Topics

AI Agents

We will cover the agent-environment interface, the rational, learning agent architectures and a practical example of a robotic agent.

Pantelis Monogioudis

Oct 20, 2022

We provide a historical perspective on AI development and the role of rules in mission-critical systems.

Pantelis Monogioudis

Oct 20, 2021

Data Science 360

What are the disciplines that we need to cross fertilize to get a system that possesses intelligence?

The four approaches towards AI

Agents

Ultimately AI will be a cloud of reasoning systems.

Pantelis Monogioudis

Feb 21, 2022

The Learning Problem

Statistical Learning Theory

The supervised learning problem statement.

Introduction to Scene Understanding

In the previous chapters we have treated the perception subsystem mainly from starting the first principles that govern supervised learning to the deep learning…

Object Detection & Semantic Segmentation Workshop

Mask R-CNN: A detailed guide with Detectron2:

Welcome to the Mask R-CNN with Detectron2 tutorial!

Mask R-CNN Demo

A quick intro to using the pre-trained model to detect and segment objects.

Detectron2 Beginner’s Tutorial

Welcome to detectron2! This is the official colab tutorial of detectron2. Here, we will go through some basics usage of detectron2, including the following: * Run inference…

Mask R-CNN Semantic Segmentation

The semantic segmentation approach described in this section is Mask R-CNN based on this paper paper. Mask R-CNN is an extension of Faster R-CNN that adds a mask head to the…

Mask R-CNN - Inspect Training Data

Inspect and visualize data loading and pre-processing code.

Mask R-CNN - Inspect Trained Model

Code and visualizations to test, debug, and evaluate the Mask R-CNN model.

Mask R-CNN - Inspect Weights of a Trained Model

This notebook includes code and visualizations to test, debug, and evaluate the Mask R-CNN model.

UNet Semantic Segmentation

Finetuning Language Models for Text Classification - Patent Dataset

This notebook was submitted by NYU student Sky Achitoff

Pantelis Monogioudis

This is a test

Natural Language Processing

“You shall know a word by the company it keeps” (J. R. Firth 1957: 11) - many modern discoveries are in fact rediscoveries from other works sometimes decades old. NLP is…

Language Models Workshop

The following notebook is a from-scratch attempt on character-level language modeling. Its instructive for you to go through it first and then go through the corresponding…

CNN Language Model

The following was developed by Harini Appansrinivasan, NYU as part of an assignment submission.

Language Models

These notes heavily borrowing from the CS224N set of notes on Language Models.

LSTM Language Model from scratch

This notebook was borrowed from Christina Kouridis’ github. The notation is different than the notation used in the LTSM section of the notes and will be changed in a next…

Simple RNN Language Model

Our aim is to predict the next character given a set of previous characters from our data string. For our RNN implementation, we would take a sequence of length 25…

Transformer-based LLMs

The aim here is not to repeat information that has been published for state of the art (SOTA) large language models elsewhere but to simply connect the decoder-based…

Introduction to NLP Pipelines

In this chapter, we will introduce the topic of processing with neural architectures language in general. This includes natural language, code etc.

Tokenization

nuxxpxjelstl Warning If this notebook seems half done, its because it is. I am still working on it. I will update it based on…

Word2Vec Embeddings

In the so called classical NLP, words were treated as atomic symbols, e.g. hotel, conference, walk and they were represented with on-hot encoded (sparse) vectors e.g.

Word2Vec from scratch

This self-contained implementation is instructive and you should go through it to understand the word2vec embedding.

Word2Vec Tensorflow Tutorial

word2vec is not a singular algorithm, rather, it is a family of model architectures and optimizations that can be used to learn word embeddings from large datasets.…

RNN-based Neural Machine Translation

These notes heavily borrowing from the CS229N 2019 set of notes on NMT.

NMT Metrics - BLEU

In 2002, IBM researchers developed the BiLingual Evaluation Understudy (BLEU) metric that remains, with its many variants to this day, one of the most quoted metrics for…

Attention in RNN-based NMT

When you hear the sentence “the soccer ball is on the field,” you don’t assign the same importance to all 7 words. You primarily take note of the words “ball” “on,” and “field…

Character-level recurrent sequence-to-sequence model

Author: fchollet
Date created: 2017/09/29
Last modified: 2020/04/26
Description: Character-level recurrent sequence-to-sequence model.

