annaweb

Large language models, such as those that power popular artificial intelligence chatbots like ChatGPT, are incredibly complex. Even though these models are being used as tools in many areas, such as customer support, code generation, and language translation, scientists still don’t fully grasp how they work.

In an effort to better understand what is going on under the hood, researchers at MIT and elsewhere studied the mechanisms at work when these enormous machine-learning models retrieve stored knowledge.

They found a surprising result: Large language models (LLMs) often use a very simple linear function to recover and decode stored facts. Moreover, the model uses the same decoding function for similar types of facts. Linear functions, equations with only two variables and no exponents, capture the straightforward, straight-line relationship between two variables.

The researchers showed that, by identifying linear functions for different facts, they can probe the model to see what it knows about new subjects, and where within the model that knowledge is stored.

Using a technique they developed to estimate these simple functions, the researchers found that even when a model answers a prompt incorrectly, it has often stored the correct information. In the future, scientists could use such an approach to find and correct falsehoods inside the model, which could reduce a model’s tendency to sometimes give incorrect or nonsensical answers.

“Even though these models are really complicated, nonlinear functions that are trained on lots of data and are very hard to understand, there are sometimes really simple mechanisms working inside them. This is one instance of that,” says Evan Hernandez, an electrical engineering and computer science (EECS) graduate student and co-lead author of a paper detailing these findings.

Hernandez wrote the paper with co-lead author Arnab Sharma, a computer science graduate student at Northeastern University; his advisor, Jacob Andreas, an associate professor in EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); senior author David Bau, an assistant professor of computer science at Northeastern; and others at MIT, Harvard University, and the Israeli Institute of Technology. The research will be presented at the International Conference on Learning Representations.

Finding facts

Most large language models, also called transformer models, are neural networks. Loosely based on the human brain, neural networks contain billions of interconnected nodes, or neurons, that are grouped into many layers, and which encode and process data.

Much of the knowledge stored in a transformer can be represented as relations that connect subjects and objects. For instance, “Miles Davis plays the trumpet” is a relation that connects the subject, Miles Davis, to the object, trumpet.

As a transformer gains more knowledge, it stores additional facts about a certain subject across multiple layers. If a user asks about that subject, the model must decode the most relevant fact to respond to the query.

If someone prompts a transformer by saying “Miles Davis plays the. . .” the model should respond with “trumpet” and not “Illinois” (the state where Miles Davis was born).

“Somewhere in the network’s computation, there has to be a mechanism that goes and looks for the fact that Miles Davis plays the trumpet, and then pulls that information out and helps generate the next word. We wanted to understand what that mechanism was,” Hernandez says.

The researchers set up a series of experiments to probe LLMs, and found that, even though they are extremely complex, the models decode relational information using a simple linear function. Each function is specific to the type of fact being retrieved.

For example, the transformer would use one decoding function any time it wants to output the instrument a person plays and a different function each time it wants to output the state where a person was born.

The researchers developed a method to estimate these simple functions, and then computed functions for 47 different relations, such as “capital city of a country” and “lead singer of a band.”

While there could be an infinite number of possible relations, the researchers chose to study this specific subset because they are representative of the kinds of facts that can be written in this way.

They tested each function by changing the subject to see if it could recover the correct object information. For instance, the function for “capital city of a country” should retrieve Oslo if the subject is Norway and London if the subject is England.

Functions retrieved the correct information more than 60 percent of the time, showing that some information in a transformer is encoded and retrieved in this way.

“But not everything is linearly encoded. For some facts, even though the model knows them and will predict text that is consistent with these facts, we can’t find linear functions for them. This suggests that the model is doing something more intricate to store that information,” he says.

Visualizing a model’s knowledge

They also used the functions to determine what a model believes is true about different subjects.

In one experiment, they started with the prompt “Bill Bradley was a” and used the decoding functions for “plays sports” and “attended university” to see if the model knows that Sen. Bradley was a basketball player who attended Princeton.

“We can show that, even though the model may choose to focus on different information when it produces text, it does encode all that information,” Hernandez says.

They used this probing technique to produce what they call an “attribute lens,” a grid that visualizes where specific information about a particular relation is stored within the transformer’s many layers.

Attribute lenses can be generated automatically, providing a streamlined method to help researchers understand more about a model. This visualization tool could enable scientists and engineers to correct stored knowledge and help prevent an AI chatbot from giving false information.

In the future, Hernandez and his collaborators want to better understand what happens in cases where facts are not stored linearly. They would also like to run experiments with larger models, as well as study the precision of linear decoding functions.

