IEBus (Inter Equipment Bus) is a communication bus specification "between equipments within a vehicle or a chassis" of Renesas Electronics. It defines OSI model layer 1 and layer 2 specification. IEBus is mainly used for car audio and car navigations, which established de facto standard in Japan, though SAE J1850 is major in United States. IEBus is also used in some vending machines, which major customer is Fuji Electric. Each button on the vending machine has an IEBus ID, i.e. has a controller. Detailed specification is disclosed to licensees only, but protocol analyzers are provided from some test equipment vendors. Its modulation method is PWM (Pulse-Width Modulation) with 6.00 MHz base clock originally, but most of automotive customers use 6.291 MHz, and physical layer is a pair of differential signalling harness. Its physical layer adopts half-duplex, asynchronous, and multi-master communication with carrier-sense multiple access with collision detection (CSMA/CD) for medium access control. It allows for up to fifty units on one bus over a maximum length of 150 meters. Two differential signalling lines are used with Bus+ / Bus− naming, sometimes labeled as Data(+) / Data(−). It is sometimes described as "IE-BUS", "IE-Bus," or "IE Bus," but these are incorrect. In formal, it is "IEBus." IEBus® and Inter Equipment Bus® are registered trademark symbols of Renesas Electronics Corporation, formerly NEC Electronics Corporation, (JPO: Reg. No.2552418 and 2552419, respectively). == History == In the middle of '80s, semiconductor unit of NEC Corporation, currently Renesas Electronics, started the study for increasing demands for automotive audio systems. IEBus is introduced as a solution for the distributed control system. In the late 1980s, several similar specifications, including the Domestic Digital Bus (D2B), the Japanese Home Bus (HBS), and the European Home System (EHS) are proposed by different companies or organizations. These were once discussed as IEC 61030, but it was withdrawn in 2006. IEBus is also a similar specification (refer to "Transfer signal format" section), but not listed in these criteria. As the result, IEBus becomes a de facto standard of car audio in Japan. Regarding the Domestic Digital Bus (D2B), it is re-defined as D2B Optical by Mercedes-Benz independently. As for Japanese Home Bus System (HBS), it is defined in 1988 as Home Bus System Standard Specification, ET-2101 by JEITA and REEA (Radio Engineering & Electronics Assiation) in Japan. It is being used by several Japanese air conditioner manufacturers (for example, M-Net from Mitsubishi and the P1/P2 or F1/F2 bus from Daikin). Fujitsu provided HBPC (Home Bus Protocol Controller) chip as MB86046B. But it is unclear whether Fujitsu (currently, Cypress) still manufactures this HBPC LSI as of 2018. Mitsumi Electric provides the MM1007 and MM1192 driver ICs for HBS. The HBS specification is also discussed in the Echonet Consortium. In 2014, a utility model patent for protocol converter from HBS to RS-485 is granted in China as "CN204006496U." Regarding the replacement of IEBus, a paper by Hyundai Autonet, currently Hyundai Mobis, describes as follows. "In communication methods for digital input capable amplifiers, Inter Equipment Bus (IEBus) was used in early times, but for now, Controller Area Network (CAN) is mainly used." == Protocol overview == A master talks to a slave. Each unit has a master and a slave address register. Only one device can talk on the bus at any given time. There is a pecking order for the types of communications which will take precedence over another. Each communication from master to slave must be replied to by the slave going back to the master with acknowledge bits each of those show ACK or NAK. If the master does not receive the ACK within a predefined time allowance for a mode, it drops the communication