Systems and methods for multi-scale depth balance models

Abstract

Systems and methods for multi-scale depth balance models. A computer-implemented method for classifying and training a neural network includes receiving an input at the neural network, where the input includes a plurality of resolution inputs of varying resolutions, outputting a plurality of feature tensors for each corresponding resolution of the plurality of resolution inputs, fusing the plurality of feature tensors using upsampling or downsampling for varying resolutions, identifying one or more prediction vectors from the plurality of feature tensors using a balance solver, and outputting a loss in response to the one or more prediction vectors.

Description

Technical Field

[0001] The present disclosure relates to computer systems having artificial intelligence capabilities including neural networks. Background Art

[0002] A DEQ (Deep Equilibrium) model can define some forward functions f such that the forward pass of the network is given by first solving an equilibrium point The backward pass can largely correspond to multiplying by the Jacobian which in its exact form requires the solution of a linear system. There are three main challenges in developing DEQ models. One can define f such that a fixed point exists and ideally is unique. A second challenge may be to define a root-finding process that can find the fixed point in the forward pass A third challenge may be to define the backward pass to solve the multiplication by the Jacobian

[0003] Past deep learning methods for computer vision tasks satisfied the first property by having layers that combine features at a specific resolution and gradually downsample the image, but these layers were explicitly defined such that memory requirements increased with the number of layers. On the other hand, DEQ models can simulate larger networks with a small memory footprint but do not learn in a hierarchical manner. The DEQ model can maintain the same resolution throughout the forward and backward passes and does not explicitly define multiple layers, so the DEQ model may have difficulty downsampling at intermediate points in the network. Summary of the Invention

[0004] According to one embodiment, a computer-implemented method for classifying and training a neural network includes receiving an input at the neural network, where the input includes a plurality of resolution inputs of varying resolutions, outputting a plurality of feature tensors for each corresponding resolution of the plurality of resolution inputs, fusing the plurality of feature tensors using upsampling or downsampling for the varying resolutions, identifying one or more prediction vectors from the plurality of feature tensors using an equilibrium solver, and outputting a loss in response to the one or more prediction vectors.

[0005] According to a second embodiment, a computer-implemented method for classifying an input to a neural network includes receiving an input at the neural network, where the input includes a plurality of resolution inputs, identifying a feature tensor for each corresponding resolution of the plurality of resolution inputs, inputting the feature tensor of each corresponding resolution into a corresponding residual block, and identifying one or more prediction vectors from the plurality of feature tensors using an equilibrium solver to output a classification of the input.

[0006] ​According to a third embodiment, a system for training a neural network includes an input interface for accessing input data of the neural network and a processor in communication with the input interface. The processor is programmed to receive the input data at the neural network, where the input includes a plurality of resolution inputs of varying resolution, output a plurality of feature tensors for each corresponding resolution of the plurality of resolution inputs, fuse the plurality of feature tensors using upsampling or downsampling for the varying resolution, identify a prediction vector in response to the plurality of feature tensors using a balance solver, where the prediction vector includes features associated with each of the plurality of feature tensors, and output the prediction vector as a classification of the input data. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 FIG. 1 shows a system 100 for training a neural network.

[0008] Figure 2 FIG. 2 shows a computer-implemented method 200 for training a neural network.

[0009] Figure 3 FIG. 3 illustrates the general structure of the MDEQ model.

[0010] Figure 4 FIG. 4 depicts the internal structure of a system 400 including residual blocks.

[0011] Figure 5 FIG. 5 depicts a schematic diagram of the interaction between a computer-controlled machine 10 and a control system 12.

[0012] Figure 6 FIG. 6 depicts Figure 1 a schematic diagram of a control system configured to control a vehicle, which may be a semi-autonomous vehicle or a semi-autonomous robot.

[0013] Figure 7 FIG. 7 depicts Figure 1 a schematic diagram of a control system configured to control a manufacturing machine (such as a part of a production line) of a manufacturing system, such as a stamping tool, a cutting tool, or a gun drill.

[0014] Figure 8 FIG. 8 depicts Figure 1 a schematic diagram of a control system configured to control a power tool having at least a semi-autonomous mode, such as a drill or a driver.

[0015] Figure 9 FIG. 9 depicts a schematic diagram of a control system configured to control an automated personal assistant Figure 1 FIG. 10 depicts a schematic diagram of a control system configured to control an automated personal assistant.

[0016] Figure 10 FIG. 11 depicts Figure 1Schematic diagram of a control system configured to control a monitoring system, such as an access control system or a surveillance system.

[0017] Figure 11 depicts Figure 1 Schematic diagram of a control system configured to control an imaging system, such as an MRI device, an x-ray imaging device, or an ultrasonic device. DETAILED DESCRIPTION

[0018] Embodiments of the present disclosure are described herein. However, it will be understood that the disclosed embodiments are merely examples, and other embodiments may take various forms and alternative forms. The figures are not necessarily to scale; some features may be enlarged or minimized to show details of particular components. Thus, the specific structural and functional details disclosed herein should not be construed as limiting, but merely as a representative basis for teaching one skilled in the art to employ the embodiments in various ways. As will be understood by those of ordinary skill in the art, the various features illustrated and described with reference to any one figure may be combined with features illustrated in one or more other figures to produce embodiments that are not explicitly illustrated or described. Combinations of the illustrated features provide representative embodiments for typical applications. However, for a particular application or implementation, various combinations and modifications of the features consistent with the teachings of the present disclosure may be desirable.

