A digital media service (DMS) is an online service provider that sells access to digital library of content such as films, software, games, images, literature, etc. While no transfer of property is made, a nearly perfect duplicate of the data (song movie, etc.) is made on a customer's computer. Content is either primarily hosted on a dedicated server, which is owned by the service provider, or it is hosted primarily on the hard drives of its customers using a P2P protocol with, perhaps, a dedicated server to supplement. == History == One example of the older business model is the iTunes Store, which still markets and prices data as individual retail products. There are no examples of the latter business model in operation yet, but one is currently in development by Global Gaming Factory X and expected to begin operation some time after they acquire The Pirate Bay domain on August 27, 2009. A key difference between the two models is that the model which relies on its customer base for offering their bandwidth for other customers to access customer hosted data can operate at significantly lower costs than a company that seeks to limit data access to a per-download fee in order to supplement the cost of using its own hosting and bandwidth. The P2P model holds the potential for companies to offer unlimited access to the largest data library in the history of the internet to its customers for a reasonably low membership rate that is relevant to the cost of operation. While the market is virtually untouched, the P2P supplemented model will need entrepreneurs who are able to overcome a series of challenges in order to compete with the older business model as well as that which is offered for free (and often against the wishes of copyright holders) by hundreds of P2P communities on the internet. These challenges include, but are not limited to: Offering better data quality, speed, convenience and ease of use, protocol, sense of security, indexing and search organization, site up time, data library size, customer support, advertising, artist/copyright holder incentives and compensation, incentives and compensation for customers hosting data and providing bandwidth, guaranteed seeding (available access to indexed data at all times), than competitors.
Sparrow (chatbot)
Sparrow is a chatbot developed by the artificial intelligence research lab DeepMind, a subsidiary of Alphabet Inc. It is designed to answer users' questions correctly, while reducing the risk of unsafe and inappropriate answers. One motivation behind Sparrow is to address the problem of language models producing incorrect, biased or potentially harmful outputs. Sparrow is trained using human judgements, in order to be more “Helpful, Correct and Harmless” compared to baseline pre-trained language models. The development of Sparrow involved asking paid study participants to interact with Sparrow, and collecting their preferences to train a model of how useful an answer is. To improve accuracy and help avoid the problem of hallucinating incorrect answers, Sparrow has the ability to search the Internet using Google Search in order to find and cite evidence for any factual claims it makes. To make the model safer, its behaviour is constrained by a set of rules, for example "don't make threatening statements" and "don't make hateful or insulting comments", as well as rules about possibly harmful advice, and not claiming to be a person. During development study participants were asked to converse with the system and try to trick it into breaking these rules. A 'rule model' was trained on judgements from these participants, which was used for further training. Sparrow was introduced in a paper in September 2022, titled "Improving alignment of dialogue agents via targeted human judgements"; however, the bot was not released publicly. DeepMind CEO Demis Hassabis said DeepMind is considering releasing Sparrow for a "private beta" some time in 2023. == Training == Sparrow is a deep neural network based on the transformer machine learning model architecture. It is fine-tuned from DeepMind's Chinchilla AI pre-trained large language model (LLM), which has 70 Billion parameters. Sparrow is trained using reinforcement learning from human feedback (RLHF), although some supervised fine-tuning techniques are also used. The RLHF training utilizes two reward models to capture human judgements: a “preference model” that predicts what a human study participant would prefer and a “rule model” that predicts if the model has broken one of the rules. == Limitations == Sparrow's training data corpus is mainly in English, meaning it performs worse in other languages. When adversarially probed by study participants it breaks the rules 8% of the time; however, this is still three times lower than the baseline prompted pre-trained model (Chinchilla).
