A graphics processing unit (GPU) is a specialized electronic circuit designed for digital image processing and to accelerate computer graphics, being present either as a component on a discrete graphics card or embedded on motherboards, mobile phones, personal computers, workstations, and game consoles. GPUs are increasingly being used for artificial intelligence (AI) processing due to linear algebra acceleration, which is also used extensively in graphics processing. Although there is no single definition of the term, and it may be used to describe any video display system, in modern use a GPU includes the ability to internally perform the calculations needed for various graphics tasks, like rotating and scaling 3D images, and often the additional ability to run custom programs known as shaders. This contrasts with earlier graphics controllers known as video display controllers which had no internal calculation capabilities, or blitters, which performed only basic memory movement operations. The modern GPU emerged during the 1990s, adding the ability to perform operations like drawing lines and text without CPU help, and later adding 3D functionality. Graphics functions are generally independent and this lends these tasks to being implemented on separate calculation engines. Modern GPUs include hundreds, or thousands, of calculation units. This made them useful for non-graphic calculations involving embarrassingly parallel problems due to their parallel structure. The ability of GPUs to rapidly perform vast numbers of calculations has led to their adoption in diverse fields including artificial intelligence (AI) where they excel at handling data-intensive and computationally demanding tasks. Other non-graphical uses include the training of neural networks and cryptocurrency mining. == History == === 1960s === Dedicated 3D graphics hardware dates back to graphic terminals such as the Adage AGT-30 from 1967 with analog matrix processors. In 1969 Evans & Sutherland (E&S) introduced the Line Drawing System-1 (LDS-1), which was the first all-digital system to provide matrix multiplication. Also in 1969, the low-cost graphics terminal IMLAC PDS-1 was introduced. It later saw use as an early 3D gaming machine with the likes of Maze War. === 1970s === In professional hardware, in 1972 PLATO IV system becomes operational at the University of Illinois Urbana-Champaign. Between around 1973 and 1978, several networked multiplayer wireframe 3D games are implemented and popularized by users of the system. Also in 1972, the E&S Continuous Tone 1 (CT1) "Watkins box" system (consisting of an E&S LDS-2 and Shaded Picture System) is delivered to Case Western Reserve University. It offered the first real-time Gouraud shading. In 1975, a joint effort between Evans & Sutherland Computer Corporation and the University of Utah's computer graphics department results in the first ever MOSFET video framebuffer, capable of color and smooth shading. E&S Continuous Tone 3 (CT3) system was delivered in 1977 to Lufthansa for pilot training using computer simulation. It was the first graphics system capable of real-time texture mapping. Ikonas made graphics systems with 8- and 24-bit graphics and 3D acceleration in the late 70s. Arcade system boards have used specialized 2D graphics circuits since the 1970s. In early video game hardware, RAM for frame buffers was expensive, so video chips composited data together as the display was being scanned out on the monitor. A specialized barrel shifter circuit helped the CPU animate the framebuffer graphics for various 1970s arcade video games from Midway and Taito, such as Gun Fight (1975), Sea Wolf (1976), and Space Invaders (1978). The Namco Galaxian arcade system in 1979 used specialized graphics hardware that supported RGB color, multi-colored sprites, and tilemap backgrounds. The Galaxian hardware was widely used during the golden age of arcade video games, by game companies such as Namco, Centuri, Gremlin, Irem, Konami, Midway, Nichibutsu, Sega, and Taito. The Atari 2600 in 1977 used a video shifter called the Television Interface Adaptor. Atari 8-bit computers (1979) had ANTIC, a video processor which interpreted instructions describing a "display list"—the way the scan lines map to specific bitmapped or character modes and where the memory is stored (so there did not need to be a contiguous frame buffer). 