Machine Learning & Foundation Models

The theorem, describing how to update the probability of a hypothesis as new evidence arrives, provided a mathematical foundation for the probabilistic and statistical methods that much of machine learning, from spam filters to modern neural networks, would later build upon.

The pair proposed algorithms mirroring human thought processes, among the earliest formal models of neural computation. Their work on human cognitive systems contributed to modern machine learning technologies.

The Canadian psychologist described a theoretical neural structure formed by interactions among nerve cells. This set the groundwork for how many machine learning algorithms work, with connected artificial neurons changing the strength of their connections based on data.

His paper on machine intelligence, better known for proposing what became called the Turing test, also sketched the idea of a machine that started simple and developed its abilities through a training process, an early statement of the learning-machine concept.

The IBM employee built a computer program that calculated the chance of winning in checkers for each side, widely regarded as the first machine learning program. It demonstrated that a program could improve at a task through experience.

Working at IBM, Samuel wrote a program that adjusted its own evaluation of board positions based on the outcomes of games it played against itself, an early demonstration of a system that improved its performance through experience rather than explicit reprogramming.

Cybertron analysed sonar signals, electrocardiograms and speech patterns using rudimentary reinforcement learning. It was repetitively trained by a human operator and had a "goof" button to make it reevaluate incorrect decisions.

The book dealt mostly with machine learning for pattern classification, summarising the state of the field. Interest in pattern recognition continued into the 1970s, as described by Duda and Hart in 1973.

Their work established mathematical conditions under which a learning system could be expected to generalise from training examples to new data, theory that later underpinned support vector machines and much of the formal analysis of machine learning methods.

Their book showed such networks could not represent certain simple functions, and the finding, combined with the hardware limits of the time, contributed to a sharp fall in funding for neural network research that lasted into the 1980s.

Their account described how research interest related to pattern recognition carried on through the 1970s. This continuity linked earlier learning-machine work to later statistical approaches.

By 1980 symbolic expert systems had come to dominate AI research, marginalising statistical approaches. This shift shaped the trajectory of machine learning before its later statistical revival.

Genetic algorithms were used as models in machine learning during these two decades. They represented an evolutionary alternative to gradient-based learning methods.

The program inductively inferred logic programs from positive and negative examples. It was an early implementation of inductive logic programming as a machine learning approach.

A report described an artificial neural network learning to recognise 26 letters, 10 digits and 4 special symbols from a computer terminal. It was an early demonstration of trained neural character recognition.

The crossbar adaptive array (CAA) was a neural network capable of self-learning, establishing self-learning as a machine learning paradigm. It extended learning beyond externally supervised training.

Their paper showed how errors could be propagated backward through a network's layers to adjust its internal parameters efficiently, providing the practical training method that almost all subsequent neural network research and applications relied upon.

The commission's finding concerned a medical school's computerised selection process. It is cited as an early case of algorithmic bias arising from data-driven decision rules.

Separating from mainstream AI, machine learning re-embraced statistics and optimisation. The decade marked its emergence as an independent discipline.

Their architecture solved a long-standing problem in training neural networks on sequential data, in which useful signals from early in a sequence were lost before reaching later layers, and it became the standard approach for speech and language tasks for the following two decades.

The method showed that neural networks with many layers could be trained effectively by first training each layer separately, helping revive research interest in deep neural networks after two decades in which shallower methods had dominated.

The media-services provider offered a prize for a program that improved the accuracy of its Cinematch recommendation algorithm by at least 10%. The contest drew wide research participation and popularised competitive machine learning benchmarking.

The dataset and its annual competition gave researchers a common, large-scale test of progress, and the scale of labelled data it provided proved necessary for the deep learning methods that would later dominate the field.

The joint team built an ensemble model that met the 10% accuracy threshold, winning $1 million. The result demonstrated the power of combining many predictors in recommendation systems.

An article in The Wall Street Journal noted the firm's use of machine learning in finance. It marked growing public attention to machine learning in investment prediction.