Fei-Fei Li
Fei-Fei Li (Chinese: 李飞飞; pinyin: Lǐ Fēifēi; born July 3, 1976) is a Chinese-born American computer scientist best known for establishing ImageNet, the dataset that enabled rapid advances in computer vision in the 2010s. She is a professor of computer science at Stanford University, with research expertise in artificial intelligence, machine learning, deep learning, computer vision, and cognitive neuroscience. Li is a co-director of the Stanford Institute for Human-Centered Artificial Intelligence and a co-director of the Stanford Vision and Learning Lab, and served as Chief Scientist of AI/ML at Google Cloud and the director of the Stanford Artificial Intelligence Laboratory from 2013 to 2018. In 2017, she co-founded AI4ALL, a nonprofit organization working to increase diversity in the field of artificial intelligence. In 2023, Li was named one of the Time 100 AI Most Influential People. Li received the Intel Lifetime Achievements Innovation Award in 2017 for her contributions to artificial intelligence, and was elected member of the National Academy of Engineering, the National Academy of Medicine in 2020 and the American Academy of Arts and Sciences in 2021. In 2025, she was named as one of the "Architects of AI" for Time's Person of the Year. On August 3, 2023, Li was appointed to the United Nations Scientific Advisory Board, established by Secretary-General Antonio Guterres. In 2024, Li was included on the Gold House's most influential Asian A100 list. In 2024, she raised $230 million for a startup called World Labs, which she and three colleagues founded to develop a "spatial intelligence" AI technology that can understand how the three-dimensional physical world works. In 2026, World Labs raised $1 Billion. == Early life and education == Li was born in Beijing, China, in 1976 and grew up in Chengdu, Sichuan. She studied at Sichuan Chengdu No.7 High School. When she was 12, her father immigrated to Parsippany, New Jersey. When she was 16, Li and her mother joined him in the United States. While attending Parsippany High School, Li worked weekends at her family's dry-cleaning shop. She graduated from Parsippany High School in 1995. She was inducted into the hall of fame at Parsippany High School in 2017. Li pursued undergraduate study at Princeton University, where she received a Bachelor of Arts with a major in physics in 1999. Li completed her senior thesis, "Auditory binaural correlogram difference: a new computational model for Huggins dichotic pitch", under the supervision of Bradley Dickinson, professor of electrical engineering. During her years at Princeton, Li returned home most weekends to help run her family's dry cleaning business and worked as a dishwasher to supplement the family income. Li pursued graduate study at the California Institute of Technology, where she received a Master of Science in electrical engineering in 2001 and a Doctor of Philosophy in electrical engineering in 2005. Li completed her dissertation, "Visual Recognition: Computational Models and Human Psychophysics", under the primary supervision of Pietro Perona and secondary supervision of Christof Koch. Her graduate studies were supported by the National Science Foundation Graduate Research Fellowship and The Paul & Daisy Soros Fellowships for New Americans. == Career and research == From 2005 to 2006, Li was an assistant professor in the Electrical and Computer Engineering Department at the University of Illinois Urbana-Champaign, and from 2007 to 2009, she was an assistant professor in the Computer Science Department at Princeton University. She joined Stanford in 2009 as an assistant professor, and was promoted to associate professor with tenure in 2012, and then full professor in 2018. At Stanford, Li served as the director of Stanford Artificial Intelligence Lab (SAIL) from 2013 to 2018. Her research has focused on computer vision, deep learning, and cognitive neuroscience, with over 300 peer-reviewed publications. She became the founding co-director of Stanford's University-level initiative - the Human-Centered AI Institute, along with co-director Dr. John Etchemendy, former provost of Stanford University. The institute aligns with Li's aims to advance AI research, education, policy, and practice to improve the human condition. While at Princeton in 2007, Li led the development of ImageNet, a massive visual database designed to advance object recognition in AI. The project involved labeling over 14 million images using Amazon Mechanical Turk and inspired the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), which catalyzed progress in deep learning and led to dramatic improvements in image classification performance. The database addressed a key bottleneck in computer vision: the lack of large, annotated datasets for training machine learning models. Today, ImageNet is credited as a cornerstone innovation that underpins advancements in autonomous vehicles, facial recognition, and medical imaging. On her sabbatical from Stanford University from January 2017 to fall of 2018, Li joined Google Cloud as its Chief Scientist of AI/ML and Vice President. At Google, her team focused on democratizing AI technology and lowering the barrier for entrance to businesses and developers, including the developments of products like AutoML. In September 2017, Google secured a contract from the Department of Defense called Project Maven, which aimed to use AI techniques to interpret images captured by drone cameras. Google told