[0019] In previous DEQ models, defining a function with a fixed point could be heuristically done by using a network structure that empirically seemed to produce stable fixed points. This could include using existing transformer-like architectures or existing convolutional architectures, with appropriate layer normalization to attempt to provide a stable system. However, such layers do not have a formal guarantee of the existence of a fixed point. For the root-finding process, DEQ can use the nonlinear Broyden method, which also does not guarantee finding a root (even if a root or fixed point exists). Additionally, the backward pass can use a similar Broyden method for linear solving, which can be guaranteed to multiply by an appropriate inverse if the appropriate inverse exists, but if there is no fixed point or if the forward pass fails to converge, the Jacobian will not correspond to a meaningful derivative.

[0020] A deep neural network having z a hidden layer f and an activation layer such that for where the weights and the previous layer input are both bound across layers, e.g., . Some of these activations f can exhibit attractor properties, e.g., there exists a fixed point such that and , i.e., f repeated application until the initial activation converges to a fixed point . If this is the case, the iterative function application can be equivalently replaced by a numerical method to directly find the fixed point. This transfers the problem from computing the forward and backward passes of multiple layers to directly computing and optimizing the fixed point via a numerical method. This can reduce memory footprint (since intermediate values in the layers do not need to be stored) and solve the problem of finding the optimal number of layers L .

[0021] The Multiscale Deep Equilibrium Model (MDEQ) is built on top of its predecessor, the Deep Equilibrium Model (DEQ). When the input to the DEQ has a single resolution, the input to the MDEQ is supplied at multiple resolutions, which allows it to learn from a spectrum of resolutions ranging from fine-grained features (high resolution) to global features (low resolution). The MDEQ also specifies the process of mixing and combining information at different scales. Additionally, the output from the DEQ has the same resolution as its input, while the MDEQ gives an output for each input scale. This makes the MDEQ model more flexible in terms of joint learning or transfer, as the outputs at different resolutions can be used to learn auxiliary losses for a single training task to learn several tasks simultaneously (i.e., using the high-resolution output for semantic segmentation and the low-resolution output for image classification), or to more easily transfer the learned model from one task to another.

[0022] The MDEQ model learns in a hierarchical manner, e.g., a person considering data at multiple scales or resolutions, which is necessary for learning in multi-tasks (e.g., computer vision tasks). The MDEQ model is implicit and it can simulate much larger networks while maintaining a relatively small memory footprint for storing and training the model.

[0023] Figure 1 FIG. shows a system 100 for training a neural network. The system 100 may include an input interface for accessing training data 192 of the neural network. For example, as Figure 1As illustrated, the input interface may be constituted by a data storage interface 180, which may access training data 192 from a data storage device 190. For example, the data storage interface 180 may be a memory interface or a permanent storage interface, such as a hard disk or SSD interface, but may also be a personal area network, local area network or wide area network interface, such as a Bluetooth, Zigbee or Wi-Fi interface or an Ethernet or fiber optic interface. The data storage device 190 may be an internal data storage device of the system 100, such as a hard disk drive or SSD, but may also be an external data storage device, such as a network-accessible data storage device.

[0024] In some embodiments, the data storage device 190 may further include a data representation 194 of an untrained version of the neural network, which may be accessed by the system 100 from the data storage device 190. However, it will be appreciated that the training data 192 and the data representation 194 of the untrained neural network may also each be accessed from different data storage devices, for example via different subsystems of the data storage interface 180. Each subsystem may have the types as described above for the data storage interface 180. In other embodiments, the data representation 194 of the untrained neural network may be generated internally by the system 100 based on the design parameters of the neural network and may thus not be explicitly stored on the data storage device 190. The system 100 may further include a processor subsystem 160, which may be configured to provide, during operation of the system 100, an iterative function as an alternative to a stack of layers of the neural network to be trained. Here, the corresponding layers of the replaced stack of layers may have weights shared with each other and may receive the output of the previous layer as input, or for the first layer of the stack of layers, receive an initial activation and a part of the input of the stack of layers. The processor subsystem 160 may further be configured to iteratively train the neural network using the training data 192. Here, the training iterations of the processor subsystem 160 may include a forward propagation part and a backward propagation part. The processor subsystem 160 may be configured to perform the forward propagation part, among other operations, in particular by: defining the forward propagation part that may be performed, determining an equilibrium point of the iterative function, at which the iterative function converges to a fixed point, where determining the equilibrium point includes using a numerical root-finding algorithm to find the root solution of the iterative function minus its input, and by providing the equilibrium point as an alternative to the output of the stack of layers in the neural network. The system 100 may further include an output interface for outputting a data representation 196 of the trained neural network, which may also be referred to as training model data 196. For example, also as in Figure 1As shown in the figure, the output interface may be constituted by the data storage interface 180. In these embodiments, the interface is an input / output ("IO") interface through which the training model data 196 may be stored in the data storage device 190. For example, the data representation 194 defining the "untrained" neural network may be at least partially replaced by the data representation 196 of the trained neural network during or after training because the parameters of the neural network, such as the weights, hyperparameters, and other types of parameters of the neural network, may be adapted to reflect the training on the training data 192. This is also illustrated in Figure 1 Figure by reference numerals 194, 196 of the same data record on the data storage device 190. In other embodiments, the data representation 196 may be stored separately from the data representation 194 defining the "untrained" neural network. In some embodiments, the output interface may be separate from the data storage interface 180, but generally may be of the type described above for the data storage interface 180.