6502 machine code subroutines could be triggered on scan lines by setting a bit on a display list instruction. ANTIC also supported smooth vertical and horizontal scrolling independent of the CPU. === 1980s === In the 1980s significant advancements were made in professional 3D graphics hardware. Perhaps most impactful was the 1981 development of the Geometry Engine, a VLSI vector processor ASIC designed by Jim Clark and Marc Hannah at Stanford University. This processor is the forerunner of modern tensor cores and other similar processors marketed for graphics and AI. The Geometry Engine went on to be used in Silicon Graphics workstations for many years. Silicon Graphics's first product, shipped in November 1983, was the IRIS 1000, a terminal with hardware-accelerated 3D graphics based on the Geometry Engine. The Geometry Engine was capable of approximately 6 million operations per second. The 1981 NEC μPD7220 was the first implementation of a personal computer graphics display processor as a single large-scale integration (LSI) integrated circuit chip. This enabled the design of low-cost, high-performance video graphics cards such as those from Number Nine Visual Technology. It became the best-known GPU until the mid-1980s. It was the first fully integrated VLSI (very large-scale integration) metal–oxide–semiconductor (NMOS) graphics display processor for PCs, supported up to 1024×1024 resolution, and laid the foundations for the PC graphics market. It was used in a number of graphics cards and was licensed for clones such as the Intel 82720, the first of Intel's graphics processing units. The Williams Electronics arcade games Robotron: 2084, Joust, Sinistar, and Bubbles, all released in 1982, contain custom blitter chips for operating on 16-color bitmaps. In 1984, Hitachi released the ARTC HD63484, the first major CMOS graphics processor for personal computers. The ARTC could display up to 4K resolution when in monochrome mode. It was used in a number of graphics cards and terminals during the late 1980s. In 1985, the Amiga was released with a custom graphics chip called Agnus including a blitter for bitmap manipulation, line drawing, and area fill. It also included a coprocessor with its own simple instruction set, that was capable of manipulating graphics hardware registers in sync with the video beam (e.g. for per-scanline palette switches, sprite multiplexing, and hardware windowing), or driving the blitter. Also in 1985, IBM released the Professional Graphics Controller, designed by later to be Nvidia co-founder Curtis Priem, which was a rudimentary 3D card with 640 × 480 256-color graphics which used a dedicated CPU to draw graphics independently of the main system. It was used as the basis of cards by a number of makers (including Matrox) and its analog RGB signaling led directly to the VGA video standard. Priem later in the 80s worked on the influential Sun Microsystems GX (also known as cgsix) accelerated 2D graphics card. In 1986, Texas Instruments released the TMS34010, the first fully programmable graphics processor. It could run general-purpose code but also had a graphics-oriented instruction set. During 1990–1992, this chip became the basis of the Texas Instruments Graphics Architecture ("TIGA") Windows accelerator cards. Following in 1987, the IBM 8514 graphics system was released. It was one of the first video cards for IBM PC compatibles that implemented fixed-function 2D primitives in electronic hardware. Sharp's X68000, released in 1987, used a custom graphics chipset with a 65,536 color palette and hardware support for sprites, scrolling, and multiple playfields. It served as a development machine for Capcom's CP System arcade board. Fujitsu's FM Towns computer, released in 1989, had support for a 16,777,216 color palette. For context, IBM also introduced its Video Graphics Array (VGA) display system in 1987, with a maximum resolution of 640 × 480 pixels. Unlike 8514/A, VGA had no hardware acceleration features. In November 1988, NEC Home Electronics announced its creation of the Video Electronics Standards Association (VESA) to develop and promote a Super VGA (SVGA) computer display standard as a successor to VGA. Super VGA enabled graphics display resolutions up to 800 × 600 pixels, a 56% increase. In 1988 SGI sold IRIS workstation graphics with 10-12 Geometry Engines and introduced the IrisVision add-in board for IBM MicroChannel bus (RS/6000) based on the Geometry Engine as well. In 1988 as well, the first dedicated polygonal 3D graphics boards in arcade machines were introduced wit
Weight initialization