and returns to its standby (listen) mode. Detailed specification of OSI model layer 2 is disclosed to licensees only, but protocol analyzers are provided from some test equipment vendors. In 2012, one of Chinese manufacturer's patent is granted as "CN202841169U". An open-source software emulator called "IEBus Studio" exists on a repository of SourceForge, but the last update was on 2008-02-24. Another open-source analyzer software called "IEBusAnalyzer" is available on GitHub repository. Some hobbyist made some tools also. === Physical layer (OSI model layer 1) specification overview === From μPD6708 data sheet. and μPD78098B Subseries user's manual, hardware. Communication system Half-duplex asynchronous communication Multi-master system All the units connected to the IEBus can transfer data to the other units. Broadcast communication function (communication between one unit and multiple units) Normally, communication is individually carried out from one unit to another. By using the broadcast communication function, however, communication can be executed from one unit to plural units as follows: Group broadcast communication: Broadcast communication to group units Simultaneous broadcast communication: Broadcast communication to all units Effective transmission rate The effective transmission rate can be selected from the following three communication modes: Mixture of the plural of modes in the same bus line is not allowed. Correct communication between different base clock is not possible. Access control CSMA/CD (Carrier Sense Multiple Access with Collision Detection) The priority of occupying IEBus is as follows: «1» Broadcast communication takes precedence over individual communication. «2» The lower the master address, the higher the priority. Communication scale Number of units: 50 MAX. Cable length: 150 m MAX. (when a twisted pair cable is used) Load capacity: MAX. 8000 pF; between Bus+ and Bus−, (6.000000 MHz base clock) MAX. 7100 pF; between Bus+ and Bus−, (6.291456 MHz base clock) Terminating resistor: 120 Ω Logic level Logic 1: Low level. Voltage difference between Bus+ and Bus− is under 20mV Logic 0: High Level. Voltage difference between Bus+ and Bus− is over 120mV In-phase input voltage high: Bus+ ≤ (VDD-1.0) V, Bus− ≥ 1.0 V === Transfer signal format === From μPD6708 data sheet. and μPD78098B Subseries user's manual, hardware. This frame format is much similar to that of Domestic Digital Bus (D2B). All fields are MSB first. ==== Functions of Control bits ==== === Bit format === Each IEBus bit consists of four periods. Preparation period: The first or subsequent low-level (logic "1") period Synchronization period: Next high-level (logic "0") period Data period: Period indicating value of bit; ether low-level (logic "1") or high-level (logic "0") Stop period: The last low-level (logic "1") period Synchronization is done by each bit. Time lengths of the synchronization period and data period are almost the same. The time of the entire bits' and each bit's specification, related to the time of each period allocated to it, differ depending both on the type of the transmit bit and on whether the unit is the master or a slave unit. == Automotive manufacturers using IEBus == Each manufacturer has its own name, but it is not an alias of IEBus. Those are specifications of wire harness which comprise control cables based on IEBus, OSI model layer 3 and above communication protocol, audio cables, interconnection couplers, and so on. === Pioneer === Pioneer Corporation employed IEBus for its original branded car audio in early '90s. In its earlier stage, it was used just for control bus between the head unit in dashboard and the CD changer usually placed in trunk room. Nowadays, the specification includes connection between head units, navigation systems, rear speaker systems, and so on. IP-Bus: Wire harness specification. === Toyota === Pioneer Corporation pushed Toyota Motor Corporation to adopt IEBus as the genuine parts. In 1994, Toyota decided to employ IEBus for its genuine specification, but it is slightly different from that of Pioneer. It is named as AVC-LAN. AVC-LAN: Wire harness specification, based on mode 2. === Honda/Acura === Pioneer Corporation also pushed Honda Motor. Honda also decided to adopt IEBus as its genuine parts specification just after Toyota do so. GA-NET II: Wire harness specification. Honda Music Link: Honda genuine gadget to connect Apple Inc. products. A hobbyist made touch screen controller on Acura TSX for a Car PC installed in the trunk. === Sirius XM Satellite Radio === Sirius XM Satellite Radio is a satellite broadcasting radio operator in US. Its digital media receiver equipment utilizes IEBus. == Evaluation boards == === SAKURA board === GR-SAKUKRA board and GR-SAKURA-FULL board are Renesas official promotion boards of RX63N chip, which enables IEBus mode 0 and 1, but not mode 2, i.e. not available for Toyota AVC-LAN. They are an Arduino pin compatible low-price ones, suitable for hobbyists. Their color of printed circuit board is SAKURA in Japanese, which means cherry blossom. To e
PagedAttention
PagedAttention is an attention algorithm for efficient serving of large language models (LLMs). It was introduced in 2023 by Woosuk Kwon and colleagues in the paper Efficient Memory Management for Large Language Model Serving with PagedAttention, alongside the vLLM serving engine. The method stores the key–value cache used during autoregressive decoding in fixed-size blocks that can be mapped to non-contiguous physical memory, borrowing ideas from virtual memory, paging, and operating system design. == Background == In transformer inference, the key–value cache grows with sequence length and the number of concurrent requests. Kwon et al. argued that earlier serving systems typically reserved contiguous cache regions in advance, which caused reserved space, internal fragmentation, and external fragmentation. In their experiments, the paper reported that the effective memory utilization of previous systems could fall as low as 20.4%. == Description == PagedAttention partitions the cache of each sequence into fixed-size KV blocks. A request's cache is represented as a sequence of logical blocks, while a block table maps those logical blocks to physical GPU-memory blocks. As a result, neighboring logical blocks do not need to be contiguous in physical memory, and new blocks can be allocated on demand as generation proceeds. The design also makes it easier to share cache state across related decoding paths. In vLLM, physical blocks can be reference-counted and shared among requests or branches, with block-granularity copy-on-write used when a shared block must be modified. The original paper applied this design to parallel sampling, beam search, and prompts with shared prefixes. == Mathematical formulation == For a query token i {\displaystyle i} in causal self-attention, the standard attention output can be written as a i j = exp ( q i ⊤ k j / d ) ∑ t = 1 i exp ( q i ⊤ k t / d ) , o i = ∑ j = 1 i a i j v j {\displaystyle a_{ij}={\frac {\exp(\mathbf {q} _{i}^{\top }\mathbf {k} _{j}/{\sqrt {d}})}{\sum _{t=1}^{i}\exp(\mathbf {q} _{i}^{\top }\mathbf {k} _{t}/{\sqrt {d}})}},\;\mathbf {o} _{i}=\sum _{j=1}^{i}a_{ij}\mathbf {v} _{j}} where q i {\displaystyle \mathbf {q} _{i}} , k j {\displaystyle \mathbf {k} _{j}} , and v j {\displaystyle \mathbf {v} _{j}} are the query, key, and value vectors, and d {\displaystyle d} is the attention dimension. If the cache is partitioned into blocks of size B {\displaystyle B} , the key and value blocks may be written as K j = ( k ( j − 1 ) B + 1 , … , k j B ) , V j = ( v ( j − 1 ) B + 1 , … , v j B ) {\displaystyle \mathbf {K} _{j}=(\mathbf {k} _{(j-1)B+1},\ldots ,\mathbf {k} _{jB}),\;\mathbf {V} _{j}=(\mathbf {v} _{(j-1)B+1},\ldots ,\mathbf {v} _{jB})} PagedAttention then performs the computation blockwise: A i j = exp ( q i ⊤ K j / d ) ∑ t = 1 ⌈ i / B ⌉ exp ( q i ⊤ K t / d ) , o i = ∑ j = 1 ⌈ i / B ⌉ V j A i j ⊤ {\displaystyle \mathbf {A} _{ij}={\frac {\exp(\mathbf {q} _{i}^{\top }\mathbf {K} _{j}/{\sqrt {d}})}{\sum _{t=1}^{\lceil i/B\rceil }\exp(\mathbf {q} _{i}^{\top }\mathbf {K} _{t}/{\sqrt {d}})}},\;\mathbf {o} _{i}=\sum _{j=1}^{\lceil i/B\rceil }\mathbf {V} _{j}\mathbf {A} _{ij}^{\top }} where A i j {\displaystyle \mathbf {A} _{ij}} is the vector of attention scores for the j {\displaystyle j} -th KV block. In the formulation given by Kwon et al., this preserves the causal attention calculation while allowing the key and value blocks to reside in non-contiguous physical memory. == Performance and use == The vLLM paper reported that, on its evaluated workloads, the use of PagedAttention and the associated memory-management design improved serving throughput by 2–4× over the compared baselines, including FasterTransformer and Orca, while preserving model outputs. In experiments on OPT-13B with the Alpaca trace, the paper also reported memory savings of 6.1–9.8% for parallel sampling and 37.6–55.2% for beam search through KV-block sharing. A 2024 survey of LLM serving systems described PagedAttention as having become an industry norm in LLM serving frameworks, citing support in TGI, vLLM, and TensorRT-LLM. == Limitations and alternatives == Subsequent work has described trade-offs in the approach. The 2025 vAttention paper argued that PagedAttention requires attention kernels to be rewritten to support paging and increases software complexity, portability issues, redundancy, and execution overhead, proposing instead a memory manager that keeps the cache contiguous in virtual memory while relying on demand paging for physical allocation. === vAttention === Unlike PagedAttention, vAttention does not introduce a different attention rule; it retains the standard attention computation Attention ( q i , K , V ) = softmax ( q i K ⊤ s c a l e ) V . {\displaystyle \operatorname {Attention} (q_{i},K,V)=\operatorname {softmax} \left({\frac {q_{i}K^{\top }}{\mathrm {scale} }}\right)V.} In the notation of Prabhu et al., the key and value tensors for a request seen so far are K , V ∈ R L ′ × ( H × D ) {\displaystyle K,V\in \mathbb {R} ^{L'\times (H\times D)}} , where L ′ {\displaystyle L'} is the context length seen so far, H {\displaystyle H} is the number of KV heads on a worker, and D {\displaystyle D} is the dimension of each KV head. In systems prior to PagedAttention, the K cache (or V cache) at each layer of a worker is typically allocated as a 4D tensor of shape [ B , L , H , D ] , {\displaystyle [B,L,H,D],} where B {\displaystyle B} is batch size and L {\displaystyle L} is the maximum context length supported by the model. vAttention preserves this contiguous virtual-memory view while deferring physical-memory allocation to runtime. A serving framework maintains separate K and V tensors for each layer, so vAttention reserves 2 N {\displaystyle 2N} virtual-memory buffers on a worker, where N {\displaystyle N} is the number of layers managed by that worker. The maximum size of one virtual-memory buffer is B S = B × S , {\displaystyle BS=B\times S,} where S {\displaystyle S} is the maximum size of a single request's per-layer K cache (or V cache) on a worker. The paper defines S = L × H × D × P , {\displaystyle S=L\times H\times D\times P,} where P {\displaystyle P} is the number of bytes needed to store one element. In this formulation, vAttention keeps the KV cache contiguous in virtual memory and relies on demand paging for physical allocation, rather than modifying the attention kernel to operate over non-contiguous KV-cache blocks.