[0025] Figure 2 A computer-implemented method 200 for training a neural network is shown. Method 200 may but does not need to correspond to Figure 1 the operation of the system 100 because it may also correspond to the operation of another type of system, device, or equipment, or because it may correspond to a computer program.

[0026] Method 200 is shown to include providing 210 a neural network in a step titled "Providing a data representation of a neural network", where providing the neural network includes providing an iterative function as an alternative to a layer stack of the neural network, where corresponding layers of the replaced layer stack have weights shared among each other and receive the output of the previous layer as input, or for the first layer of the layer stack, receive an initial activation and a portion of the input of the layer stack. Method 200 is further shown to include accessing 220 training data of the neural network in a step titled "Accessing training data". Method 200 is further shown to include iteratively training 230 the neural network using the training data in a step titled "Iteratively training a neural network using training data", where the training 230 can include a forward propagation portion and a backward propagation portion. Performing the forward propagation portion by method 200 can include determining 240 an equilibrium point of the iterative function in a step titled "Determining an equilibrium point using a root-finding algorithm", at which the iterative function converges to a fixed point, where determining the equilibrium point includes using a numerical root-finding algorithm to find a root solution of the iterative function minus its input, and including providing 250 the equilibrium point as an alternative to the output of the layer stack in a neural network in a step titled "Providing the equilibrium point as an alternative to the output of the layer stack". Method 200 can further include outputting 260 the trained neural network after training and in a step titled "Outputting the trained neural network". A Deep Equilibrium (DEQ) neural network can be further described in a patent application titled "DEEP NEURAL NETWORK WITH EQUILIBRIUM SOLVER" having application number X, XXX, XXX, which is incorporated herein by reference in its entirety.

[0027] Figure 3 The general structure of the MDEQ model 300 is illustrated. As in previous DEQ models, transformations that can mark fixed points of interest to the system, and x can be a precomputed input representation provided to and z for the internal state of the model.

[0028] For context, in the original DEQ model, the DEQ in the forward pass (inference) sequence can include an input and . Hyperparameters can include a basic layer function . The algorithm can:

[0029] 1. Initialize memory .

[0030] 2. Define a function .

[0031] 3. Call a subroutine 。

[0032] The DEQ model can output 。Specifically, (root finding) can be calculated via any root finder method. In a non-limiting manner, this can include any variant of Newton's method, e.g., the classical Newton-Raphson method, Broyden's method, Steffensen's method, etc.

[0033] During the DEQ backward pass training sequence, the system can utilize the backpropagation error and 、 from the forward pass as inputs. The algorithm for the backward pass can run the following steps:

[0034] 1. Define the function

[0035] 2. Calculate the partial derivatives and 。

[0036] 3. A) Calculate and 。

[0037] Or

[0038] B) Solve the linear system via a suitable subroutine.

[0039] The MDEQ model described further below is built on the following forward pass sequence and backward pass sequence by specifying the function used in the above algorithm. The hyperparameters in the MDEQ model can include n , the number of resolutions. The input (e.g., the input) can include (the input at each scale, where the resolutions are in descending order) and (the hidden state having the same size as ). For each resolution , the algorithm can:

[0040]

[0041]

[0042]

[0043] Norm can be defined as a set of normalization operations (similar to batch normalization).

[0044] Then for each resolution , the MDEQ model can apply a multi-resolution fusion step that mixes features from other scales:

[0045]

[0046] One or more 2-stride 3×3 convolutions (e.g., is one or more) are applied to the higher resolution, and followed by a 1×1 convolution with bilinear interpolation applied to the lower resolution.

[0047] f The output of is a set of hidden states

[0048] However, these steps can be modified as long as the first step is a transformation that is separately applied to each resolution to maintain its shape, and the second step is a transformation that mixes information between different resolutions. This function can then be used in the forward and backward passes above. After the forward pass, the model can obtain the root of , which in the case of MDEQ is a set of resolutions

[0049] One of the main concepts of MDEQ is a transformation that is driven to equilibrium The system can use a design where features at different resolutions 303 are first obtained through residual blocks 307. The features can be derived from an input image 301 with varying resolutions 303 or resolution images 303, as shown in Figure 3 Each resolution 303 can have its own corresponding residual block 307. The residual blocks 307 can be shallow and structurally the same. Streams at different resolutions can be processed side by side or simultaneously. At resolution i , the residual block can receive the internal state along with the input representation , and output a transformed feature tensor 305 at the same resolution.

