaches in order to obtain better insight into the reliability and trustworthiness of the point cloud processors being analyzed. For example, the standard testing approach could be improved based on the outputs of this invention by collecting further point cloud data to test that is near the conditions identified as yielding incorrect estimates by the invention. Additionally, when point clouds are found that violate the rules, these point clouds can be used to augment the training or test set of the neural network, thereby improving network performance.
Additionally, existing methods for verification of neural networks are concerned exclusively with image (pixel) data and with the typical machine learning benchmark datasets for image based classification tasks. In the context of classification, the problem of defining correctness in a mathematically precise way is very different from the case of geometric tasks of the kind considered when working with point cloud (instead of image) data. For this reason, looking at stronger geometric correctness properties does not make sense for the applications being considered by the prior art in this area. Moreover, the satisfaction modulo theories (SMT) or mixed integer linear programing (MILP) based tools that have been employed in the prior art are not capable of reasoning about the trigonometric functions that occur prominently in the specification of geometric correctness properties (the rules described above) for point cloud data. Because an alternative approach is being used in the method according to embodiments of the present disclosure, which instead uses the dReal SMT solver of Literature Reference No. 4 that is capable of manipulating trigonometric functions, one can formulate and reason about the natural correctness conditions related to point cloud data, as described above.
The present invention can be applied to any program or product line that utilizes neural network based systems for processing point cloud data, such as autonomous or semi-autonomous vehicle platforms that utilize lidar or stereo camera systems. As described above, the present invention can be used either for quality assurance of neural network based point cloud processors or to generate point cloud data that can be utilized to improve training. In particular, the invention is analogous to source code analysis systems, such as described in U.S. Pat. No. 10,423,518 (which is hereby incorporated by reference as though fully set forth herein), in that it takes as input the source code description of the neural network based point cloud processor, carries out an analysis of that source code to check against “rules” (in this case, the rules correspond to the properties of tasks (1)-(3) of interest described above), and then shares the outputs of that analysis with the user.
In one non-limiting example, the invention described herein is used to evaluate whether or not an autonomous system platform is actually safe to operate. A goal of the present invention is to help prevent a system that is likely to fail from being deployed. Thus, the invention is configured to be used at design time (or to provide data to regulators for certification purposes). Analysis outputs inform for what kind of data failures are or are not likely to occur. An example of a failure is an autonomous vehicle hitting a pedestrian. In this type of situation, it is likely that the perception system failed to correctly detect a pedestrian and failed to brake before hitting the pedestrian.
Finally, while this invention has been described in terms of several embodiments, one of ordinary skill in the art will readily recognize that the invention may have other applications in other environments. It should be noted that many embodiments and implementations are possible. Further, the following claims are in no way intended to limit the scope of the present invention to the specific embodiments described above. In addition, any recitation of “means for” is intended to evoke a means-plus-function reading of an element and a claim, whereas, any elements that do not specifically use the recitation “means for”, are not intended to be read as means-plus-function elements, even if the claim otherwise includes the word “means”. Further, while particular method steps have been recited in a particular order, the method steps may occur in any desired order and fall within the scope of the present invention.

Claims

What is claimed is:

1. A system for proving correctness properties of a neural network for providing estimates for point cloud data, the system comprising:

one or more processors and a non-transitory computer-readable medium having executable instructions encoded thereon such that when executed, the one or more processors perform operations of:

receiving, as input, a description of a neural network for generating estimates from a set of point cloud data, wherein the description of the neural network is in one of a source code format or a serialized format;

parsing the description of the neural network to obtain a symbolic representation;

based on a combination of the symbolic representation and a set of analysis parameters, generating an analysis output indicating whether the neural network satisfies a correctness property in generating the estimates from the set of point cloud data,

wherein the analysis output is one of a mathematical proof artifact proving that the set of analysis parameters is satisfied, a list of one or more point clouds for which the set of analysis parameters is violated, and a report that progress could not be made by the analysis.

2. The system as set forth in claim 1, wherein the analysis output acts as a stopping condition for preventing deployment of the neural network.

3. The system as set forth in claim 1, wherein the estimates comprise an estimate of a slope of a patch of the set of point cloud data.

4. The system as set forth in claim 1, wherein the estimates comprise an estimate of a surface normal of a patch of the set of point cloud data.

5. The system as set forth in claim 1, wherein the estimates comprise an estimate regarding whether a patch of the set of point cloud data is ground or non-ground.

6. The system as set forth in claim 1, wherein the set of analysis parameters describe geometric properties of the set of point cloud data.

7. A computer implemented method for proving correctness properties of a neural network for providing estimates for point cloud data, the method comprising an act of:

causing one or more processors to execute instructions encoded on one or more associated memories, each associated memory being a non-transitory computer-readable medium, such that upon execution, the one or more processors perform operations of:

8. The method as set forth in claim 7, wherein the analysis output acts as a stopping condition for preventing deployment of the neural network.

9. The method as set forth in claim 7, wherein the estimates comprise an estimate of a slope of a patch of the set of point cloud data.

10. The method as set forth in claim 7, wherein the estimates comprise an estimate of a surface normal of a patch of the set of point cloud data.

11. The method as set forth in claim 7, wherein the estimates comprise an estimate regarding whether a patch of the set of point cloud data is ground or non-ground.

12. The method as set forth in claim 7, wherein the set of analysis parameters describe geometric properties of the set of point cloud data.

13. A computer readable program for proving correctness properties of a neural network for providing estimates for point cloud data, the computer readable program comprising:

computer-readable instructions stored on a non-transitory computer-readable medium that are executable by a computer having one or more processors for causing the processor to perform operations of:

14. The computer readable program as set forth in claim 13, wherein the analysis output acts as a stopping condition for preventing deployment of the neural network.

15. The computer readable program as set forth in claim 13, wherein the estimates comprise an estimate of a slope of a patch of the set of point cloud data.

16. The computer readable program as set forth in claim 13, wherein the estimates comprise an estimate of a surface normal of a patch of the set of point cloud data.

17. The computer readable program as set forth in claim 13, wherein the estimates comprise an estimate regarding whether a patch of the set of point cloud data is ground or non-ground.

18. The computer readable program as set forth in claim 13, wherein the set of analysis parameters describe geometric properties of the set of point cloud data.