List of Tesla Autopilot crashes
Tesla Autopilot, a Level 2 advanced driver assistance system (ADAS), was released in October 2015 and the first fatal crashes involving the system occurred less than one year later. The fatal crashes attracted attention from news publications and United States government agencies, including the National Transportation Safety Board (NTSB) and National Highway Traffic Safety Administration (NHTSA), which has argued the Tesla Autopilot death rate is higher than the reported estimates. In addition to fatal crashes, there have been many nonfatal ones. Causes behind the incidents include the ADAS failing to recognize other vehicles, insufficient Autopilot driver engagement, and violating the operational design domain. As of October 2025, there have been hundreds of nonfatal incidents involving versions of Autopilot and sixty-five reported fatalities, fifty-four of which NHTSA investigations or expert testimony later verified and two that NHTSA's Office of Defect Investigations determined as happening during the engagement of Full Self-Driving (FSD) after 2022. Collectively, these cases culminated in a general recall in December 2023 of all vehicles equipped with Autopilot, which Tesla claims it resolved by an over-the-air software update. Immediately after closing its investigation in April 2024, NHTSA opened a recall query to determine the effectiveness of the recall. == Notable fatal crashes == === Handan, Hebei, China (January 20, 2016) === On January 20, 2016, Gao Yaning, the driver of a Tesla Model S in Handan, Hebei, China, was killed when his car crashed into a stationary truck. The Tesla was following a car in the far left lane of a multi-lane highway; the car in front moved to the right lane to avoid a truck stopped on the left shoulder, and the Tesla, which the driver's father believes was in Autopilot mode, did not slow before colliding with the stopped truck. According to footage captured by a dashboard camera, the stationary street sweeper on the left side of the expressway partially extended into the far left lane, and the driver did not appear to respond to the unexpected obstacle. Initially, Yaning was held responsible for the collision by local traffic police and, in September 2016, his family filed a lawsuit in July against the Tesla dealer who sold the car. The family's lawyer stated the suit was intended "to let the public know that self-driving technology has some defects. We are hoping Tesla when marketing its products, will be more cautious. Do not just use self-driving as a selling point for young people." Tesla released a statement which said they "have no way of knowing whether or not Autopilot was engaged at the time of the crash" since the car telemetry could not be retrieved remotely due to damage caused by the crash. In 2018, the lawsuit was stalled because telemetry was recorded locally to a SD card and was not able to be given to Tesla, who provided a decoding key to a third party for independent review. Tesla stated that "while the third-party appraisal is not yet complete, we have no reason to believe that Autopilot on this vehicle ever functioned other than as designed." Chinese media later reported that the family sent the information from that card to Tesla, which admitted Autopilot was engaged two minutes before the crash. Tesla since then removed the term "Autopilot" from its Chinese website. === Williston, Florida, US (May 7, 2016) === On May 7, 2016, Tesla driver Joshua Brown was killed in a crash with an 18-wheel tractor-trailer in Williston, Florida. By late June 2016, the NHTSA opened a formal investigation into the fatal autonomous accident, working with the Florida Highway Patrol. According to the NHTSA, preliminary reports indicate the crash occurred when the tractor-trailer made a left turn in front of the 2015 Tesla Model S at an intersection on a non-controlled access highway, and the car failed to apply the brakes. The car continued to travel after passing under the truck's trailer. The Tesla was eastbound in the rightmost lane of US 27, and the westbound tractor-trailer was turning left at the intersection with NE 140th Court, approximately 1 mi (1.6 km) west of Williston; the posted speed limit is 65 mph (105 km/h). The diagnostic log of the Tesla indicated it was traveling at a speed of 74 mi/h (119 km/h) when it collided with and traveled under the trailer, which was not equipped with a side underrun protection system. A reconstruction of the accident estimated the driver would have had approximately 10.4 seconds to detect the truck and take evasive action. The underride collision sheared off the Tesla's greenhouse, destroying everything above the beltline, and caused fatal injuries to the driver. In the approximately nine seconds after colliding with the trailer, the Tesla traveled another 886.5 feet (270.2 m) and came to rest after colliding with two chain-link fences and a utility pole. The NHTSA's preliminary evaluation was opened to examine the design and performance of any automated driving systems in use at the time of the crash, which involves a population of an estimated 25,000 Model S cars. On July 8, 2016, the NHTSA requested Tesla Inc. to hand over to the agency detailed information about the design, operation and testing of its Autopilot technology. The agency also requested details of all design changes and updates to Autopilot since its introduction, and Tesla's planned updates scheduled for the next four months. According to Tesla, "neither autopilot nor the driver noticed the white side of the tractor-trailer against a brightly lit sky, so the brake was not applied." The car attempted to drive full speed under the trailer, "with the bottom of the trailer impacting the windshield of the Model S". Tesla also stated that this was Tesla's first known Autopilot-related death in over 130 million miles (208 million km) driven by its customers while Autopilot was activated. According to Tesla there