[0050] The input 301 can undergo a series of transformations to form x 303, x 303 will be provided to In contrast to the original DEQ model that acts on a sequence and is provided with a single representation of the input, MDEQ can be provided with input representations at n resolutions (e.g., n= 4). The transformation at each resolution receives its own input representation,

[0051] For , . The input representation at a lower resolution can be derived from the original (high-resolution) input via a 2-strided convolution.

[0052] While the original DEQ has an internal state with a single tensor z , the MDEQ state includes a set of tensors 305 at n resolutions:

[0053] (The hidden state has the same size as ).

[0054] Each of the different tensors 305 has a different dimension and a different resolution. As such, the equilibrium solver 311 can be based on previous methods of other DEQ models. The system can initialize the internal state by setting i for all scales . Since the system is performing root finding, the system does not need to vectorize the multi-resolution tensors. The set of tensors can be maintained as a set of n tensors, whose corresponding equilibrium states can be solved and backpropagated simultaneously, where each resolution induces its own gradient.

[0055] Compared to previous DEQ models, the solver 311 can also be modified. The high dimensionality of the images can make storing such updates very expensive, despite their low rank. To address this, the system improves the memory efficiency of the forward and backward passes. The new solver can keep the most recent m low-rank updates at any step and discard the earlier ones. The solver 311 can output a single prediction vector 313 for all resolutions. The prediction vector 313 can include the features of each vector at the corresponding resolution.

[0056] Existing implicit models assume that the loss 317 is defined on a single stream of implicit hidden states with a uniform input and output shape. As such, it is not clear how such models can transfer across structurally different tasks (e.g., pre-training for image classification and fine-tuning for semantic segmentation). Additionally, there is no natural way to define auxiliary losses because there are no "layers" and the forward and backward computation traces are decoupled. The loss 317 in the MDEQ model can be associated with each prediction vector because the MDEQ model can assign the loss to more than one vector. As such, the vectors can always have a loss associated with at least one of them.

[0057] The MDEQ model can expose convenient "interfaces" to its states at multiple resolutions. One resolution (the highest) can be the same as the resolution of the input and can be used to define the loss for dense prediction tasks such as semantic segmentation. For example, this can include deriving specific classifications of the associated parts of the identified objects in the image, such as face recognition, eye recognition, hair color, etc. Another resolution (e.g., the lowest) can be a vector in which the spatial dimensions are collapsed and can be used to define the loss for image-level tagging tasks such as image classification. This can suggest a clean protocol for training the same model for different tasks either jointly (e.g., multi-task learning, where structurally different supervision flows through multiple heads) or sequentially (e.g., pre-training for image classification through one head and fine-tuning for semantic segmentation through another head).

[0058] Because batch normalization may not be directly suitable for implicit models because it estimates the population statistics based on layers, which are implicit in the MDEQ setting, and the Jacobian matrix of the transformation can scale poorly such that the fixed point is significantly harder to solve. Therefore, MDEQ can utilize group normalization, which groups the input channels and performs normalization within each group (e.g., at each resolution). Group normalization can be independent of the batch size and provides more natural support for transfer learning (e.g., pre-training and fine-tuning for structurally different tasks). For stability, MDEQ may not utilize the learnable affine parameters of group normalization.

[0059] Rather than utilizing the conventional spatial dropout used by interpretive vision models applied with a random mask to a given layer in the network, MDEQ can adopt the variational dropout technique found in RNNs (recurrent neural networks), where the exact same mask can be applied at all calls of and reset at the start of each training iteration.

[0060] For all resolutions i , the multi-scale features can be initialized to . However, this can cause a certain instability during the training of MDEQ, especially during its start phase, which may be due to the sharp change in the slope of the original ReLU non-linearity where the derivative is undefined. To address this, during the initial phase of training, the MDEQ model can replace the last ReLU in both the residual block and the multi-scale fusion with softplus. These may later switch back to ReLU. Softplus provides a smooth approximation to ReLU but has a slope around (where controls the curvature).

[0061] The MDEQ model can be applied to convolutions with a small receptive field on potentially very large images (e.g., 2048 X 1024 images), such as the two 3×3 convolutional filters in the residual block of

[0062] Figure 4 The internal structure of system 400 including a residual block is disclosed. The residual block can utilize group normalization 413 instead of batch normalization. Below, the residual block at resolution i can be formally expressed as follows, where for each resolution : Output of

[0063]

[0064]

[0065]

[0066] Therefore, a two-dimensional (2D) convolutional layer 405 with group normalization can output a convolutional kernel wrapped with the layer to help generate an output tensor. A rectifier 407 can be utilized for the convolutional kernel output from block 405. The activation function can include rectified linear units (ReLU) 407, 411. A second 2D convolutional layer 409 with group normalization can receive the internal state along with the input representation and output a transformed feature tensor at the same resolution, as specified in the above formula. The residual block can apply a 2-stride 3×3 convolution to higher resolutions and a 1×1 convolution followed by bilinear interpolation to lower resolutions.

