IP Library › Granted Patent US 12,333,796
Granted Patent B2
US 12,333,796 · App. 17/845,732 · Granted Jun 17, 2025

Bayesian compute unit with reconfigurable sampler and methods and apparatus to operate the same

Inventors: Srivatsa Rangachar Srinivasa (Hillsboro, OR); Tanay Karnik (Portland, OR); Dileep Kurian (Portland, OR); Ranganath Krishnan (Hillsboro, OR); Jainaveen Sundaram Priya (Hillsboro, OR); Indranil Chakraborty (West Lafayette, IN)
Assignee: Intel Corporation
G06V10/84G06F7/5443G06V10/82
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,333,796
App. No.
17/845,732
Granted
Jun 17, 2025
Kind
B2
Abstract

Methods, apparatus, systems, and articles of manufacture providing a Bayesian compute unit with reconfigurable sampler and methods and apparatus to operate the same are disclosed. An example apparatus includes a number generator to generate a sequence of numbers; a multiplier to generate a plurality of products by multiplying respective numbers of the sequence of the numbers by a variance value; and an adder to generate a plurality of weights by adding a mean value to the plurality of products, the plurality of weights corresponding to a single probability distribution.

Claims (67)

1. An apparatus for an artificial intelligence-based model, the apparatus comprising:

memory to store a variance value and a mean value, the mean value and the variance value resulting from training of the artificial intelligence-based model; and

a plurality of neurons, each neuron including:

number generator circuitry to generate a first random number and a second random number;

multiplier circuitry to:

generate a first product by multiplying the first random number by the variance value; and

generate a second product by multiplying the second random number by the variance value;

adder circuitry to:

generate a first weight by adding the mean value to the first product; and

generate a second weight by adding the mean value to the second product, the first and second weights corresponding to a single probability distribution;

a first processing element to apply the first weight to a first activation; and

a second processing element to apply the second weight to a second activation.

2. The apparatus of claim 1 , wherein memory includes a first input memory and a second input memory:

the first input memory to store a first part of the first activation and a first part of the second activation;

the second input memory to store a second part of the first activation and a second part of the second activation; and

wherein the second activation is different than the first activation.

3. The apparatus of claim 2 , wherein:

the first processing element is to generate first output values by applying (a) the first weight to the first part of the first activation and (b) second weight to the first part of the second activation;

the second processing element to generate second output values by applying (a) the first weight to the second part of the first activation and (b) the second weight to the second part of the second activation; and

output memory to store the first output values and the second output values.

4. The apparatus of claim 3 , wherein the first activation corresponds to a first forward pass and the second activation corresponds to a second forward pass.

5. The apparatus of claim 3 , wherein the first activation corresponds to a first image in a batch and the second activation corresponds to a second image in the batch.

6. The apparatus of claim 1 , wherein the number generator circuitry is to generate a sequence to have a number of pseudo random numbers, the number corresponding to a total number of activations to be processed by at least one of the first processing element or the second processing element.

7. The apparatus of claim 1 , wherein the first and second processing elements are implemented in a first compute node, the number generator circuitry to generate a sequence to have a number of pseudo random numbers, the number corresponding to a total number of activations to be processed by the first compute node and a second compute node.

8. The apparatus of claim 1 , wherein the mean value is a first mean value and the variance value is a first variance value, further including a mixture model processor to generate samples that correspond to a mixture model based on (a) the first weight and the second weight and (b) a second mean value different from the first mean value and a second variance value different from the first variance value.

9. A non-transitory computer readable medium comprising instructions which cause one or more processors to at least:

cause storage of a variance value and a mean value, the mean value and the variance value resulting from training of an artificial intelligence-based model;

generate a first random number and a second random number;

generate a first product by multiplying the first random number by the variance value;

generate a second product by multiplying the second random number by the variance value;

generate a first weight by adding the mean value to the first product;

generate a second weight by adding the mean value to the second product, the first and second weights corresponding to a single probability distribution;

apply the first weight to a first activation; and

apply the second weight to a second activation.