is a fatality every 94 million miles (150 million km) among all type of vehicles in the U.S. It is estimated that billions of miles will need to be traveled before Tesla Autopilot can claim to be safer than humans with statistical significance. Researchers say that Tesla and others need to release more data on the limitations and performance of automated driving systems if self-driving cars are to become safe and understood enough for mass-market use. The truck's driver told the Associated Press that he could hear a Harry Potter movie playing in the crashed car, and said the car was driving so quickly that "he went so fast through my trailer I didn't see him. [The film] was still playing when he died and snapped a telephone pole a quarter-mile down the road." According to the Florida Highway Patrol, they found in the wreckage an aftermarket portable DVD player. (It is not possible to watch videos on the Model S touchscreen display while the car is moving.) A laptop computer was recovered during the post-crash examination of the wreck, along with an adjustable vehicle laptop mount attached to the front passenger's seat frame. The NHTSA concluded the laptop was probably mounted, and the driver may have been distracted at the time of the crash. In January 2017, the NHTSA Office of Defects Investigations (ODI) released a preliminary evaluation, finding that the driver in the crash had seven seconds to see the truck and identifying no defects in the Autopilot system; the ODI also found that the Tesla car crash rate dropped by 40 percent after Autosteer installation, but later also clarified that it did not assess the effectiveness of this technology or whether it was engaged in its crash rate comparison. The NHTSA Special Crash Investigation team published its report in January 2018. According to the report, for the drive leading up to the crash, the driver engaged Autopilot for 37 minutes and 26 seconds, and the system provided 13 "hands not detected" alerts, to which the driver responded after an average delay of 16 seconds. The report concluded "Regardless of the operational status of the Tesla's ADAS technologies, the driver was still responsible for maintaining ultimate control of the vehicle. All evidence and data gathered concluded that the driver neglected to maintain complete control of the Tesla leading up to the crash." In July 2016, the NTSB announced it had opened a formal investigation into the fatal accident while Autopilot was engaged. The NTSB is an investigative body that only has the power to make policy recommendations. An agency spokesman said, "It's worth taking a look and seeing what we can learn from that event, so that as that automation is more widely introduced we can do it in the safest way possible." The NTSB opens annually about 25 to 30 highway investigations. In September 2017, the NTSB released its report, determining that "the probable cause of the Williston, Florida, crash was the truck driver's failure to yield the right of way to the car, combine
Artificial intelligence in customer experience
Artificial intelligence in customer experience is the use and development of artificial intelligence (AI) to aid and improve customer experience (sometimes abbreviated to CX AI). Chatbots are often seen as the first step in the development of AI within the industry, but more tailored offerings are slowly becoming available. The use of artificial intelligence in the space has since become more diverse than simply chatbots, with AI underpinning entire CX cloud platforms now used at major corporations. Contact center as a service (CCaaS) has become a core solution of the CX (customer experience) industry, with the CCaaS market size expected to reach $17.19 Billion by 2030 in the United States alone. == History == As with many AI applications, CX AI early implementation case studies have demonstrated that AI can increase the quality of customer interactions and therefore the overall experience that organizations can provide. This in turn has suggested a higher return on investment and/or revenue as a result. The beginning of the revolution of customer experience and the use of machine learning was with chatbots. The use of this type of AI can be traced back to Alan Turing in 1950, when the Church–Turing thesis suggested that computers could use "formal reasoning" to reach conclusions. In 2017, Meta produced one of the first breakthroughs for everyday use of AI for customer experience when it allowed Facebook users to create their own messaging bots for free on its Facebook messenger platform. The main focus of this was to both automate and improve customer experience and interaction. In 2023, CCaaS vendors began announcing the integration of ChatGPT’s generative AI into their CX solutions. Generative AI adds a layer of semantics into AI outputs. This was a major breakthrough for conversational AI. Using natural language processing and conversational AI, chatbots could enhance the level of service they could provide, speaking to customers in an easy-to-understand and conversational tone. == Applications == Currently the main location for the application of CX AI in the sector is in contact centers. Historically, contact centers were simply known as call centers, but in recent years differentiation developed between the two terms. Call centers provide phone support, while contact centers also provide support via digital channels in addition to analogue phone systems. Contact centers are therefore seen as a complete customer service solution, where as call centers simply cover one aspect of customer interactions. As a part of improving CX, AI is also improving the employee experience. AI is able to automate tasks to free up time for contact center agents to focus on higher priority tasks. For example, AI can be used for auto summarization. This means that instead of human agents having to summarize customer interactions now AI can do it, saving organizations time and money.