[0067] After these blocks, the second part is a multi-resolution fusion step that mixes feature maps across different scales. The transformed feature undergoes from the current scale i to each other scale not equal to i j ​Upsampling or downsampling. In the MDEQ architecture, downsampling can be performed via (one or more) 2-stride 3×3 convolutional 2d, and upsampling is performed via a 1×1 convolution followed by bilinear interpolation. The final output scale can be formed by summing over the transformed feature maps provided from all incoming scales i (along with )). Thus, the output feature tensor at each scale is a mixture of the transformed features from all scales. This may force the features at all scales to be consistent and drive the entire system to a coordinated equilibrium, thus harmonizing the representations across scales. j

[0068] Figure 5 FIG. depicts a schematic diagram of the interaction between a computer-controlled machine 10 and a control system 12. The computer-controlled machine 10 may include a neural network as described in Figure 1-4 . The computer-controlled machine 10 includes an actuator 14 and a sensor 16. The actuator 14 may include one or more actuators, and the sensor 16 may include one or more sensors. The sensor 16 is configured to sense the condition of the computer-controlled machine 10. The sensor 16 may be configured to encode the sensed condition into a sensor signal 18 and transmit the sensor signal 18 to the control system 12. Non-limiting examples of the sensor 16 include video, radar, LiDAR, ultrasonic, and motion sensors. In one embodiment, the sensor 16 is an optical sensor configured to sense an optical image of the environment near the computer-controlled machine 10.

[0069] The control system 12 is configured to receive the sensor signal 18 from the computer-controlled machine 10. As will be elaborated below, the control system 12 may further be configured to calculate an actuator control command 20 depending on the sensor signal and transmit the actuator control command 20 to the actuator 14 of the computer-controlled machine 10.

[0070] As shown in Figure 5 , the control system 12 includes a receiving unit 22. The receiving unit 22 may be configured to receive the sensor signal 18 from the sensor 16 and transform the sensor signal 18 into an input signal x. In an alternative embodiment, in the absence of the receiving unit 22, the sensor signal 18 is directly received as the input signal x. Each input signal x may be a part of each sensor signal 18. The receiving unit 22 may be configured to process each sensor signal 18 to generate each input signal x. The input signal x may include data corresponding to the image recorded by the sensor 16.

[0071] ​The control system 12 includes a classifier 24. The classifier 24 can be configured to classify an input signal x into one or more labels using a machine learning (ML) algorithm such as the neural network described above. The classifier 24 is configured to be parameterized by parameters such as the parameters described above (e.g., parameter ). The parameters can be stored in and provided by the non-volatile storage device 26. The classifier 24 is configured to determine an output signal y from the input signal x. Each output signal y includes information assigning one or more labels to each input signal x. The classifier 24 can transmit the output signal y to the conversion unit 28. The conversion unit 28 is configured to convert the output signal y into an actuator control command 20. The control system 12 is configured to transmit the actuator control command 20 to the actuator 14, and the actuator 14 is configured to actuate the computer-controlled machine 10 in response to the actuator control command 20. In another embodiment, the actuator 14 is configured to actuate the computer-controlled machine 10 directly based on the output signal y.

[0072] When the actuator 14 receives the actuator control command 20, the actuator 14 is configured to perform an action corresponding to the associated actuator control command 20. The actuator 14 can include control logic configured to transform the actuator control command 20 into a second actuator control command for controlling the actuator 14. In one or more embodiments, instead of or in addition to the actuator, the actuator control command 20 can be used to control a display.

[0073] In another embodiment, instead of or in addition to the computer-controlled machine 10 including the sensor 16, the control system 12 includes the sensor 16. Instead of or in addition to the computer-controlled machine 10 including the actuator 14, the control system 12 can also include the actuator 14.

[0074] As shown in Figure 5 , the control system 12 also includes a processor 30 and a memory 32. The processor 30 can include one or more processors. The memory 32 can include one or more memory devices. The classifier 24 (e.g., the ML algorithm) of one or more embodiments can be implemented by the control system 12, which includes the non-volatile storage device 26, the processor 30, and the memory 32.

[0075] The non-volatile storage device 26 may include one or more persistent data storage devices such as hard disk drives, optical drives, tape drives, non-volatile solid state devices, cloud storage, or any other device capable of persistently storing information. The processor 30 may include one or more devices selected from high performance computing (HPC) systems, which include high performance cores, microprocessors, microcontrollers, digital signal processors, microcomputers, central processing units, field programmable gate arrays, programmable logic devices, state machines, logic circuits, analog circuits, digital circuits, or any other device that manipulates signals (analog or digital) based on computer-executable instructions residing in the memory 32. The memory 32 may include a single memory device or multiple memory devices, including but not limited to random access memory (RAM), volatile memory, non-volatile memory, static random access memory (SRAM), dynamic random access memory (DRAM), flash memory, cache memory, or any other device capable of storing information.

[0076] The processor 30 may be configured to read into the memory 32 and execute computer-executable instructions of one or more ML algorithms and / or methods embodied in the non-volatile storage device 26 and implementing one or more embodiments. The non-volatile storage device 26 may include one or more operating systems and applications. The non-volatile storage device 26 may store data compiled and / or interpreted from computer programs created using a variety of programming languages and / or technologies, which include, without limitation and individually or in combination, Java, C, C++, C#, Objective C, Fortran, Pascal, Java Script, Python, Perl, and PL / SQL.