10. The computer readable medium of claim 9 , wherein the instructions cause the one or more processors to:

cause storage of a first part of the first activation and a first part of the second activation;

cause storage of a second part of the first activation and a second part of the second activation; and

wherein the second activation is different than the first activation.

11. The computer readable medium of claim 10 , wherein the instructions cause the one or more processors to:

generate first output values by applying (a) the first weight the first part of the first activation and (b) the second weight to the first part of the second activation;

generate second output values by applying (a) the first weight to the second part of the first activation and (b) the second weight to the second part of the second activation; and

cause storage of the first output values and the second output values.

12. The computer readable medium of claim 11 , wherein the first activation corresponds to a first forward pass and the second activation corresponds to a second forward pass.

13. The computer readable medium of claim 11 , wherein the first activation corresponds to a first image in a batch and the second activation corresponds to a second image in the batch.

14. The computer readable medium of claim 9 , wherein the instructions cause the one or more processors to generate a sequence to have a number of pseudo random numbers, the number corresponding to a total number of activations to be processed by a compute node.

15. The computer readable medium of claim 9 , wherein the instructions cause the one or more processors to generate a sequence to have a number of pseudo random numbers, the number corresponding to a total number of activations to be processed by a first compute node and a second compute node.

16. The computer readable medium of claim 9 , wherein the mean value is a first mean value and the variance value is a first variance value, the instructions to cause the one or more processors to generate samples that correspond to a mixture model based on (a) the first weight and the second weight and (b) a second mean value different than the first mean value and a second variance value different than the first variance value.

17. A system to implement neural network to apply a plurality of weights in an artificial intelligence-based model, the system comprising:

a first hardware-implemented compute node including:

a first programmable sampling unit to generate a first weight and a second weight, the first and second weights corresponding to a single probability distribution;

a first processing element to generate first output values by applying (a) the first weight to a first part of a first activation and (b) the second weight to the first part of a second activation; and

a second processing element to generate second output values by applying (a) the first weight to a second part of the first activation and (b) the second weight to the second part of the second activation; and

a second hardware-implemented compute node including:

a second programmable sampling unit to generate a third weight and a fourth weight, the third and fourth weights corresponding to the single probability distribution;

a third processing element to generate third output values by applying (a) the third weight to a first part of a third activation and (b) the fourth weight to the first part of a fourth activation; and

a fourth processing element to generate fourth output values by applying (a) the third weight to a second part of the third activation and (b) the fourth weight to the second part of the fourth activation.

18. The system of claim 17 , wherein the first compute node further includes:

a first input memory to store the first part of the first activation and the first part of the second activation; and

a second input memory to store a second part of the first activation and the second part of the second activation.

19. The system of claim 18 , wherein the first compute node further includes:

a first output memory to store the first output values; and

a second output memory to store the second output values.

20. The system of claim 19 , wherein the first activation corresponds to a first forward pass and the second activation corresponds to a second forward pass.

21. The system of claim 19 , wherein the first activation corresponds to a first image in a batch and the second activation corresponds to a second image in the batch.