employees who protested the company's work on Project Maven that their role was "specifically scoped to be for non-offensive purposes". In June 2018, Google told employees it would not seek renewal of the contract. In internal emails which were later leaked to reporters, Li expressed enthusiasm for the Google Cloud role in Project Maven, but warned against mentioning its AI component, saying that military AI is linked in the public mind with the danger of autonomous weapons. Asked about those leaked emails, Li told The New York Times, "I believe in human-centered AI to benefit people in positive and benevolent ways. It is deeply against my principles to work on any project that I think is to weaponize AI." In the fall of 2018, Li left Google and returned to Stanford University to continue her professorship. In 2023, Li co-led the launch of the RAISE-Health (Responsible AI for Safe and Equitable Health) initiative at Stanford University in collaboration with Stanford medicine. The initiative aims to develop frameworks for the responsible use of artificial intelligence in healthcare, including clinical care, biomedical research, and patient safety. According to her Stanford profile, she has been on partial academic leave from January 2024 through the end of 2025 to focus on entrepreneurial ventures. In 2024, Li said there was a disparity between private-sector investment in AI and support for academic and government research, and called for greater public funding for scientific uses of the technology and for studying its risks. Li is also known for her non-profit work as the co-founder and chairperson of nonprofit organization AI4ALL, whose mission is to educate the next generation of AI technologists, thinkers and leaders by promoting diversity and inclusion through human-centered AI principles. The program was created in collaboration with Melinda French Gates and Jensen Huang. Prior to establishing AI4ALL in 2017, Li and her former student Olga Russakovsky, currently an assistant professor in Princeton University, co-founded and co-directed the precursor program at Stanford called SAILORS (Stanford AI Lab OutReach Summers). SAILORS was an annual summer camp at Stanford dedicated to 9th grade high school girls in AI education and research, established in 2015 till it changed its name to AI4ALL @Stanford in 2017. In 2018, AI4ALL has successfully launched five more summer programs in addition to Stanford, including Princeton University, Carnegie Mellon University, Boston University, University of California Berkeley, and Canada's Simon Fraser University. We are at a turning point. AI's influence continues to grow, but representation and inclusion of a diversity of researchers in the field does not. It's critical that we seize this moment to create structures that will support long-term, positive changes. This won't happen via a single mechanism or quick fix. It starts with early education and extends to the existing structures of power within academia, work cultures among current AI researchers, and gatekeeping functions of research publishing, to name a few levers of change. Li has been described as a "researcher bringing humanity to AI". Li was elected as a member of the American Academy of Arts and Sciences in 2021, the National Academy of Engineering in 2020, and the National Academy of Medicine in 2020. In a November 2023 interview with The Guardian, Li said that while she would not refer to herself as the "godmother
Andrew Ng
Andrew Yan-Tak Ng (Chinese: 吳恩達; born April 18, 1976) is a British-American computer scientist and technology entrepreneur focusing on machine learning and artificial intelligence (AI). Ng was a cofounder and head of Google Brain and was the former Chief Scientist at Baidu. Ng is an adjunct professor at Stanford University (formerly associate professor and Director of its Stanford AI Lab or SAIL). Ng has also worked in online education, cofounding Coursera and DeepLearning.AI. He has spearheaded many efforts to "democratize deep learning" teaching over 8 million students through his online courses. Ng is renowned globally in computer science, recognized in Time magazine's 100 Most Influential People in 2012 and Fast Company's Most Creative People in 2014. His influence extends to being named in the Time100 AI Most Influential People in 2023. In 2018, he launched and currently heads the AI Fund, initially a $175-million investment fund for backing artificial intelligence startups. He has founded Landing AI, which provides AI-powered SaaS products. On April 11, 2024, Amazon announced Ng's appointment to its board of directors. == Early life and education == Andrew Yan-Tak Ng was born in London, in 1976 to Ronald Paul Ng, a hematologist and lecturer at UCL Medical School, and Tisa Ho, an arts administrator working at the London Film Festival. His parents were both immigrants from Hong Kong. His family moved back to Hong Kong and he spent his early childhood there. In 1984 he and his family moved to Singapore. Ng attended and graduated from Raffles Institution. In 1997, he earned his undergraduate degree with a triple major in computer science, statistics, and economics from Carnegie Mellon University in Pittsburgh, Pennsylvania. Between 1996 and 1998 he also conducted research on reinforcement learning, model selection, and feature selection at the AT&T Bell Labs. In 1998, Ng earned his master's degree in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology (MIT) in Cambridge, Massachusetts. At