[0077] When executed by the processor 30, the computer-executable instructions of the non-volatile storage device 26 may cause the control system 12 to implement one or more ML algorithms and / or methods as disclosed herein. The non-volatile storage device 26 may also include ML data (including data parameters) that support the functions, features, and processes of one or more embodiments described herein.

[0078] The program code embodying the algorithms and / or methods described herein can be distributed, alone or in combination, as a program product in a variety of different forms. The program code can be distributed using a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to implement aspects of one or more embodiments. A computer-readable storage medium that is non-transitory in nature can include volatile and non-volatile, removable and non-removable tangible media implemented in any method or technology for the storage of information such as computer-readable instructions, data structures, program modules, or other data. The computer-readable storage medium can further include RAM, ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid state storage technology, portable compact disc read-only memory (CD-ROM), or other optical storage device, cassette tape, magnetic tape, disk storage device, or other magnetic storage device, or any other medium that can be used to store the desired information and that can be read by a computer. The computer-readable program instructions can be downloaded to a computer, another type of programmable data processing apparatus, or another device from a computer-readable storage medium, or via a network to an external computer or external storage device.

[0079] The computer-readable program instructions stored in the computer-readable medium can be used to direct a computer, other type of programmable data processing apparatus, or other device to function in a particular manner, such that the instructions stored in the computer-readable medium produce an article of manufacture including instructions for implementing the functions, acts, and / or operations specified in the flowchart or diagram. In certain alternative embodiments, the functions, acts, and / or operations specified in the flowchart and diagram can be reordered, processed serially, and / or processed simultaneously in accordance with one or more embodiments. Additionally, any flowchart and / or diagram can include more or fewer nodes or blocks than those illustrated in accordance with one or more embodiments.

[0080] Suitable hardware components such as application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), state machines, controllers, or other hardware components or devices, or combinations of hardware, software, and firmware components can be used to embody a process, method, or algorithm, in whole or in part.

[0081] Figure 6 A schematic diagram depicting a control system 12 configured to control a vehicle 50, which can be at least partially autonomous or at least partially autonomous robot. As in Figure 5As shown, vehicle 50 includes actuator 14 and sensor 16. Sensor 16 can include one or more video sensors, radar sensors, ultrasonic sensors, LiDAR sensors, and / or position sensors (e.g., GPS). One or more of the one or more specific sensors can be integrated into vehicle 50. Alternatively or in addition to the one or more specific sensors identified above, sensor 16 can include a software module that is configured to determine the state of actuator 14 when executed. A non-limiting example of the software module includes a weather information software module that is configured to determine the current or future state of the weather near or at other locations of vehicle 50.

[0082] Classifier 24 of control system 12 of vehicle 50 can be configured to detect an object near vehicle 50 depending on input signal x. In such an embodiment, output signal y can include information characterizing the object near vehicle 50. An actuator control command 20 can be determined based on this information. Actuator control command 20 can be used to avoid a collision with the detected object.

[0083] In an embodiment where vehicle 50 is at least a partially autonomous vehicle, actuator 14 can be embodied in the brakes, propulsion system, engine, transmission, or steering system of vehicle 50. An actuator control command 20 can be determined such that actuator 14 is controlled so that vehicle 50 avoids a collision with the detected object. The detected object can also be classified according to what classifier 24 deems them most likely to be - such as a pedestrian or a tree. Actuator control command 20 can be determined depending on the classification.

[0084] In other embodiments where vehicle 50 is at least a partially autonomous robot, vehicle 50 can be a mobile robot that is configured to perform one or more functions such as flying, swimming, diving, and walking. The mobile robot can be at least a partially autonomous lawn mower or at least a partially autonomous cleaning robot. In such an embodiment, an actuator control command 20 can be determined such that the propulsion unit, steering unit, and / or braking unit of the mobile robot can be controlled so that the mobile robot can avoid a collision with the identified object.

[0085] In another embodiment, vehicle 50 is at least a partially autonomous robot in the form of a gardening robot. In such an embodiment, vehicle 50 can use an optical sensor as sensor 16 to determine the state of plants in the environment near vehicle 50. Actuator 14 can be a nozzle configured to spray a chemical. Depending on the identified plant species and / or the identified state of the plant, an actuator control command 20 can be determined so that actuator 14 sprays an appropriate amount of an appropriate chemical onto the plant.

[0086] Vehicle 50 can be at least partially autonomous robot in the form of a household appliance. Non-limiting examples of household appliances include washing machines, stoves, ovens, microwave ovens, or dishwashers. In such a vehicle 50, sensor 16 can be an optical sensor configured to detect the state of an object that will undergo processing by the household appliance. For example, in the case where the household appliance is a washing machine, sensor 16 can detect the state of the laundry inside the washing machine. Actuator control command 20 can be determined based on the detected state of the laundry.

[0087] Figure 7 A schematic diagram of a control system 12 configured to control a control system 100 (e.g., a manufacturing machine) such as a stamping tool, a cutting tool, or a gun drill of a manufacturing system 102 (such as part of a production line) is depicted. The control system 12 can be configured to control an actuator 14 that is configured to control the control system 100 (e.g., a manufacturing machine).