In deep learning, weight initialization or parameter initialization describes the initial step in creating a neural network. A neural network contains trainable parameters that are modified during training: weight initialization is the pre-training step of assigning initial values to these parameters. The choice of weight initialization method affects the speed of convergence, the scale of neural activation within the network, the scale of gradient signals during backpropagation, and the quality of the final model. Proper initialization is necessary for avoiding issues such as vanishing and exploding gradients and activation function saturation. Note that even though this article is titled "weight initialization", both weights and biases are used in a neural network as trainable parameters, so this article describes how both of these are initialized. Similarly, trainable parameters in convolutional neural networks (CNNs) are called kernels and biases, and this article also describes these. == Constant initialization == We discuss the main methods of initialization in the context of a multilayer perceptron (MLP). Specific strategies for initializing other network architectures are discussed in later sections. For an MLP, there are only two kinds of trainable parameters, called weights and biases. Each layer l {\displaystyle l} contains a weight matrix W ( l ) ∈ R n l − 1 × n l {\displaystyle W^{(l)}\in \mathbb {R} ^{n_{l-1}\times n_{l}}} and a bias vector b ( l ) ∈ R n l {\displaystyle b^{(l)}\in \mathbb {R} ^{n_{l}}} , where n l {\displaystyle n_{l}} is the number of neurons in that layer. A weight initialization method is an algorithm for setting the initial values for W ( l ) , b ( l ) {\displaystyle W^{(l)},b^{(l)}} for each layer l {\displaystyle l} . The simplest form is zero initialization: W ( l ) = 0 , b ( l ) = 0 {\displaystyle W^{(l)}=0,b^{(l)}=0} Zero initialization is usually used for initializing biases, but it is not used for initializing weights, as it leads to symmetry in the network, causing all neurons to learn the same features. In this page, we assume b = 0 {\displaystyle b=0} unless otherwise stated. Recurrent neural networks typically use activation functions with bounded range, such as sigmoid and tanh, since unbounded activation may cause exploding values. (Le, Jaitly, Hinton, 2015) suggested initializing weights in the recurrent parts of the network to identity and zero bias, similar to the idea of residual connections and LSTM with no forget gate. In most cases, the biases are initialized to zero, though some situations can use a nonzero initialization. For example, in multiplicative units, such as the forget gate of LSTM, the bias can be initialized to 1 to allow good gradient signal through the gate. For neurons with ReLU activation, one can initialize the bias to a small positive value like 0.1, so that the gradient is likely nonzero at initialization, avoiding the dying ReLU problem. == Random initialization == Random initialization means sampling the weights from a normal distribution or a uniform distribution, usually independently. === LeCun initialization === LeCun initialization, popularized in (LeCun et al., 1998), is designed to preserve the variance of neural activations during the forward pass. It samples each entry in W ( l ) {\displaystyle W^{(l)}} independently from a distribution with mean 0 and variance 1 / n l − 1 {\displaystyle 1/n_{l-1}} . For example, if the distribution is a continuous uniform distribution, then the distribution is U ( ± 3 / n l − 1 ) {\displaystyle {\mathcal {U}}(\pm {\sqrt {3/n_{l-1}}})} . === Glorot initialization === Glorot initialization (or Xavier initialization) was proposed by Xavier Glorot and Yoshua Bengio. It was designed as a compromise between two goals: to preserve activation variance during the forward pass and to preserve gradient variance during the backward pass. For uniform initialization, it samples each entry in W ( l ) {\displaystyle W^{(l)}} independently and identically from U ( ± 6 / ( n l + 1 + n l − 1 ) ) {\displaystyle {\mathcal {U}}(\pm {\sqrt {6/(n_{l+1}+n_{l-1})}})} . In the context, n l − 1 {\displaystyle n_{l-1}} is also called the "fan-in", and n l + 1 {\displaystyle n_{l+1}} the "fan-out". When the fan-in and fan-out are equal, then Glorot initialization is the same as LeCun initialization. === He initialization === As Glorot initialization performs poorly for ReLU activation, He initialization (or Kaiming initialization) was proposed by Kaiming He et al. for networks with ReLU activation. It samples each entry in W ( l ) {\displaystyle W^{(l)}} from N ( 0 , 2 / n l − 1 ) {\displaystyle {\mathcal {N}}(0,2/n_{l-1})} . === Orthogonal initialization === (Saxe et al. 2013) proposed orthogonal initialization: initializing weight matrices as uniformly random (according to the Haar measure) semi-orthogonal matrices, multiplied by a factor that depends on the activation function of the layer. It was designed so that if one initializes a deep linear network this way, then its training time until convergence is independent of depth. Sampling a uniformly random semi-orthogonal matrix can be done by initializing X {\displaystyle X} by IID sampling its entries from a standard normal distribution, then calculate ( X X ⊤ ) − 1 / 2 X {\displaystyle \left(XX^{\top }\right)^{-1/2}X} or its transpose, depending on whether X {\displaystyle X} is tall or wide. For CNN kernels with odd widths and heights, orthogonal initialization is done this way: initialize the central point by a semi-orthogonal matrix, and fill the other entries with zero. As an illustration, a kernel K {\displaystyle K} of shape 3 × 3 × c × c ′ {\displaystyle 3\times 3\times c\times c'} is initialized by filling K [ 2 , 2 , : , : ] {\displaystyle K[2,2,:,:]} with the entries of a random semi-orthogonal matrix of shape c × c ′ {\displaystyle c\times c'} , and the other entries with zero. (Balduzzi et al., 2017) used it with stride 1 and zero-padding. This is sometimes called the Orthogonal Delta initialization. Related to this approach, unitary initialization proposes to parameterize the weight matrices to be unitary matrices, with the result that at initialization they are random unitary matrices (and throughout training, they remain unitary). This is found to improve long-sequence modelling in LSTM. Orthogonal initialization has been generalized to layer-sequential unit-variance (LSUV) initialization. It is a data-dependent initialization method, and can be used in convolutional neural networks. It first initializes weights of each convolution or fully connected layer with orthonormal matrices. Then, proceeding from the first to the last layer, it runs a forward pass on a random minibatch, and divides the layer's weights by the standard deviation of its output, so that its output has variance approximately 1. === Fixup initialization === In 2015, the introduction of residual connections allowed very deep neural networks to be trained, much deeper than the ~20 layers of the previous state of the art (such as the VGG-19). Residual connections gave rise to their own weight initialization problems and strategies. These are sometimes called "normalization-free" methods, since using residual connection could stabilize the training of a deep neural network so much that normalizations become unnecessary. Fixup initialization is designed specifically for networks with residual connections and without batch normalization, as follows: Initialize the classification layer and the last layer of each residual branch to 0. Initialize every other layer using a standard method (such as He initialization), and scale only the weight layers inside residual branches by L − 1 2 m − 2 {\displaystyle L^{-{\frac {1}{2m-2}}}} . Add a scalar multiplier (initialized at 1) in every branch and a scalar bias (initialized at 0) before each convolution, linear, and element-wise activation layer. Similarly, T-Fixup initialization is designed for Transformers without layer normalization. === Others === Instead of initializing all weights with random values on the order of O ( 1 / n ) {\displaystyle O(1/{\sqrt {n}})} , sparse initialization initialized only a small subset of the weights with larger random values, and the other weights zero, so that the total variance is still on the order of O ( 1 ) {\displaystyle O(1)} . Random walk initialization was designed for MLP so that during backpropagation, the L2 norm of gradient at each layer performs an unbiased random walk as one moves from the last layer to the first. Looks linear initialization was designed to allow the neural network to behave like a deep linear network at initialization, since W R e L U ( x ) − W R e L U ( − x ) = W x {\displaystyle W\;\mathrm {ReLU} (x)-W\;\mathrm {ReLU} (-x)=Wx} . It initializes a matrix W {\displaystyle W} of shape R n 2 × m {\displaystyle \mathbb {R} ^{{\frac {n}{2}}\times m}} by any method, such as orthogonal initialization, t
Gene expression programming
Gene expression programming (GEP) in computer programming is an evolutionary algorithm that creates computer programs or models. These computer programs are complex tree structures that learn and adapt by changing their sizes, shapes, and composition, much like a living organism. And like living organisms, the computer programs of GEP are also encoded in simple linear chromosomes of fixed length. Thus, GEP is a genotype–phenotype system, benefiting from a simple genome to keep and transmit the genetic information and a complex phenotype to explore the environment and adapt to it. == Background == Evolutionary algorithms use populations of individuals, select individuals according to fitness, and introduce genetic variation using one or more genetic operators. Their use in artificial computational systems dates back to the 1950s where they were used to solve optimization problems (e.g. Box 1957 and Friedman 1959). But it was with the introduction of evolution strategies by Rechenberg in 1965 that evolutionary algorithms gained popularity. A good overview text on