MIT, he built the first publicly available, automatically indexed web-search engine for research papers on the web. It was a precursor to CiteSeerX/ResearchIndex, but specialized in machine learning. In 2002, he received his Doctor of Philosophy (Ph.D.) in Computer Science from the University of California, Berkeley, under the supervision of Michael I. Jordan. His thesis is titled "Shaping and policy search in reinforcement learning" and is well-cited to this day. == Career == === Academia and teaching === Ng started working as an assistant professor at Stanford University in 2002 and as an associate professor in 2009. Ng is a professor at Stanford University departments of Computer Science and electrical engineering. He served as the director of the Stanford Artificial Intelligence Laboratory (SAIL), where he taught students and undertook research related to data mining, big data, and machine learning. His machine learning course CS229 at Stanford is the most popular course offered on campus with over 1,000 students enrolling some years. As of 2020, three of the most popular courses on Coursera are Ng's: Machine Learning (#1), AI for Everyone (#5), Neural Networks and Deep Learning (#6). In 2008, his group at Stanford was one of the first in the US to start advocating the use of GPUs in deep learning. The rationale was that an efficient computation infrastructure could speed up statistical model training by orders of magnitude, ameliorating some of the scaling issues associated with big data. At the time it was a controversial and risky decision, but since then and following Ng's lead, GPUs have become a cornerstone in the field. Since 2017, Ng has been advocating the shift to high-performance computing (HPC) for scaling up deep learning and accelerating progress in the field. In 2012, along with Stanford computer scientist Daphne Koller he cofounded and was CEO of Coursera, a website that offers free online courses to everyone. It took off with over 100,000 students registered for Ng's popular CS229A course. Today, several million people have enrolled in Coursera courses, making the site one of the leading massive open online courses (MOOCs) in the world. === Industry === From 2011 to 2012, he worked at Google, where he founded and directed the Google Brain Deep Learning Project with Jeff Dean, Greg Corrado, and Rajat Monga. In 2014, he joined Baidu as chief scientist, and carried out research related to big data and AI. There he set up several research teams for things like facial recognition and Melody, an AI chatbot for healthcare. He also developed for the company the AI platform called DuerOS and other technologies that positioned Baidu ahead of Google in the discourse and development of AI. In March 2017, he announced his resignation from Baidu. He soon afterward launched DeepLearning.AI, an online series of deep learning courses (including the AI for Good Specialization). Then Ng launched LandingAI, which provides AI-powered SaaS products. In January 2018, Ng unveiled the AI Fund, raising $175 million to invest in new startups. In November 2021, LandingAI secured a $57 million round of series A funding led by McRock Capital, to help enterprises adopt AI. In October 2024, Ng's AI Fund made its first investment in India, backing AI healthcare startup Jivi, which uses AI for diagnoses, treatment recommendations, and administrative tasks. The investment highlights the growth of India's AI sector, expected to reach $22 billion by 2027. === Research === Ng researches primarily in machine learning, deep learning, machine perception, computer vision, and natural language processing; and is one of the world's most famous and influential computer scientists. He's frequently won best paper awards at academic conferences and has had a huge impact on the field of AI, computer vision, and robotics. During graduate school, together with David M. Blei and Michael I. Jordan, Ng co-authored the influential paper that introduced latent Dirichlet allocation (LDA) for his thesis on reinforcement learning for drones. His early work includes the Stanford Autonomous Helicopter project, which developed one of the most capable autonomous helicopters in the world. He was the leading scientist and principal investigator on the STAIR (Stanford Artificial Intelligence Robot) project, which resulted in Robot Operating System (ROS), a widely used open source software robotics platform. His vision to build an AI robot and put a robot in every home inspired Scott Hassan to back him and create Willow Garage. He is also one of the founding team members for the Stanford WordNet project, which uses machine learning to expand the Princeton WordNet database created by Christiane Fellbaum. In 2011, Ng founded the Google Brain project at Google, which developed large-scale artificial neural networks using Google's distributed computing infrastructure. Among its notable results was a neural network trained using deep learning algorithms on 16,000 CPU cores, which learned to recognize cats after watching only YouTube videos, and without ever having been told what a "cat" is. The project's technology is also currently used in the Android operating system's speech recognition system. === Views on AI === Ng thinks that the real threat is contemplating the future of work: "Rather than being distracted by evil killer robots, the challenge to labor caused by these machines is a conversation that academia and industry and government should have." He has emphasized