The Annotated Transformer

Attention is All You Need

Multi-head self-attention

Earlier we have seen examples with the token bear being in multiple grammatical patterns that also influence its meaning. To capture such multiplicity we can use multiple…

Positional Embeddings

In the RNN architectures,the decoder state at time step \(t\) was a function of the decoder state at time step \(t-1\) and the input token at time step \(t\). In other…

Scaling

import numpy as np
import matplotlib.pyplot as plt

# Creating an 8-element numpy vector with random gaussian values
# vector = np.random.randn(8)
vector = np.array([0.17148…

Single-head self-attention

In the simple attention mechanism, the attention weights are computed deterministically from the input context. We call the combination of context-free embedding (eg…

Transformers and Self-Attention

For the explanation of decoder-based architectures such as those used by GPT, please see the repo https://github.com/pantelis/femtotransformers and the embedded comments…

Automated Reasoning

We have seen in an earlier chapter where we introduced a dynamical system governing the state evolution of the environment that a state is composed of variables and such fact…

Logical Agents

In this chapter we see how agents equipped with the ability to represent internally the state of the environment and reason about the effectiveness of possible actions using…

Logical Inference

The wumpus world despite its triviality, contains some deeper abstractions that are worth summarizing.

World Models

For each problem we can define a number of world models each representing every possible state (configuration) of the environment that the agent may be in.

Automated Planning

Planning combines two major areas of AI: logic and search.

The Unified Planning Library

In this demo we will scratch the surface of the UP: we will set it up, we manually create the blocksworld domain using the UP API, we create a problem using a bit of python…

Planning Domain Definition Language (PDDL)

In the chapter of propositional logic we have seen the combinatorial explosion problem that results from the need to include in the reasoning / inference step all possible…

Logistics Planning in PDDL

We will use the Logistics domain to illustrate how to represent a planning task in PDDL.

Manufacrturing Robot Planning in PDDL

This is a real case that we tackled for a manufacturing company. This company devises supply chains to make pieces of medical equipments. A supply chain consists of…

The A* Algorithm

Dijkstra’s algorithm is very much related to the Uniform Cost Search algorithm and in fact logically they are equivalent as the algorithm explores uniformly all nodes that…

Forward Search Algorithms

If you are missing some algorithmic background, afraid not. There is a free and excellent book to help you with the background behind this chapter. Read Chapters 3 and 4 for…

Planning with Search

In the PDDL section we saw that a sequence of actions that the agent needs to execute to reach the goal can be obtained using domain-independent planners. This section…

\(A^*\) Interactive Demo

This demo is instructive of the various search algorithms we will cover here. You can introduce using your mouse obstacles in the canvas and see how the various search…

Bellman Expectation Backup

In this section we describe how to calculate the value functions by establishing a recursive relationship similar to the one we did for the return. We replace the…

Bellman Optimality Backup

Now that we can calculate the value functions efficiently via the Bellman expectation recursions, we can now solve the MDP which requires maximize either of the two…

Policy Iteration

In this section we develop dynamic programming algorithms for the so called planning problem (which is RL without learning) where we are dealing with a perfectly known MDP.

Policy Iteration

# Uncomment to run the code locally 
# !git clone https://github.com/dennybritz/reinforcement-learning.git reinforcement_learning

Cloning into…

Value Iteration

We already have seen that in the Gridworld example in the policy iteration section , we may not need to reach the optimal state value function \(v_*(s)\) to obtain an…

Markov Decision Processes

Many of the algorithms presented here like policy and value iteration have been developed in older repos such as this and this. This site is being migrated to be compatible…

Introduction to MDP

We start by reviewing the agent-environment interface with this evolved notation and provide additional definitions that will help in grasping the concepts behind DRL. We…

Introduction to MDP

We start by reviewing the agent-environment interface with this evolved notation and provide additional definitions that will help in grasping the concepts behind DRL. We…

Jack’s Car Rental

https://github.com/zy31415/jackscarrental

Cleaning Robot - Deterministic MDP

Cleaning Robot - Stochastic MDP

The following code shows the estimation of the q value function for a policy, the optimal q_star and the optimal policy for the cleaning robot problem in the strochastic case.

Finding optimal policies in Gridworld

POMDP Example

source

The recycling robot.