The gathering, following Bletchley, Seoul and Paris, brought together governments, companies and research bodies. Discussion centred on widening access to training and inference compute for researchers outside the small group of countries hosting the largest data centres.

Naver's HyperCLOVA X and LG's Exaone family continued development as part of a national push for AI systems trained on Korean-language data and hosted on domestic infrastructure, reducing reliance on foreign providers for public-sector and consumer applications.

The operators reportedly directed the model to carry out reconnaissance and exploitation steps with limited human review, according to the developer that detected and disrupted the activity, illustrating how agentic AI systems could accelerate intrusion operations.

The release merged fast-response and extended-reasoning modes into a single system that allocated more computation to harder problems automatically, intended to simplify the previous practice of choosing between separate model variants for different tasks.

Backed by the Public Investment Fund, the venture set out to build data centres, cloud services and Arabic-language models, part of a wider Gulf push to convert oil revenue into large-scale AI computing capacity.

The gathering brought together African governments, companies and international partners to discuss compute access, skills and locally relevant applications, part of a wider effort to ensure the continent was not left solely as a consumer of AI developed elsewhere.

Announced at the Paris AI Action Summit, the plan aimed to combine public and private funding toward AI data centres and computing capacity across the EU, responding to concern that Europe lagged the US and China in large-scale AI infrastructure.

The open-weight model matched leading closed systems on several reasoning benchmarks at a reported fraction of the usual training cost, prompting investors to question the scale of spending committed by major AI developers and their hardware suppliers.

The specification let developers build a common interface between language models and outside systems such as databases, file systems and other software, and was subsequently adopted by competing AI developers and toolmakers beyond Anthropic's own products.

The Google DeepMind researchers were recognised for developing a machine learning system able to predict the three-dimensional structure of proteins from their amino acid sequence, a capability subsequently applied widely across biological and pharmaceutical research.

The research surveyed AI-focused faculty across several universities worldwide. The figure was cited as evidence of demographic imbalance among those building machine learning systems.

Machine learning technology was applied to help make diagnoses and support the search for a cure. The pandemic accelerated the deployment of machine learning in medicine.

The publisher released a book generated with machine learning tools. It marked an early commercial use of automated text generation in scholarly publishing.

By 2019 GPUs, often with AI-specific enhancements, had become the standard hardware for training commercial cloud AI models. This hardware shift was central to the feasibility of deep learning.

The vehicle's perception system failed to detect a pedestrian, who died after the collision. The incident highlighted the safety limitations of machine learning in autonomous systems.

The bot rapidly picked up racist and sexist language from its interactions and was withdrawn. It became a standard example of the risks of learning from uncurated public data.

TPUs were purpose-built accelerators for machine learning workloads. Since their introduction they have become a key component of AI infrastructure, especially in cloud environments.

Go had been considered far harder for computers than chess owing to its vast number of possible positions. AlphaGo combined deep neural networks with tree search and reinforcement learning from self-play, a combination that shaped much subsequent work on game-playing and decision-making systems.

The image classification error caused public controversy over bias in machine learning systems. The gorilla label was removed, and as of 2023 the system still could not recognise gorillas.

The reported study analysed paintings and may have revealed previously unrecognised influences among artists. It illustrated machine learning's spread into the humanities.

GANs pitted two networks against each other to produce realistic synthetic data. The architecture opened up generative modelling as a major machine learning capability.

The Sun Microsystems co-founder forecast that automated machine learning medical diagnostic software would displace most physicians within two decades. The claim typified high expectations for medical machine learning.

OpenAI estimated the hardware compute used in the largest deep learning projects from AlexNet in 2012 to AlphaZero in 2017 and found a 300,000-fold increase, with a doubling time of 3.4 months. This scaling underpinned the deep learning era.

Advances in both algorithms and computer hardware enabled the training of networks with many layers of nonlinear hidden units. This combination made deep learning practical and allowed it to surpass earlier machine learning approaches.