22. The system of claim 17 , wherein the single probability distribution corresponds to a mean and a variance.

23. The system of claim 17 , wherein the single probability distribution corresponds to a mixture model distribution.

24. The system of claim 17 , wherein the first compute node and the second compute node are to generate the first, second, third, and fourth weights simultaneously.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Aug 14, 2022
From: KARNIK, TANAY; KURIAN, DILEEP; SRINIVASA, SRIVATSA RANGACHAR; PRIYA, JAINAVEEN SUNDARAM; KRISHNAN, RANGANATH; CHAKRABORTY, INDRANIL
To: INTEL CORPORATION
Reel/Frame 060801/0916 →
Continuity (1)
Related Publication 20220319162A1 · Oct 6, 2022
References Cited (21)
US 20180349158A1 · Swersky et al. · 2018 [cited by applicant]
US 20190347551A1 · Lobacheva · 2019 [cited by examiner]
US 20200218982A1 · Annau et al. · 2020 [cited by applicant]
US 20200410364A1 · Willers et al. · 2020 [cited by applicant]
Girshick, “Fast R-CNN,” arXiv:1504.08083v2, in Proceeding IEEE International Conference Computer Vision (ICCV), dated Sep. 27, 2015, 9 pages. [cited by applicant]
Krizhevsky et al., “ImageNet Classification with Deep Convolutional Neural Networks,” Part of Advances in Neural Information Processing Systems 25 (NIPS 2012), 9 pages. [cited by applicant]
Szegedy et al., “Intriguing properties of neural networks,” arXiv:1312.6199v4, dated Feb. 19, 2014, 10 pages. [cited by applicant]
Alcorn et al., “Strike (with) a Pose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects,” arXiv:1811.11553v3, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Apr. 1… [cited by applicant]
Lee et al., “A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks,” arXiv:1807.03888v2, 32nd Conference on Neural Information Processing Systems (NIPS 2018), Montreal, Canada., da… [cited by applicant]
Blum et al., “The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation,” arXiv:1904.03215v3, Sep. 9, 2019, 14 pages. [cited by applicant]
Kendall et al., “Bayesian SegNet: Model Uncertainty in Deep Convolutional Encoder-Decoder Architectures for Scene Understanding,” arXiv:1511.02680v2, dated Oct. 10, 2016, 11 pages. [cited by applicant]
Kendall et al.“What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?,” arXiv:1703.04977v2, 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, Oct. 5, 2017, 12 pa… [cited by applicant]
Krishnan et a., “BAR: Bayesian Activity Recognition using variational inference,” arXiv:1811.03305v2, Third workshop on Bayesian Deep Learning (NeurIPS 2018), Montréal, Canada, dated Dec. 1, 2018, 8 pages. [cited by applicant]
Subedar et al., “Uncertainty-aware Audiovisual Activity Recognition using Deep Bayesian Variational Inference,” arXiv:1811.10811v3, in Proceedings of the IEEE International Conference on Computer Vision, dated Sep. 20, … [cited by applicant]
Kingma et al., “Variational Dropout and the Local Reparameterization Trick,” Part of Advances in Neural Information Processing Systems 28 (NIPS 2015), 9 pages. [cited by applicant]
Wen et al., “Flipout: Efficient Pseudo-Independent Weight Perturbations on Mini-Batches,” arXiv:1803.04386v2, published as a conference paper at ICLR 2018, 16 pages. [cited by applicant]
Cai et al., “VIBNN: Hardware Acceleration of Bayesian Neural Networks,” Session 5B Neural Networks, ASPLOS'18, Mar. 24-28, 2018, Williamsburg, VA, 13 pages. [cited by applicant]
Krishnan et al., “Improving model calibration with accuracy versus uncertainty optimization,” 34th Conference on Neural Information Processing Systems (NeurIPS 2020), Vancouver, Canada, 12 pages. [cited by applicant]
Filos et al., “A Systematic Comparison of Bayesian Deep Learning Robustness in Diabetic Retinopathy Tasks,” arXiv:1912.10481v, Fourth workshop on Bayesian Deep Learning (NeurIPS 2019), Vancouver, Canada, 12 pages. [cited by applicant]
Krishnan et al., “Improving MFVI in Bayesian Neural Networks with Empirical Bayes: a Study with Diabetic Retinopathy Diagnosis,” Fourth workshop on Bayesian Deep Learning (NeurIPS 2019), Vancouver, Canada, 7 pages. [cited by applicant]
Naeini et al., “Obtaining Well Calibrated Probabilities Using Bayesian Binning,” Association for the Advancement of Artificial Intelligence (www.aaai.org), 2015, 7 pages. [cited by applicant]