the importance of expanding access to AI education, stating that empowering people around the world to use AI tools is essential to building AI applications. In a December 2023 Financial Times interview, Ng highlighted concerns regarding the impact of potential regulations on open-source AI, emphasizing how reporting, licensing, and liability risks could unfairly burden smaller firms and stifle innovation. He argued that regulating basic technologies like open-source models could hinder progress without markedly enhancing safety. Ng advocated for carefully designed regulations to prevent obstacles to the development and distribution of beneficial AI technologies. In a June 2024 interview with the Financial Times, Ng expressed concerns about proposed AI legislation in California that would have required developers to implement safety mechanisms such as a "kill switch" for advanced models. He described the bill as creating "massive liabilities for science-fiction risks" and said it "stokes fear in anyone daring to innovate." Other critics argued the bill would impose burdens on open-source developers and smaller AI companies. The bill was ultimately vetoed by Governor Gavin Newsom in September 2024. == Online education: massive open online course == In 2011, Stanford launched a total of three massive open online course (MOOCs) on machine learning (CS229a), databases, and AI, taught by Ng
Catholic Church and artificial intelligence
The Catholic Church views artificial intelligence as a significant technological development that must be governed by strict ethical principles rooted in human dignity and the common good. In January 2025, the Church issued the doctrinal note Antiqua et nova co-issued by the Dicastery for the Doctrine of the Faith and the Dicastery for Culture and Education. It addresses the "relationship between artificial intelligence and human intelligence" and offers reflections on the "anthropological and ethical challenges raised by AI". In August 2025, Time magazine included Pope Leo XIV in its 2025 list of the World’s Most Influential People in Artificial Intelligence. In May 2026, Pope Leo XIV approved the creation of a new Vatican commission on artificial intelligence. He released his first papal encyclical, titled Magnifica humanitas, on the topic later in the month.
BeyondCorp
BeyondCorp is an implementation of zero-trust computer security concepts creating a zero trust network. It is created by Google. == Background == It was created in response to the 2009 Operation Aurora. An open source implementation inspired by Google's research paper on an access proxy is known as "transcend". Google documented its Zero Trust journey from 2014 to 2018 through a series of articles in the journal ;login:. Google called their ZT network "BeyondCorp". Google implemented a Zero Trust architecture on a large scale, and relied on user and device credentials, regardless of location. Data was encrypted and protected from managed devices. Unmanaged devices, such as BYOD, were not given access to the BeyondCorp resources. == Design and technology == BeyondCorp utilized a zero trust security model, which is a relatively new security model that it assumes that all devices and users are potentially compromised. This is in contrast to traditional security models, which rely on firewalls and other perimeter defenses to protect sensitive data. === Trust === The corporate network grants no inherent trust, and all internal apps are accessed via the BeyondCorp system, regardless of whether the user is in a Google office or working remotely. BeyondCorp is related to Zero Trust architecture as it implements a true Zero Trust network, where all access is granted on identity, device, and authentication, based on robust underlying device and identity data sources. BeyondCorp works by using a number of security policies including authentication, authorization, and access control to ensure that only authorized users can access corporate resources. Authentication verifies the identity of the user, authorization determines whether the user has permission to access the requested resource, and access control policies restrict what the user can do with the resource. ==== Trust Inferrer ==== One of the main components in BeyondCorp's implementation is the Trust Inferrer. The Trust Inferrer is a security component (typically software) that looks at information about a user's device, like a computer or phone, to decide how much it can be trusted to access certain resources like important company documents. The Trust Inferrer checks things like the security of the device, whether it has the right software installed, and if it belongs to an authorized user. Based on all this information, the Trust Inferrer decides what the device can access and what it can't. === Security mechanisms === Unlike traditional VPNs, BeyondCorp's access policies are based on information about a device, its state, and its associated user. BeyondCorp considers both internal networks and external networks to be completely untrusted, and gates access to applications by dynamically asserting and enforcing levels, or “tiers,” of access. === Device Inventory Database === BeyondCorp utilized a Device Inventory Database and Device Identity that uniquely identifies a device through a digital certificate. Any changes to the device are recorded in the Device Inventory Database. The certificate is used to uniquely identify a device; however, additional information is required to