AI Overview

Outline

Machine Learning

Machine learning(ML) is the development of algorithms for machines to infer unknown information from known. Learning algorithms are used to teach robots how to interpret data to make their own decisions. Every ML algorithm is designed to solve a specific task.

Training

Intelligent robots make data-driven predictions, so they must be given the data to drive these decisions. The initial set of data given to a network for learning are massive sets of data that machines analyze to find trends, patterns, and relationships in order to learn.


Once adequately trained, a machine is given a set of data given to a network to evaluate its performance to reveal how well it will perform on unfamiliar data and show what it has learned.


Modern Training Datasets: MNIST, Cornell, and Jacquard

Problem Types

Regression: A continuous function is predicted to relate input variables to output variables

(Ex. Predicting future prices, revenue, or temperature)


Classification: A categorical mapping function is predicted to categorize/classify input variables

(Ex: Spam mail detection)



Formulating the Problem

Supervised Learning


Training data contains input variables with corresponding targets, the machine must determine the relationship between them


Example Algorithms:


  • Neural Network- A series of increasingly complex functions to map from input to target

  • Logistic regression- Predicts classification in a finite setting by modelling probability of each outcome (frequently in a binary setting)


Manufacturing Applications:


  • Modelling dynamic capabilities of a robotic arm

  • Grasping and object manipulation

  • Determining quality

  • Collision detection

  • Scene understanding and analysis


Unsupervised Learning


Training data includes high dimensional inputs without targets, and the machine has to find relationships and structure in just the data

Example Algorithms:

  • Clustering- Divides training data into groups of similar data by analyzing desired qualities

  • Association- Discovers relationships between training data to predict relationships between test data

  • Principal Component Analysis- Compresses data to more easily interpret


Manufacturing Applications:

  • Anticipating physical outcome of situations

  • Establishing inverse dynamics required for a task

  • Determining how actions will impact surrounding environment

Reinforcement Learning


Repeated attempts with a trial and error feedback loop to learn the best method for a task


Example Algorithms:

  • Q Learning- Begins with random actions and adjusts based on accuracy of its output<

  • Temporal Difference- Predicts future reward based on current and past loss

Manufacturing Applications:

  • Complex decision making

  • Ability to adapt to changing environments

Neural Networks

The term neural network is derived from its inspiration, the neural network of the human brain. Neural networks allow machines to intake, analyze, and infer information using various complex layers of functions, equations, and inputs. Activation functions begin the computations by calling hidden layers using parameters to control weights and biases. Activation functions determine the type of output and map to it from the input.





Every neural network learns the most optimal weights and biases to minimize its overall loss based on a cost function. Cost is the amount of error between the actual output and the expected output. The gradient descent of the cost function tells us how to best minimize cost.

Backpropagation is an algorithm that calculates this gradient by finding and averaging the changes that must be made to the weights and biases. This algorithm feeds forward through the network to compute the output, back propagates to compute the error at each layer, and then updates the weights and biases accordingly to minimize cost.







Modern Architectures

Convolutional Neural Networks (CNN)

Specializing in object recognition from images, convolutional layers contain filters of increasing complexity to develop a feature map highlighting key features of a given image


Modern Models


-AlexNet (2012)- First CNN to win Image Net for object recognition by using multiple GPU's for training. This method allowed for a much larger model to be trained at the fastest speeds of its time.

-VGG (2015)- Based off of AlexNet with deeper layers and smaller filters

-ResNet(2016)- An even larger improvement of AlexNet using residual connections to allow the gradient to flow through a network directly


Manufacturing Applications


-Antipodal Robotic Grasping using Generative Residual Convolutional Neural Network

-Memory Efficient Grasping Point Detection of Nontrivial Objects


Recurrent Neural Networks (RNN)

Recurrent Neural Networks(RNN) Develops a form of memory to specialize in processing sequential data and images to make predictions



Modern Models

-Recursion- Generalized RNN that uses a tree structure rather than a chain, specifically good for natural language processing and computer vision

-Long Short Term Memory- Combats vanishing gradient using cell state gates (memory) to determine relevance of data using a sigmoid function, ensuring important information is remembered irrelevant data is discarded


-Gated Recurrent Units(GRU)-Combats vanishing gradient using a hidden state containing an update gate and reset gate to determine relevancy and reconfigure accordingly


Manufacturing Applications

-DeepMPC: Learning Deep Latent Features for Model Predictive Control

-Seamless Integration and Coordination of Cognitive Skills in Humanoid Robots: A Deep Learning Approach

-Brain4Cars: Car That Knows Before You Do via Sensory-Fusion Deep Learning Architecture





Resources