Document AI
Document AI, also known as Document Intelligence, refers to a field of technology that employs machine learning (ML) techniques, such as natural language processing (NLP). These techniques are used to develop computer models capable of analyzing documents in a manner akin to human review. Through NLP, computer systems are able to understand relationships and contextual nuances in document contents, which facilitates the extraction of information and insights. Additionally, this technology enables the categorization and organization of the documents themselves. The applications of Document AI extend to processing and parsing a variety of semi-structured documents, such as forms, tables, receipts, invoices, tax forms, contracts, loan agreements, and financial reports. == Key features == Machine learning is utilized in Document AI to extract information from both printed and digital documents. This technology recognizes images, text, and characters in various languages, aiding in the extraction of insights from unstructured documents. The use of this technology can improve the speed and quality of decision-making in document analysis. Additionally, the automation of data extraction and validation can contribute to increased efficiency in document analysis processes. Since the early 2020s, the integration of large language models has extended Document AI beyond extraction toward generative tasks, including the automated drafting of forms, contracts, and document summaries. == Example == A business letter contains information in the form of text, as well as other types of information, such as the position of the text. For instance, a typical letter contains two addresses before the body of the text. The address at the very top (sometimes aligned to the right) is the sender address. This is normally followed by the date of the letter, with the place of writing. After this, the receiver address is listed. The distinction between the sender address and the receiver address is conveyed solely by the position of the address on the page, i.e. there is no textual indication like Sender: in front of the addresses. == Data dimensions and ML architecture == Data is typically distinguished into spatial data and time-series data, the former includes things like images, maps and graphs, while the latter includes signals such as stock prices or voice recordings. Document AI combines text data, which has a time dimension, with other types of data, such as the position of an address in a business letter, which is spatial. Historically in machine learning spatial data was analyzed using a convolutional neural network, and temporal data using a recurrent neural network. With the advent of dimension-type agnostic transformer architecture, these two different types of dimension can be more easily combined, Document AI is an example of this. == Benchmarks == Several public datasets are used to evaluate Document AI systems. FUNSD (Form Understanding in Noisy Scanned Documents) contains 199 annotated forms with token- and block-level labels for form understanding tasks. CORD (Consolidated Receipt Dataset) supports key information extraction from receipts. DocVQA contains approximately 50,000 questions over 12,000 document images for layout-aware visual question answering. == Common uses == Document AI systems are used to automate document processing and information extraction in business and financial workflows, including invoice and receipt processing, data entry automation, anomaly detection, mortgage processing, loan portfolio monitoring, credit risk management, and fraud detection such as counterfeit currency and fraudulent checks. They are also applied in regulatory compliance and contract analysis, including assessing changes in legal and regulatory documents. In real estate, Document AI supports document classification and structured information extraction for standardized processing and analytics. With the adoption of generative AI, Document AI systems can also generate and pre-fill structured documents such as contracts or business forms from natural language prompts.