grant access privileges to a resource. === Access Control Engine === Another important component of BeyondCorp's implementation is the Access Control Engine. Think of this as the brain of the Zero Trust architecture. The Access Control Engine is like a traffic cop standing at an intersection. Its job is to make sure that only authorized devices and users are allowed to access specific resources (like files or applications) on the network. It checks the access policy (the rules that say who can access what), the device's state (like whether it has the right software updates or security settings), and the resources being requested. Then it makes a decision on whether to grant or deny access based on all of this information. It helps ensure that only the right people and devices are allowed access to the network, which helps keep things secure. The Access Control Engine utilizes the output from the Trust Inferrer and other data that is fed into its system. == Usage == One of the first things Google did to implement a Zero Trust architecture was to capture and analyze network traffic. The purpose of analyzing the traffic was to build a baseline of what typical network traffic looked like. In doing so, BeyondCorp also discovered unusual, unexpected, and unauthorized traffic. This was very useful because it gave the BeyondCorp engineers critical information that assisted them in reengineering the system in a secure manner. Some of the benefits BeyondCorp realized by adopting a Zero Trust architecture include the ability to allow their employees to work securely from any location. It reduces the risk of data breaches since data and applications are protected and users and devices are constantly being verified. The Zero Trust architecture is scalable and can be adapted to the changing needs of the businesses and their users. Especially relevant in today's work-from-home era, BeyondCorp allows employees to access enterprise resources securely from any location, without the need for traditional VPNs.
Catastrophic interference
Catastrophic interference, also known as catastrophic forgetting, is the tendency of an artificial neural network to abruptly and drastically forget previously learned information upon learning new information. Neural networks are an important part of the connectionist approach to cognitive science. The issue of catastrophic interference when modeling human memory with connectionist models was originally brought to the attention of the scientific community by research from McCloskey and Cohen (1989), and Ratcliff (1990). It is a radical manifestation of the 'sensitivity-stability' dilemma or the 'stability-plasticity' dilemma. Specifically, these problems refer to the challenge of making an artificial neural network that is sensitive to, but not disrupted by, new information. Lookup tables and connectionist networks lie on the opposite sides of the stability plasticity spectrum. The former remains completely stable in the presence of new information but lacks the ability to generalize, i.e. infer general principles, from new inputs. On the other hand, connectionist networks like the standard backpropagation network can generalize to unseen inputs, but they are sensitive to new information. Backpropagation models can be analogized to human memory insofar as they have a similar ability to generalize, but these networks often exhibit less stability than human memory. Notably, these backpropagation networks are susceptible to catastrophic interference. This is an issue when modelling human memory, because unlike these networks, humans typically do not show catastrophic forgetting. == Discovery == The term catastrophic interference was originally coined by McCloskey and Cohen (1989) but was also brought to the attention of the scientific community by research from Ratcliff (1990). === The Sequential Learning Problem: McCloskey and Cohen (1989) === McCloskey and Cohen (1989) noted the problem of catastrophic interference during two different experiments with backpropagation neural network modelling. Experiment 1: Learning the ones and twos addition facts In their first experiment they trained a standard backpropagation neural network on a single training set consisting of 17 single-digit ones problems (i.e., 1 + 1 through 9 + 1, and 1 + 2 through 1 + 9) until the network could represent and respond properly to all of them. The error between the actual output and the desired output steadily declined across training sessions, which reflected that the network learned to represent the target outputs better across trials. Next, they trained the network on a single training set consisting of 17 single-digit twos problems (i.e., 2 + 1 through 2 + 9, and 1 + 2 through 9 + 2) until the network could represent, respond properly to all of them. They noted that their procedure was similar to how a child would learn their addition facts. Following each learning trial on the twos facts, the network was tested for its knowledge on both the ones and twos addition facts. Like the ones facts, the twos facts were readily learned by the network. However, McCloskey and Cohen noted the network was no longer able to properly answer the ones addition problems even after one learning trial of the twos addition problems. The output pattern produced in response to the ones facts often resembled an output pattern for an incorrect number more closely than the output pattern for a correct number. This is considered to be a drastic amount of