evolutionary algorithms is the book "An Introduction to Genetic Algorithms" by Mitchell (1996). Gene expression programming belongs to the family of evolutionary algorithms and is closely related to genetic algorithms and genetic programming. From genetic algorithms it inherited the linear chromosomes of fixed length; and from genetic programming it inherited the expressive parse trees of varied sizes and shapes. In gene expression programming the linear chromosomes work as the genotype and the parse trees as the phenotype, creating a genotype/phenotype system. This genotype/phenotype system is multigenic, thus encoding multiple parse trees in each chromosome. This means that the computer programs created by GEP are composed of multiple parse trees. Because these parse trees are the result of gene expression, in GEP they are called expression trees. Masood Nekoei, et al. utilized this expression programming style in ABC optimization to conduct ABCEP as a method that outperformed other evolutionary algorithms.ABCEP == Encoding: the genotype == The genome of gene expression programming consists of a linear, symbolic string or chromosome of fixed length composed of one or more genes of equal size. These genes, despite their fixed length, code for expression trees of different sizes and shapes. An example of a chromosome with two genes, each of size 9, is the string (position zero indicates the start of each gene): 012345678012345678 L+a-baccdcLabacd where “L” represents the natural logarithm function and “a”, “b”, “c”, and “d” represent the variables and constants used in a problem. == Expression trees: the phenotype == As shown above, the genes of gene expression programming have all the same size. However, these fixed length strings code for expression trees of different sizes. This means that the size of the coding regions varies from gene to gene, allowing for adaptation and evolution to occur smoothly. For example, the mathematical expression: ( a − b ) ( c + d ) {\displaystyle {\sqrt {(a-b)(c+d)}}\,} can also be represented as an expression tree: where "Q” represents the square root function. This kind of expression tree consists of the phenotypic expression of GEP genes, whereas the genes are linear strings encoding these complex structures. For this particular example, the linear string corresponds to: 01234567 Q-+abcd which is the straightforward reading of the expression tree from top to bottom and from left to right. These linear strings are called k-expressions (from Karva notation). Going from k-expressions to expression trees is also very simple. For example, the following k-expression: 01234567890 Qb+baQba is composed of two different terminals (the variables “a” and “b”), two different functions of two arguments (“” and “+”), and a function of one argument (“Q”). Its expression gives: == K-expressions and genes == The k-expressions of gene expression programming correspond to the region of genes that gets expressed. This means that there might be sequences in the genes that are not expressed, which is indeed true for most genes. The reason for these noncoding regions is to provide a buffer of terminals so that all k-expressions encoded in GEP genes correspond always to valid programs or expressions. The genes of gene expression programming are therefore composed of two different domains – a head and a tail – each with different properties and functions. The head is used mainly to encode the functions and variables chosen to solve the problem at hand, whereas the tail, while also used to encode the variables, provides essentially a reservoir of terminals to ensure that all programs are error-free. For GEP genes the length of the tail is given by the formula: t = h ( n max − 1 ) + 1 {\displaystyle t=h(n_{\max }-1)+1} where h is the head's length and nmax is maximum arity. For example, for a gene created using the set of functions F = {Q, +, −, ∗, /} and the set of terminals T = {a, b}, nmax = 2. And if we choose a head length of 15, then t = 15 (2–1) + 1 = 16, which gives a gene length g of 15 + 16 = 31. The randomly generated string below is an example of one such gene: 0123456789012345678901234567890 b+a-aQab+//+b+babbabbbababbaaa It encodes the expression tree: which, in this case, only uses 8 of the 31 elements that constitute the gene. It's not hard to see that, despite their fixed length, each gene has the potential to code for expression trees of different sizes and shapes, with the simplest composed of only one node (when the first element of a gene is a terminal) and the largest composed of as many nodes as there are elements in the gene (when all the elements in the head are functions with maximum arity). It's also not hard to see that it is trivial to implement all kinds of genetic