“This is an exciting work that reveals a missing piece in our understanding of how large language models recall factual knowledge during inference. Previous work showed that LLMs build information-rich representations of given subjects, from which specific attributes are being extracted during inference. This work shows that the complex nonlinear computation of LLMs for attribute extraction can be well-approximated with a simple linear function,” says Mor Geva Pipek, an assistant professor in the School of Computer Science at Tel Aviv University, who was not involved with this work.

This research was supported, in part, by Open Philanthropy, the Israeli Science Foundation, and an Azrieli Foundation Early Career Faculty Fellowship."

#metaglossia_mundus

The advent of the digital era has transformed communication, streamlining it through the channels of social media. Our commonplace lexicon now incorporates numerous neologisms—freshly coined words—conveying notions that would have been unfamiliar in the past. Within our non-exhaustive compilation, numerous words result from the fusion of two existing terms to characterize a new concept.

Some amalgamate Latin etymologies, while others have transitioned from nouns to verbs in popularity. To formulate this roster of 30 words absent from the lexicon 30 years ago, we have scrutinized several online English dictionaries for definitions and meanings, with exclusions made for terms like “social media.”

Expressions with pre-existing meanings that underwent transformation, such as “muggle,” were also omitted. Originally a colloquial reference to marijuana, “muggle” acquired a new significance through its adoption by J. K. Rowling in the Harry Potter series, now denoting individuals lacking specific skills. Following are the 30 words (along with their origins) that weren’t in use before the 1990s but are now commonly used:

1. Hater
   First used circa: 1991
   Meaning: A negative or critical person
   Notably popularized by the 1997 song “Playa Hater” by Notorious B.I.G, with roots traced back to Cypress Hill’s 1991 song “Psycobetabuckdown” in the hip-hop dictionary The Right Rhymes.
2. Smartphone
   First used circa: 1992
   Meaning: A mobile phone that performs many computer functions
   Though commonly associated with the 2007 release of the first iPhone, the term “smartphone” originated in 1992 with the creation of the Simon Personal Communicator by IBM, combining cell phone and personal digital assistant (PDA) functions.
3. Boop
   First used circa: 1992
   Meaning: Affectionately poking someone on the nose and saying “Boop!”
   Popularized by a 1992 episode of “The Simpsons,” where the character Bart playfully pokes his sister’s nose while making the sound “boop.”
4. Upcycle
   First used circa: 1994
   Meaning: Reusing materials to create a product of higher quality than the original
   Coined in a 1994 interview with Thornton Kay, discussing waste disposal systems, and gained traction with the 2002 publication of the book “Cradle to Cradle: The Way We Make Things.”
5. Wi-Fi
   First used circa: 1997
   Meaning: Wireless network protocols allowing devices to connect to the internet
   Invented in 1997, Wi-Fi led to the establishment of standards for communication in wireless local area networks (WLANs).
6. Emoji
   First used circa: 1997
   Meaning: A small digital image used to express an idea or emotion
   Originating from the Japanese words for “picture” and “character,” the first known emoji appeared on a mobile phone in 1997 by J-Phone (now SoftBank).
7. Cyberbullying
   First used circa: 1998
   Meaning: The use of electronic communication to bully a person
   Documented in 1998, the term gained prominence as online harassment, including cases like the 2007 suicide of 13-year-old Megan Meier, was linked to tragic outcomes.
8. Blog
   First used circa: 1999
   Meaning: A regularly updated web page written in an informal style
   Short for “weblog,” attributed to Jorn Barger, with the term “blog” commonly shortened by programmer Peter Merholz.
9. Vlog
   First used circa: 2000
   Meaning: A blog in the form of a video
   Credited to Adam Kontras, the first vlog was posted on January 2, 2000, featuring Kontras bringing a cat into a hotel with a “No Pets” policy.
10. FOMO
    First used circa: 2000
    Meaning: Fear of missing out
    Theories suggest creation in 2000 as a marketing term by Dan Herman or in 2004 by Harvard business student Patrick McGinnis, who popularized the term describing overly ambitious managers.
11. Bromance
    First used circa: 2001
    Meaning: A close friendship between men
    Coined around 2001, “bromance” combines “bro” (brother) and “romance,” encapsulating the idea of a deep friendship, often referred to as a “brother from another mother.”
12. Showmance
    First used circa: 2001
    Meaning: A romantic relationship that develops between cast members
    Originating in 2001 within the context of reality TV, “showmance” denotes a romance between cast members or production crew that typically lasts only during the show’s duration.
13. Google
    First used circa: 2002
    Meaning: Searching for information using the search engine Google
    While the Google search engine was established in 1998, the term “google” as a transitive verb emerged in 2002, notably used in the TV show “Buffy The Vampire Slayer.”
14. Selfie
    First used circa: 2002
    Meaning: A photograph that one has taken of oneself
    Although the act of taking self-portraits dates back to 1839, the term “selfie” was coined in 2002 when an inebriated Australian described his self-portrait as a “selfie,” later declared the word of the year by Oxford Dictionaries in 2013.
15. Sexting
    First used circa: 2004
    Meaning: Sending sexually explicit photographs or messages via text
    Coined in 2004 during a text messaging scandal involving David Beckham, “sexting” gained mainstream use in 2008, revealing the prevalence of sending explicit content via text messages.
16. Facebook
    First used circa: 2004
    Meaning: A social network
    Launched in 2004 as “Thefacebook,” the term “facebook” transformed into a verb, signifying the act of spending time on the social networking website.
17. Podcast
    First used circa: 2004
    Meaning: A digital audio program available for download
    Originating from audio blogging in the 1980s, “podcast” was first coined by Ben Hammersley in The Guardian in 2004 to describe downloadable analytical radio programming.
18. Paywall
    First used circa: 2005
    Meaning: A system in which access to all or part of a website is restricted to paid subscribers
    Although The Wall Street Journal implemented a paywall in 1997, the term “paywall” was officially recorded in 2005, representing the practice of restricting website access to paying subscribers.
19. Locavore
    First used circa: 2005
    Meaning: People who mainly eat food produced within a 250-mile radius of where they live
    Coined by Jessica Prentice in 2005, “locavore” describes individuals who prioritize consuming locally sourced food, derived from the Latin words “locus” (place) and “vorare” (swallow).
20. Turnt
    First used circa: 2005
    Meaning: Extremely excited, wild, or drunk
    Initially documented in 2005 by Urban Dictionary, “turnt” gained visibility on Twitter in 2008, with its precise origin, particularly speculated to have connections to hip-hop music.
21. YouTuber
    First used circa: 2006
    Meaning: A frequent user of the video-sharing website YouTube
    Coined in 2006, “YouTuber” emerged as a term for individuals actively engaged with the YouTube platform, officially recognized by the Oxford English Dictionary in 2016.
22. Clickbait
    First used circa: 2006
    Meaning: Content the main purpose of which is to encourage people to click on a link
    Originating around 2006 and officially added to the Oxford English Dictionary in 2016, “clickbait” refers to content designed to entice clicks, a term that gained prominence with the rise of online media.
23. Crowdsourcing
    First used circa: 2006
    Meaning: Obtaining information by enlisting the services of a large number of people
    Coined in a 2006 Wired article titled “Rise of Crowdsourcing” by Jeff Howe, describing the utilization of a large pool of individuals for creating content and solving problems.
24. Hashtag
    First used circa: 2007
    Meaning: A word or phrase preceded by a hash sign
    Contrary to popular belief, the hashtag was introduced to Twitter in 2007 by user Chris Messina, suggesting its use for grouping discussions with the symbol #.
25. Birther
    First used circa: 2008
    Meaning: A person who claims President Barack Obama was born outside the US
    Coined during the 2008 presidential campaign, “birther” describes individuals who falsely asserted that Barack Obama was not a natural-born U.S. citizen.
26. Bitcoin
    First used circa: 2008
    Meaning: A type of digital currency
    Although the cryptocurrency Bitcoin was created in 2009, the term was first documented in 2008 when the domain bitcoin.org was registered.
27. Photobomb
    First used circa: 2008
    Meaning: Sneaking into the background of people’s photographs
    While added to the Merriam-Webster Dictionary in 2017, “photobomb” traces back to a 2008 entry on Urban Dictionary, defining the act of intentionally intruding into the background of a photo.
28. Adulting
    First used circa: 2008
    Meaning: To do the things that adults regularly have to do
    Gaining popularity with millennials, “adulting” refers to the process of performing typical adult responsibilities and was first observed in a 2008 tweet.
29. Omnishamble
    First used circa: 2009
    Meaning: Something that is totally disorganized and a complete failure
    Coined in 2009 in the British political satire “The Thick of It,” “omnishamble” combines “omni” meaning “all” with “shambles.”
30. Bingeable
    First used circa: 2013
    Meaning: A show having multiple episodes or parts that can be watched in rapid succession
    Coined around 2013 and officially recognized by the Merriam-Webster Dictionary in 2018, “bingeable” describes a show with episodes conducive to rapid and continuous watching, particularly with the advent of streaming services like Netflix."

#metaglossia_mundus