Finite State Machine of a a recycling robot and MDP dynamics LUT

Optimal Capacity Control

In this section we outline a capacity control policy that is routinely used in various industries (airlines, car rentals, hospitality) to make reservations towards a…

Policy Evaluation (Prediction)

The policy \(\pi\) is evaluated when we have produced the state-value function \(v_\pi(s)\) for all states. In other words when we know the expected discounted returns that…

Policy Improvement (Control)

In the policy improvement step we are given the value function and simply apply the greedy heuristic to it.

Reinforcement Learning

Different Approaches to solve known and unknown MDPs

Generalized Policy Iteration

As we saw in the dynamic programming (DP) solution MDP problem, policy iteration is an algorithm that consists of two simultaneous, interacting processes: one making the…

\(\epsilon\)-greedy Monte-Carlo (MC) Control

In this section we outline methods that can result in optimal policies when the MDP is unknown and we need to learn its underlying functions / models - also known as the mode…

SARSA Gridworld Example

SARSA Gridworld

The SARSA Algorithm

SARSA implements a \(Q(s,a)\) value-based GPI and naturally follows as an enhancement from the \(\epsilon-greedy\) policy improvement step of MC control.

Policy Gradient - Pong Game

""" Trains an agent with (stochastic) Policy Gradients on Pong. Uses OpenAI Gym. """
import numpy as np
import cPickle as pickle
import gym

# hyperparameters
H = 200 #…

Policy Gradient Algorithms - REINFORCE

Given that RL can be posed as an MDP, in this section we continue with a policy-based algorithm that learns the policy directly by optimizing the objective function and can…

Monte-Carlo Prediction

In this chapter we find optimal policy solutions when the MDP is unknown and we need to learn its underlying value functions - also known as the model free prediction…

Example of \(Q(s,a)\) Prediction

Suppose an agent is learning to play the toy environment shown above. This is a essentially a corridor and the agent has to learn to navigate to the end of the corridor to…

Temporal Difference (TD) Prediction

If one had to identify one idea as central and novel to reinforcement learning, it would undoubtedly be temporal-difference(TD) learning. TD learning is a combination of…

Welcome !

Learn the concepts and engineer AI agents with real-time perceptive and language understanding abilities.

Topics

AI Agents

A systems approach to AI

The four approaches towards AI

Rules, rule the world

Data Science 360

The four approaches towards AI

The Learning Problem

Introduction to Scene Understanding

Object Detection & Semantic Segmentation Workshop

Mask R-CNN: A detailed guide with Detectron2:

Mask R-CNN Demo

Detectron2 Beginner’s Tutorial

Mask R-CNN Semantic Segmentation

Mask R-CNN - Inspect Training Data

Mask R-CNN - Inspect Trained Model

Mask R-CNN - Inspect Weights of a Trained Model

UNet Semantic Segmentation

Finetuning Language Models for Text Classification - Patent Dataset

This is a test

Natural Language Processing

Language Models Workshop

CNN Language Model

Language Models

LSTM Language Model from scratch

Simple RNN Language Model

Transformer-based LLMs

Introduction to NLP Pipelines

Tokenization

Word2Vec Embeddings

Word2Vec from scratch

Word2Vec Tensorflow Tutorial

RNN-based Neural Machine Translation

NMT Metrics - BLEU

Attention in RNN-based NMT

Character-level recurrent sequence-to-sequence model

The Annotated Transformer

Multi-head self-attention

Positional Embeddings

Scaling

Single-head self-attention

Transformers and Self-Attention

Automated Reasoning

Logical Agents

Logical Inference

World Models

Automated Planning

The Unified Planning Library

Planning Domain Definition Language (PDDL)

Logistics Planning in PDDL

Manufacrturing Robot Planning in PDDL

The A* Algorithm

Forward Search Algorithms

Planning with Search

\(A^*\) Interactive Demo

Bellman Expectation Backup

Bellman Optimality Backup

Policy Iteration

Policy Iteration

Value Iteration

Markov Decision Processes

Introduction to MDP

Introduction to MDP

Jack’s Car Rental

Cleaning Robot - Deterministic MDP

Cleaning Robot - Stochastic MDP

Finding optimal policies in Gridworld

POMDP Example

The recycling robot.

Optimal Capacity Control

Policy Evaluation (Prediction)

Policy Improvement (Control)

Reinforcement Learning

Generalized Policy Iteration

\(\epsilon\)-greedy Monte-Carlo (MC) Control

SARSA Gridworld Example

The SARSA Algorithm

Policy Gradient - Pong Game