Suggestions
Kapil Sachdeva
Fellow of Technical Staff at HID Global
Kapil Sachdeva is a highly skilled professional currently serving as a Fellow of Technical Staff at HID Global.13 With a LinkedIn username of ksachdeva, he has established himself as a passionate and dedicated engineer in his field.1
Professional Background
Kapil's expertise spans several areas of technology, including:
- Language Runtimes
- Bayesian Statistics
- Machine Learning
- Cryptography
- Embedded Systems2
His role at HID Global involves working as a Principal Engineer focused on Secure Firmware.5 This position allows him to apply his extensive knowledge in various technological domains.
Characteristics and Work Ethic
Colleagues describe Kapil as a passionate engineer who invests significant thought and energy into his designs.1 His curious nature drives him to stay current with the latest developments in his field, ensuring that his work remains cutting-edge and innovative.
Location and Contact Information
Kapil Sachdeva is based in Austin, TX, United States.3 While specific contact details are not provided here, he can be reached through professional networking platforms using his LinkedIn username, ksachdeva.
Recent Activities
Kapil has been active in sharing his knowledge and insights on professional platforms. In April 2023, he made a post related to deep learning, variational autoencoders (VAE), and machine learning, demonstrating his ongoing engagement with advanced topics in artificial intelligence and data science.4
Kapil Sachdeva's role as a Fellow of Technical Staff at HID Global, combined with his diverse expertise and passionate approach to engineering, positions him as a valuable asset in the technology sector, particularly in areas requiring advanced technical knowledge and innovative problem-solving skills.
Highlights
This tutorial provides an in-depth and visual explanation of the three famous Bounding Box loss functions using the simulated bounding box regression problem.
The simulation experiment (1,715,000 regression cases) is implemented using the new PyTorch functional API to compute per-sample gradients.
Colab notebook: https://lnkd.in/gdvJgZUm
#objectdetection #deeplearning #machinelearning #datascience https://lnkd.in/gJhJsAiN
Quite often VAE is explained as - “Oh, it is just an Auto-encoder with latent variable distribution!”
While the above explanation is not inaccurate, it kind of undermines the primary reason it came into existence. These four tutorials (in sequence) should be helpful if you want to dig a bit deeper.
Tutorial 1 - What is KL Divergence? https://lnkd.in/gjMSuUGv
- What does it mean to compare two probability distributions?
- Why do we have a “log” in the formulation of KL Divergence?
- Why should we see it as an expected value of log likelihood ratio?
- Why is it called “Divergence” and not “Distance”?
- What are the two types of KL divergence?
- What are the different behaviors (mode vs mean seeking) exhibited by different types of KL divergence?
Tutorial 2 - What is Evidence Lower Bound (ELBO)? https://lnkd.in/gfVydAqF
- In variational inference, we approximate the intractable posterior distribution using optimization. Optimization requires a loss function. Could we use KL Divergence?
- Why KL Divergence (as it is) would not be usable?
- Step by Step simplification of KL divergence into different components to see if there is something else we could leverage or pose our question in a different but “somewhat” equivalent/constrained form.
- What is log evidence and why is it called Lower bound on log evidence?
- Why components of ELBO can be seen as Reconstruction Loss and KL divergence between our approximated posterior and the prior?
Tutorial 3 - What is a Variational Auto-encoder? https://lnkd.in/grhMpqkj
- What is amortized inference?
- What is a Latent Variable Model and computational challenges associated with it?
- Neural networks (the function approximators) for amortization inference?
- Why predict the parameters of distribution of Latent (Random) Variables?
- A neural network (called Encoder) to predict the parameters of Latent Variable.
- How to learn the weights of Encoder neural networks? What is the loss function? KL? ELBO?
- Why add an additional neural network (called Decoder) to get the reconstruction term? To get reconstruction term in ELBO?
Tutorial 4 - What is a Re-parameterization trick? https://lnkd.in/gceB4S2Q
- Why is the stochastic layer in VAE problematic? High variance in gradient computation?
- A deeper explanation of various terms and symbols in the ELBO used in VAE.
- Why gradient computation for the encoder and decoder network using ELBO is difficult? Why Leibniz integral rule can’t be applied?
- Two concepts to rescue -> Change of Variable -> Law Of The Unconscious Statistician (LOTUS)
- How to use change of variable and LOTUS in externalizing the randomness when doing gradient computation?
- Why is it called the Re-parameterization trick?
#deeplearning #vae #machinelearning