Visual hull
A visual hull is a geometric entity created by shape-from-silhouette 3D reconstruction technique introduced by A. Laurentini. This technique assumes the foreground object in an image can be separated from the background. Under this assumption, the original image can be thresholded into a foreground/background binary image, which we call a silhouette image. The foreground mask, known as a silhouette, is the 2D projection of the corresponding 3D foreground object. Along with the camera viewing parameters, the silhouette defines a back-projected generalized cone that contains the actual object; this cone is called a silhouette cone. The intersection of the two silhouette cones defines a visual hull. which is a bounding geometry of the actual 3D object. When the reconstructed geometry is only used for rendering from a different viewpoint, the implicit reconstruction together with rendering can be done using graphics hardware. == In two dimensions == A technique used in some modern touchscreen devices employs cameras placed in the corners situated opposite infrared LEDs. The one-dimensional projection (shadow) of objects on the surface may be used to reconstruct the convex hull of the object. Visual hull generation method has also been used within experimental tele-meeting systems that aim to allow a user in a remote location to interact with virtual objects. The method uses multiple cameras to capture the real-world movements and interactions of the "sender", employing hardware-accelerated volumetric visual hull representation to create 3D volume from 2D multi-view images. Its ultimate aim is to allow 3D collaboration between the two users in the virtual realm, with the visual hull technique reducing the computational power required to allow this type of interaction and enabling the use of consumer goods such as the Wii Remote as a tool for interaction.
Tensor network
Tensor networks or tensor network states are a class of variational wave functions used in the study of many-body quantum systems and fluids. Tensor networks extend one-dimensional matrix product states to higher dimensions while preserving some of their useful mathematical properties. The wave function is encoded as a tensor contraction of a network of individual tensors. The structure of the individual tensors can impose global symmetries on the wave function (such as antisymmetry under exchange of fermions) or restrict the wave function to specific quantum numbers, like total charge, angular momentum, or spin. It is also possible to derive strict bounds on quantities like entanglement and correlation length using the mathematical structure of the tensor network. This has made tensor networks useful in theoretical studies of quantum information in many-body systems. They have also proved useful in variational studies of ground states, excited states, and dynamics of strongly correlated many-body systems. == Diagrammatic notation == In general, a tensor network diagram (Penrose diagram) can be viewed as a graph where nodes (or vertices) represent individual tensors, while edges represent summation over an index. Free indices are depicted as edges (or legs) attached to a single vertex only. Sometimes, there is also additional meaning to a node's shape. For instance, one can use trapezoids for unitary matrices or tensors with similar behaviour. This way, flipped trapezoids would be interpreted as complex conjugates to them. == History == Foundational research on tensor networks began in 1971 with a paper by Roger Penrose. In "Applications of negative dimensional tensors" Penrose developed tensor diagram notation, describing how the diagrammatic language of tensor networks could be used in applications in physics. In 1992, Steven R. White developed the density matrix renormalization group (DMRG) for quantum lattice systems. The DMRG was the first successful tensor network and associated algorithm. In 2002, Guifré Vidal and Reinhard Werner attempted to quantify entanglement, laying the groundwork for quantum resource theories. This was also the first description of the use of tensor networks as mathematical tools for describing quantum systems. In 2004, Frank Verstraete and Ignacio Cirac developed the theory of matrix product states, projected entangled pair states, and variational renormalization group methods for quantum spin systems. In 2006, Vidal developed the multi-scale entanglement renormalization ansatz (MERA). In 2007 he developed entanglement renormalization for quantum lattice systems. In 2010, Ulrich Schollwock developed the density-matrix renormalization group for the simulation of one-dimensional strongly correlated quantum lattice systems. In 2014, Román Orús introduced tensor networks for complex quantum systems and machine learning, as well as tensor network theories of symmetries, fermions, entanglement and holography. == Connection to machine learning == Tensor networks have been adapted for supervised learning, taking advantage of similar mathematical structure in variational studies in quantum mechanics and large-scale machine learning. This crossover has spurred collaboration between researchers in artificial intelligence and quantum information science. In June 2019, Google, the Perimeter Institute for Theoretical Physics, and X (company), released TensorNetwork, an open-source library for efficient tensor calculations. The main interest in tensor networks and their study from the perspective of machine learning is to reduce the number of trainable parameters (in a layer) by approximating a high-order tensor with a network of lower-order ones. Using the so-called tensor train technique (TT), one can reduce an N-order tensor (containing exponentially many trainable parameters) to a chain of N tensors of order 2 or 3, which gives us a polynomial number of parameters.