[0088] The sensor 16 (e.g., a manufacturing machine) of the system 100 can be an optical sensor configured to capture one or more characteristics of a manufactured product 104. A classifier 24 can be configured to determine the state of the manufactured product 104 based on the one or more captured characteristics. The actuator 14 can be configured to control the control system 100 (e.g., a manufacturing machine) depending on the state of the manufactured product 104 determined for a subsequent manufacturing step of the manufactured product 104. The actuator 14 can be configured to control the function of the control system 100 (e.g., a manufacturing machine) on a subsequent manufactured product 106 in the system 100 (e.g., a manufacturing machine) depending on the determined state of the manufactured product 104.

[0089] Figure 8 A schematic diagram of a control system 12 configured to control a power tool 150 having at least a partially autonomous mode, such as a drill or a driver, is depicted. The control system 12 can be configured to control an actuator 14 that is configured to control the power tool 150.

[0090] The sensor 16 of the power tool 150 can be an optical sensor configured to capture one or more characteristics of the work surface 152 and / or the fastener 154 driven into the work surface 152. The classifier 24 can be configured to determine the state of the work surface 152 and / or the fastener 154 relative to the work surface 152 based on one or more captured characteristics. The state can be that the fastener 154 is flush with the work surface 152. Alternatively, the state can be the hardness of the work surface 152. The actuator 14 can be configured to control the power tool 150 such that the drive function of the power tool 150 is adjusted depending on the determined state of the fastener 154 relative to the work surface 152 or one or more captured characteristics of the work surface 152. For example, if the state of the fastener 154 is flush with the work surface 152, the actuator 14 can interrupt the drive function. As another non-limiting example, the actuator 14 can apply additional or less torque depending on the hardness of the work surface 152.

[0091] Figure 9 A schematic diagram depicting a control system 12 configured to control an automated personal assistant 900 is shown. The control system 12 can be configured to control an actuator 14, which is configured to control the automated personal assistant 900. The automated personal assistant 900 can be configured to control household appliances such as a washing machine, a stove, an oven, a microwave oven, or a dishwasher.

[0092] The sensor 16 can be an optical sensor and / or an audio sensor. The optical sensor can be configured to receive video images of the gesture 904 of the user 902. The audio sensor can be configured to receive voice commands of the user 902.

[0093] The control system 12 of the automated personal assistant 900 can be configured to determine an actuator control command 20, which is configured to control the control system 12. The control system 12 can be configured to determine the actuator control command 20 based on the sensor signal 18 of the sensor 16. The automated personal assistant 900 is configured to transmit the sensor signal 18 to the control system 12. The classifier 24 of the control system 12 can be configured to perform a gesture recognition algorithm to identify the gesture 904 made by the user 902, determine the actuator control command 20, and transmit the actuator control command 20 to the actuator 14. The classifier 24 can be configured to retrieve information from a non-volatile storage device in response to the gesture 904 and output the retrieved information in a form suitable for the user 902 to receive.

[0094] Figure 10FIG. depicts a schematic diagram of a control system 12 configured to control a surveillance system 250. The surveillance system 250 may be configured to physically control access through a door 252. A sensor 16 may be configured to detect a scenario related to determining whether to grant access. The sensor 16 may be an optical sensor configured to generate and transmit image and / or video data. The control system 12 may use such data to detect a person's face.

[0095] A classifier 24 of the control system 12 of the surveillance system 250 may be configured to interpret the image and / or video data by matching the identities of known persons stored in a non-volatile storage device 26, thereby determining the identity of the person. The classifier 24 may be configured to generate an actuator control command 20 in response to the interpretation of the image and / or video data. The control system 12 is configured to transmit the actuator control command 20 to an actuator 14. In this embodiment, the actuator 14 may be configured to lock or unlock the door 252 in response to the actuator control command 20. In other embodiments, non-physical logical access control is also possible.

[0096] The surveillance system 250 may also be a monitoring system. In such an embodiment, the sensor 16 may be an optical sensor configured to detect a scenario under surveillance, and the control system 12 is configured to control a display 254. The classifier 24 is configured to determine the classification of the scenario, such as whether the scenario detected by the sensor 16 is suspicious. The control system 12 is configured to transmit the actuator control command 20 to the display 254 in response to the classification. The display 254 may be configured to adjust the displayed content in response to the actuator control command 20. For example, the display 254 may highlight an object classified as suspicious by the classifier 24.

[0097] Figure 11 FIG. depicts a schematic diagram of a control system 12 configured to control an imaging system 1100, such as an MRI device, an x-ray imaging device, or an ultrasonic device. The sensor 16 may be, for example, an imaging sensor. The classifier 24 may be configured to determine the classification of all or part of a sensed image. The classifier 24 may be configured to determine or select an actuator control command 20 in response to the classification obtained by a trained neural network. For example, the classifier 24 may interpret a region of the sensed image as a potential anomaly. In this case, an actuator control command 20 may be determined or selected such that a display 302 displays the imaging and highlights the potential anomaly region.

