Neurophysiologist Warren McCulloch and mathematician Walter Pitts wrote a paper on how neurons might work, modeling a simple neural network using electrical circuits. This is seen as the starting point for the concept of Artificial Neural Networks.

Research Paper: A Logical Calculus of the Ideas Immanent In Nervous Activity

Donald Hebb proposed the theory that neural pathways are strengthened each time they are used, a concept fundamental to the ways in which humans learn. This formed the basis of Hebbian learning, a principle used in several neural network models.

Wiki: Hebbian Theory
Book: The Organization of Behavior

Marvin Minsky, who would later become one of the pioneers of artificial intelligence, built the SNARC (Stochastic Neural Analog Reinforcement Calculator), the first neural network machine, suggesting that machines could potentially simulate human intelligence.

Wiki: SNARC
Article

Frank Rosenblatt invented the Perceptron, a type of linear classifier that forms the basis for many neural networks. The Perceptron was the first model that could learn from its mistakes, making it a milestone in the machine learning field.

Research Paper: The Perceptron: A Perceived and Recognizing Automation

Bernard Widrow and Marcian Hoff of Stanford developed models called ADALINE and MADALINE. These models were early ancestors to the modern neural networks and were the first learning machines to use an adaptive filter.

NOTE:
While it's true that Bernard Widrow and Ted Hoff created the ADALINE model (Adaptive Linear Neuron), the MADALINE (Multiple ADALINE) was actually introduced later in 1962.

Wiki: According to Wiki it was in 1960
Standford: According to Standford it wasin 1959
Research Paper: Adaptive Switching Circuits

Cover and Hart introduced the k-nearest neighbor algorithm, a simple but effective classification and regression method. This was one of the first instance-based learning algorithms, where the function is approximated locally and all computation is deferred until classification. It's a non-parametric method which remains widely used today.

Research Paper: Nearest Neighbor Pattern Classification

In their book "Perceptrons," Marvin Minsky and Seymour Papert presented limitations of perceptrons and the Rosenblatt's perceptron learning theorem. They argued that perceptrons couldn't learn an XOR function, which led to a shift in research away from neural networks to symbolic methods for the next decade, commonly referred to as the "AI winter"

PDF Book: Perceptron
Amazon Book: Perceptron

Dunn JC introduced the C-means (Fuzzy C-means) algorithm in 1973. It's a method of clustering that allows data points to belong to multiple clusters with varying degrees of membership. This ability to have "soft" cluster assignments differentiated it from earlier "hard" clustering methods like K-means and opened the door for more flexible machine learning models.

Wiki: Fuzzy Clustering
Research Paper: A Fuzzy Relative of the ISODATA Process and Its Use in Detecting Compact Well-Separated Clusters

The CART (Classification And Regression Trees) algorithm, another important decision tree algorithm, was introduced by Breiman, Friedman, Olshen, and Stone. The CART algorithm improved on earlier decision tree algorithms with the addition of pruning techniques to avoid overfitting, making it more applicable for machine learning tasks.

Wiki: CART (Decision Tree)
Amazon Book: Classification And Regression Trees (1st Edition) 

In 1980, the first conference dedicated solely to machine learning, ICML, was held. This conference marked a major milestone by bringing together professionals and researchers from different fields to discuss advancements and make collaborations in the field of machine learning.

Wiki: List of all ICML Conference

Leslie Valiant proposed the 'Probably Approximately Correct' (PAC) concept in learning theory. The PAC model provides a mathematical definition of learning and was a significant theoretical advance. This approach gave rise to the field of computational learning theory.

Wiki: Probably Approximately Correct Learning
Research Paper: A Theory of the Learnable
Amazon Book: Probably Approximately Correct (1st Edition)

A crucial development in machine learning came about in 1986 with the introduction of the back-propagation algorithm for training multi-layer neural networks. It became a foundational method for training neural networks, published by David Rumelhart, Geoffrey Hinton, and Ronald Williams.

Wiki: Backpropogation

Research Paper: Learning Representations by Back-Propogating Errors

In 1987, a model-free reinforcement learning algorithm was introduced by Chris Watkins. The method, called Q-Learning, would go on to become an essential component in much of the future research in reinforcement learning.

Wiki: Q-Learning
Research Paper: Learning from Delayed Rewards
**Chris Watkins first developed the algorithm in his PhD thesis, which was submitted in 1987. However, the thesis was not published until 1989.