modification (mutation, inversion, insertion, recombination, and so on) with the guarantee that all resulting offspring encode correct, error-free programs. == Multigenic chromosomes == The chromosomes of gene expression programming are usually composed of more than one gene of equal length. Each gene codes for a sub-expression tree (sub-ET) or sub-program. Then the sub-ETs can interact with one another in different ways, forming a more complex program. The figure shows an example of a program composed of three sub-ETs. In the final program the sub-ETs could be linked by addition or some other function, as there are no restrictions to the kind of linking function one might choose. Some examples of more complex linkers include taking the average, the median, the midrange, thresholding their sum to make a binomial classification, applying the sigmoid function to compute a probability, and so on. These linking functions are usually chosen a priori for each problem, but they can also be evolved elegantly and efficiently by the cellular system of gene expression programming. == Cells and code reuse == In gene expression programming, homeotic genes control the interactions of the different sub-ETs or modules of the main program. The expression of such genes results in different main programs or cells, that is, they determine which genes are expressed in each cell and how the sub-ETs of each cell interact with one another. In other words, homeotic genes determine which sub-ETs are called upon and how often in which main program or cell and what kind of connections they establish with one another. === Homeotic genes and the cellular system === Homeotic genes have exactly the same kind of structural organization as normal genes and they are built using an identical process. They also contain a head domain and a tail domain, with the difference that the heads contain now linking functions and a special kind of terminals – genic terminals – that represent the normal genes. The expression of the normal genes results as usual in different sub-ETs, which in the cellular system are called ADFs (automatically defined functions). As for the tails, they contain only genic terminals, that is, derived features generated on the fly by the algorithm. For example, the chromosome in the figure has three normal genes and one homeotic gene and encodes a main program that invokes three different functions a total of four times, linking them in a particular way. From this example it is clear that the cellular system not only allows the unconstrained evolution of linking functions but also code reuse. And it shouldn't be hard to implement recursion in this system. === Multiple main programs and multicellular systems === Multicellular systems are composed of more than one homeotic gene. Each homeotic gene in this system puts together a different combination of sub-expression trees or ADFs, creating multiple cells or main programs. For example, the program shown in the figure was created using a cellular system with two cells and three normal genes. The applications of these multicellular systems are mu
2024 Abu Dhabi Autonomous Racing League
On 27 April 2024, the inaugural race of the Abu Dhabi Autonomous Racing League was held at the Yas Marina Circuit in Abu Dhabi. The race, originally scheduled to last eight laps, was ultimately shortened to six laps due to various complications, including subpar performance. It involved four self-driving race cars, only two of which – German cars Hailey and Constructor AI – finished the race; the other two did not finish. == Background == === Abu Dhabi Autonomous Racing League (A2RL) === The A2RL is an autonomous racing championship based in Abu Dhabi and organized by ASPIRE, part of the Advanced Technology Research Council. It is one of two active autonomous car racing championships, the second being the US-based Indy Autonomous Challenge. Unlike the IAC, which primarily focuses on time trials, simulated races, and challenges for teams, the A2RL's car races are closer to a standard grand prix formula race format. Both use Dallara-supplied racecars; the IAC uses the AV-24 chassis derived from Indy NXT's IL-15, while the A2RL chassis is designated EAV-24 and is derived from the SF-23 chassis used in Japanese Super Formula races. === Entrants === In total, eight teams were part of the A2RL in 2024, but only four would compete in the race proper. The list of teams in 2024 is: Fly Eagle (China/UAE) Code19 Racing (United States) Constructor University (Germany) Kinetiz (Singapore/UAE) Humda Lab (Hungary) PoliMove (Italy) Unimore (Italy) Technical University of Munich (Germany) Most teams come from universities and many, such as PoliMove and TUM, already have experience with autonomous