[0098] The processes, methods, or algorithms disclosed herein may be delivered to or implemented by a processing device, controller, or computer, which may include any existing programmable electronic control unit or dedicated electronic control unit. Similarly, the processes, methods, or algorithms may be stored in a variety of forms as data and instructions executable by a controller or computer, including but not limited to information permanently stored on a non-writable storage medium such as a ROM device and information variably stored on a writable storage medium such as a floppy disk, magnetic tape, CD, RAM device, and other magnetic and optical media. The processes, methods, or algorithms may also be implemented in a software executable object. Alternatively, suitable hardware components such as application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), state machines, controllers, or other hardware components or devices, or combinations of hardware, software, and firmware components, may be used to embody the processes, methods, or algorithms, in whole or in part.

[0099] While the foregoing has described exemplary embodiments, it is not intended that these embodiments describe all possible forms encompassed by the claims. The words used in the specification are words of description rather than limitation, and it is understood that various changes may be made without departing from the spirit and scope of the disclosure. As previously described, the features of the various embodiments may be combined to form additional embodiments of the invention that may not be explicitly described or illustrated. While the various embodiments may have been described as providing advantages over other embodiments or over prior art implementations in one or more desired characteristics or being preferred to other embodiments or prior art implementations, one of ordinary skill in the art recognizes that one or more features or characteristics may be compromised depending on the specific application and implementation to achieve the desired overall system attributes. These attributes may include, but are not limited to, cost, strength, durability, life cycle cost, marketability, appearance, packaging, size, suitability, weight, manufacturability, ease of assembly, etc. Accordingly, to the extent that any embodiment is described as less desirable in one or more features compared to other embodiments or prior art implementations, such embodiments are not outside the scope of the disclosure and may be desirable for a particular application.

Claims

1. A computer-implemented method for classifying and training a neural network, comprising: Receive an input at a neural network, the input being an image, where the input includes a plurality of resolution inputs of varying resolutions; Output a plurality of feature tensors for each corresponding resolution of the plurality of resolution inputs; Fuse the plurality of feature tensors using upsampling or downsampling of the corresponding resolution; Identify one or more prediction vectors from the plurality of feature tensors using a balance solver; and Output a loss in response to the one or more prediction vectors.

2. The computer-implemented method according to claim 1, wherein outputting the loss is in response to a backpropagation sequence that utilizes only a single layer.

3. The computer-implemented method according to claim 1, wherein Each of the plurality of resolution inputs is input to a corresponding residual block for a specific resolution.

4. The computer-implemented method according to claim 1, wherein the resolution input comprises pixel images of different sizes.

5. The computer-implemented method according to claim 1, wherein The method includes using a plurality of residual blocks to output feature tensors for each corresponding resolution of the plurality of resolution inputs.

6. The computer-implemented method according to claim 1, wherein one of the feature tensors corresponding to the highest resolution of the plurality of resolution inputs is only downsampled.

7. The computer-implemented method according to claim 1, wherein one of the feature tensors corresponding to the lowest resolution of the plurality of resolution inputs is only upsampled.

8. The computer-implemented method according to claim 1, wherein the fusion of the plurality of feature tensors is performed in a single layer.

9. The computer-implemented method according to claim 1, wherein the downsampling is performed by a 2-stride 3-by-3 two-dimensional convolutional layer.

10. The computer-implemented method according to claim 1, wherein The upsampling is performed by a 1-by-1 convolution followed by bilinear interpolation.

11. A computer-implemented method for classifying an input to a neural network, comprising: Receive an input at a neural network, the input being an image, where the input includes a plurality of resolution inputs; Identify feature tensors for each corresponding resolution of the plurality of resolution inputs; Input the feature tensors of each corresponding resolution into corresponding residual blocks; and Identify one or more prediction vectors from the feature tensors using a balance solver to output a classification of the input.

12. The computer-implemented method according to claim 11, wherein the method comprises fusing the feature tensors for each corresponding residual block by downsampling or upsampling.

13. The computer-implemented method according to claim 11, wherein the plurality of resolution inputs are varying resolutions of the image.

14. The computer-implemented method according to claim 11, wherein the neural network comprises only a single layer.

15. The computer-implemented method according to claim 11, wherein the method comprises applying group normalization to group the multiple-resolution inputs and performing normalization within each corresponding resolution.

16. The computer-implemented method according to claim 11, wherein the method comprises outputting a loss in response to training one or more prediction vectors of the neural network using backpropagation, wherein the backpropagation comprises identifying the derivative of the loss associated with hyperparameters of the neural network.

17. The computer-implemented method according to claim 11, wherein the method comprises fusing feature tensors using upsampling or downsampling for each corresponding resolution.

18. A system comprising a neural network, comprising: An input interface for accessing input data of a neural network, the input being an image; and A processor in communication with the input interface, the processor being programmed to: Receive the input data at the neural network, where the input includes a plurality of resolution inputs of varying resolutions; Output a plurality of feature tensors for each corresponding resolution of the plurality of resolution inputs; Fuse the plurality of feature tensors using upsampling or downsampling; Use a balance solver to identify prediction vectors in response to the plurality of feature tensors, where the prediction vectors include features associated with each of the plurality of feature tensors; and Output the prediction vectors as a classification of the input data.

19. The system according to claim 18, wherein the multiple-resolution inputs comprise varying resolutions of an image.

20. The system according to claim 18, wherein, Each of the plurality of feature tensors for each corresponding resolution is sent to a corresponding residual block.

Images