error. Furthermore, the problems 2+1 and 1+2, which were included in both training sets, even showed dramatic disruption during the first learning trials of the twos facts. Experiment 2: Replication of Barnes and Underwood (1959) study In their second connectionist model, McCloskey and Cohen attempted to replicate the study on retroactive interference in humans by Barnes and Underwood (1959). They trained the model on A-B and A-C lists and used a context pattern in the input vector (input pattern), to differentiate between the lists. Specifically the network was trained to respond with the right B response when shown the A stimulus and A-B context pattern and to respond with the correct C response when shown the A stimulus and the A-C context pattern. When the model was trained concurrently on the A-B and A-C items then the network readily learned all of the associations correctly. In sequential training the A-B list was trained first, followed by the A-C list. After each presentation of the A-C list, performance was measured for both the A-B and A-C lists. They found that the amount of training on the A-C list in Barnes and Underwood study that lead to 50% correct responses, lead to nearly 0% correct responses by the backpropagation network. Furthermore, they found that the network tended to show responses that looked like the C response pattern when the network was prompted to give the B response pattern. This indicated that the A-C list apparently had overwritten the A-B list. This could be likened to learning the word dog, followed by learning the word stool and then finding that you think of the word stool when presented with the word dog. McCloskey and Cohen tried to reduce interference through a number of manipulations including changing the number of hidden units, changing the value of the learning rate parameter, overtraining on the A-B list, freezing certain connection weights, changing target values 0 and 1 instead 0.1 and 0.9. However, none of these manipulations satisfactorily reduced the catastrophic interference exhibited by the networks. Overall, McCloskey and Cohen (1989) concluded that: at least some interference will occur whenever new learning alters the weights involved in representing old learning the greater the amount of new learning, the greater the disruption in old knowledge interference was catastrophic in the backpropagation networks when learning was sequential but not concurrent === Constraints Imposed by Learning and Forgetting Functions: Ratcliff (1990) === Ratcliff (1990) used multiple sets of backpropagation models applied to standard recognition memory procedures, in which the items were sequentially learned. After inspecting the recognition performance models he found two major problems: Well-learned information was catastrophically forgotten as new information was learned in both small and large backpropagation networks. Even one learning trial with new information resulted in a significant loss of the old information, paralleling the findings of McCloskey and Cohen (1989). Ratcliff also found that the resulting outputs were often a blend of the previous input and the new input. In larger networks, items learned in groups (e.g. AB then CD) were more resistant to forgetting than were items learned singly (e.g. A then B then C...). However, the forgetting for items learned in groups was still large. Adding new hidden units to the network did not reduce interference. Discrimination between the studied items and previously unseen items decreased as the network learned more. This finding contradicts studies on human memory, which indicated that discrimination increases with learning. Ratcliff attempted to alleviate this problem by adding 'response nodes' that would selectively respond to old and new inputs. However, this method did not work as these response nodes would become active for all inputs. A model which used a context pattern also failed to increase discrimination between new and old items. == Proposed solutions == The main cause of catastrophic interference seems to be overlap in the representations at the hidden layer of distributed neural networks. In a distributed representation, each input tends to create changes in the weights of many of the nodes. Catastrophic forgetting occurs because when many of the weights where "knowledge is stored" are changed, it is unlikely for prior knowledge to be kept intact. During sequential learning, the inputs become mixed, with the new inputs being superimposed on top of the old ones. Another way to conceptualize this is by visualizing learning as a movement through a weight space. This weight space can be likened to a spatial representation of all of the possible combinations of weights that the network could possess. When a network first learns to represent a set of patterns, it finds a point in the weight space that allows it to recognize all of those patterns. However, when the network then learns a new set of patterns, it will move to a place in the weight space for which the only concern is the recognition of the new patterns. To recognize both sets of patterns, the network must find a place in the weight space suitable for recognizing both the new and the old patterns. Below are a number of techniques which have empirical support in successfully reducing catastrophic interference in backpropagation neural networks: === Orthogonality === Many of the early techniques in reducing representational overlap involved making either the input vecto