Richard Sutton and Andrew Barto published "Reinforcement Learning: An Introduction", a highly influential book that provided a clear and simple account of the key ideas and algorithms of reinforcement learning. Their work has been instrumental in making reinforcement learning accessible to a broader audience and set the foundation for further developments in the field.

Amazon Book: Reinforcement Learning: An Introduction 

Leo Breiman, a statistician at the University of California, Berkeley, proposed an ensemble method known as Bootstrap Aggregating (or "bagging"). Bagging involves using multiple decision trees and aggregating their outputs to make predictions. This marked a significant advancement in the use of ensemble methods in machine learning, which combine multiple models to achieve better predictive performance than any of the individual models could achieve alone.

Wiki: Bootstrap Aggregation
Research Paper: Bagging Predictors

Sepp Hochreiter and Jürgen Schmidhuber proposed Long Short-Term Memory (LSTM) networks, a kind of recurrent neural network (RNN) designed to avoid the long-term dependency problem. LSTM has been widely used in deep learning for tasks that require remembering information over long periods, such as time series prediction and natural language processing.

Wiki: Long Short Term Memory
Research Paper: Long Short Term Memory

Christopher Bishop, Markus Svensén, and Christopher K. I. Williams proposed the Generative Topographic Mapping (GTM) algorithm, a principled framework for visualizing high-dimensional data. This made it easier for machine learning practitioners to understand and interpret complex datasets, marking a step forward in exploratory data analysis.

Wiki: Generative Topographic Map
Research Paper: GTM: The Generative Topographic Mapping

Leo Breiman proposed the random forest algorithm, an ensemble learning method for classification and regression that operates by constructing a multitude of decision trees. Random forests are a significant development in machine learning, as they provide high prediction accuracy and feature importance estimation, and are robust against overfitting.

Wiki: Random Forest
Research Paper: Random Forests

David M. Blei, Andrew Y. Ng, and Michael I. Jordan introduced a generative probabilistic model for collections of discrete data such as text corpora. LDA is a three-level hierarchical Bayesian model, in which each item of a collection is modeled as a finite mixture over an underlying set of topics.

Wiki: LDA
Research Paper: Latent Dirichlet Allocation

The Netflix Prize was a machine learning and data mining competition for movie rating prediction. It awarded $1 million to the BellKor's Pragmatic Chaos team for improving Netflix's recommendation algorithm's accuracy by over 10%. This competition drew worldwide attention to the capabilities of machine learning algorithms in real-world applications.

About Netflix Prize: Paper

Microsoft launched the Kinect as a motion-sensing input device for the Xbox 360. It used machine learning in real-time to process depth data. Kinect's technology brought computer vision to the consumer market, and the device had widespread implications for how machine learning could be implemented in consumer technology.

Research Paper: Microsoft Kinect Sensor and Its Effect
** There was a research paper written on the working on Kinect in 2012

In a landmark paper, Google researchers demonstrated the power of using large-scale distributed systems to train deep learning models. This paper showcased Google's distributed computing capabilities and how they could be used to train large deep learning models on massive datasets, a trend that continues to this day.

Research Paper: Large Scale Distributed Deep Networks

Tomas Mikolov, et al. from Google introduced word2vec, a group of related models that are used to produce word embeddings - numerical forms of linguistic context. This revolutionized Natural Language Processing (NLP) by enabling machines to understand words in relation to other words, based on the "distance" between word vectors.

Google Research: Word2Vec (Colab)
Research Paper: Efficient Estimation of Word Representations in Vector Space

Google's DeepMind developed AlphaGo, a computer program that defeated world champion Go player Lee Sedol in a five-game match. This was considered a major breakthrough in AI, as the game of Go, with its vast number of potential moves, had long been considered a challenging benchmark for AI.

Google Deep Mind: AlphaGo
Research Paper: Mastering the Game of Go with Deep Neural Networks and Tree Search

OpenAI's GPT (Generative Pretrained Transformer) and Google's BERT (Bidirectional Encoder Representations from Transformers) were released, setting new standards in language modeling. These transformer-based models demonstrated unprecedented performance on a variety of NLP tasks and led to the emergence of large language models as a key area of research in machine learning.

GPT Research Paper: Improving Language Understanding by Generative Pre-Training
BERT Research Paper: BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding

OpenAI introduced GPT-3, a transformer-based language model with 175 billion machine learning parameters. This was a massive leap from its predecessor, GPT-2, which had 1.5 billion parameters. GPT-3 demonstrated remarkable capabilities in generating human-like text, raising both hopes and concerns about the future of AI.

Research Paper: Language Models are Few-Shot Learners