VSD - Machine Intelligence in EDA/CAD

Explores machine intelligence in design automation, covering timing, noise, and verification. Classifies learning by input, architecture, and output, with supervised/unsupervised/reinforcement, neural networks, deep learning, and regression, classification, clustering, dimensionality reduction.

Discover the architecture of machine intelligence in eda/cad, contrasting rule-based programming with machine learning and deep learning. Explore neural networks like feedforward, cnn, gans, and applications in verification and speech.

Master the MI foundation for machine intelligence in EDA/CAD by aligning mathematics with Python tooling, and learn to set up a Colaboratory lab for Python programming.

Learn python scripting for machine intelligence in eda/cad using Google Colab notebooks and cells. Explore lists, tuples, dictionaries, strings, slicing, indexing, type casting, and indentation with exercises on vector multiplication.

Explore how TensorFlow builds a compute graph with nodes and arcs, connect inputs, and run a session to perform operations and produce outputs like 20.

Participants explore TensorFlow sessions, computation graphs, and fetch list with feed dictionaries to drive execution in the live Q&A. They highlight TensorFlow's parallelism, graph optimization, and deep learning workflows.

Apply regression to estimate wire resistance from design automation data using linear, polynomial, and multinomial models. Emphasize dataset normalization to align feature scales for reliable training and validation.

Explore a simple machine learning model with a weight and bias, trained by gradient descent to minimize a loss function across 8000 samples, accounting for noise and outliers.

Live q&a and labs explore the gradient descent algorithm, its learning rate, convergence behavior, and the impact of feature selection and dimensionality on model training.

The lecture explains overflow and uninitialized issues in linear regression. It covers learning rate tuning, data normalization, and binary, grid, and random search for hyperparameters in ml solution flow.

Train a linear regression model to estimate resistance using normalized data, 80/20 train-test split, placeholders and variables, gradient descent optimization, and export weights and bias for deployment.

Forecast resistance values by normalizing inputs, applying the model, and denormalizing outputs via training data. Improve accuracy by using all three features and exploring polynomial regression to address underfitting.

Learn to apply supervised learning for regression and resistance estimation by selecting features, understanding training versus validation data, and leveraging domain expertise to improve model accuracy.

Explore how underfitting and overfitting affect model generalization, and learn strategies to improve accuracy by adjusting model complexity, features, data collection, and regularization.

Explore the wire error model and underfitting through live Q&A, contrasting training and validation accuracy with cost, and discussing timing analysis and pessimistic predictions in EDA/CAD.

Investigate how machine learning can improve parasitics extraction and RC estimation, discuss open problems, and propose ideas like adversarial networks inside Virtuoso to boost productivity.

Learn wire capacitance estimation (WiCE) using machine learning with normalized features like length and thickness, exploring loss functions such as mean squared error and hands-on labs.

Predict capacitance with a three-variable model, visualize predictions in a 3D plot with color representing thickness, and ensure reproducibility by using seeds.

Conducts a live q&a on WiCE, illustrating how dimensionality from hundreds to two thousand features affects model training and evaluation, and discusses training cost versus validation to gauge potential.

Explore classification as a supervised learning task, compare decision boundaries across logistic regression, SVM, and neural networks, and apply binary classification to design and anomaly detection in EDA/CAD.

Explore decision boundaries, contrasting linear and nonlinear classification, through the VLSI cell classification dataset of 12,000 cells labeled sequential or combinational; apply logistic regression and visualize high-dimensional patterns in 3D.

Explore logistic regression as a simple classifier using the sigmoid to produce probabilities, define decision boundaries, and contrast cross-entropy with mean squared error in the VCC model.

Explore binary cell classification with logistic regression, using 1755-dimension feature vectors, random data shuffling, and softmax with cross-entropy loss; evaluate with accuracy and false positives/negatives analysis.

Learn how a confusion matrix uses true positives, true negatives, false positives, and false negatives to evaluate model performance, highlighting accuracy pitfalls in imbalanced data.

Understand how the support vector machine solves complex separation by lifting data into higher dimensions to reveal a linear boundary, the largest margin classifier, for machine intelligence in EDA/CAD.