racing, primarily from competing in the IAC. All teams had two months to code and test their AIs. Unlike most international open-wheel racing tournaments, such as Formula 1 or Formula E, no free practice sessions were undertaken. === TII Pre-race demonstration === Prior to the race itself, a mock 1v1 duel between former F1 driver Danill Kvyat and a self-driving car from the non-competing TII Racing team took place; the autonomous car was green and had number 01, while Kvyat's car was red and had number 00. Kvyat spent most of the duel in the pits. Kvyat himself said: "I'm not racing autonomous cars here. It won't be a flat-out race". == Qualifying == === Qualifying report === As only four of the eight entrants would compete in the main event, qualifying time trials were held to determine the four main race competitors, as well as their positions in the grid. Only the cars with the four best lap times over three time trial sessions held on Friday and Saturday would qualify. Multiple errors and setbacks occurred during qualifying. In the first session, Maveric AI, Code19's car, left the track and stopped just after turn 14 due to connectivity issues. Fly Eagle's car, Feiying, had multiple upsets; at one point, Feiying ran into localization issues and began swerving left and right before stopping just before turn 10. Later, Feiying swerved again and nearly hit the wall at the back straight, near the support pits, due to further localization issues. Sparkz, the Kinetiz team's car, swerved and crashed into the wall near yacht berths 51-56 after turn 11, damaging the front right wheel's axle and partially detaching the forward wings. Sparkz would be the only car to not have a set time at the end of the time trials. PoliMove car Eva braked hard without warning at the straight, the LED status indicator turning off, suggesting the AI computer had a system crash or shut itself down. After the sun went down, during the second session, Hailey, the car from the TUM team, went off-track after turn 9 and stopped, its status indicator flashing red, meaning Hailey's AI disengaged itself. Eva had further issues, once again braking hard and spinning out into turn 1. Later, the same thing happened to Feiying; it later swerved left and right and stopped due to further localization issues. The morning after, during the third and final session, Hailey went off-track after turn 5, and were unable to regain the pole position. === Qualifying classification === == Attack/Defend challenge == === Attack/Defend challenge report === In this part of the event, cars would be put on a series of 1v1 duels to see how well they could defend their position or attack to gain one higher. During one such duel, an incident occurred where Hailey rear-ended Eva, sending both off the track and prematurely ending the duel. The challenge was otherwise uneventful. === Attack/Defend challenge results === == Main race == === Race report === Eventually, at around 20:30 Gulf Standard Time on the night of 27 April, the main event (termed the "Grand Final" on-stream) would begin. The starting order was Eva first, Gianna second, Hailey third, and Constructor AI last. The race began with a rolling start. As a safety measure, the first two laps were conducted under virtual safety car (VSC) to make sure the cars stayed together, making them de facto formation laps, even if they counted towards race distance. However, Hailey ended up stopping at the final turn and strayed too far from the cars ahead, and as a result, the VSC conditions were extended for another lap. According to the livestream's on-screen graphics, Hailey was upwards of one minute and 22.3 seconds behind Gianna after the former started moving again. On lap 4, halfway through the planned race, and with Hailey more than 30 seconds behind Gianna, the VSC was lifted, and the green flag finally dropped. At first, the two Italian cars were leading the pack, Eva was the race leader with Gianna 3.2 seconds behind, however, as it entered the chicane, Eva hit the brakes and spun out, with Gianna briefly stopping as it passed Eva. Eva's spin automatically triggered a full-course yellow flag. Normally, under yellow flag conditions, overtaking is not permitted, but with Eva stopped and being moved off the track, it was theoretically permitted to overtake Eva. However, presumably due to an oversight in the AI's code, the cars assumed overtaking Eva, despite being off the track, was not permitted. As a result, both Gianna and Constructor AI stopped as they did not want to overtake Eva due to the yellow flag, with Hailey following suit as it approached. Constructor AI's status indicator was solid red, suggesting the AI had disengaged; however, Gianna's status indicator remained solid purple, showing the AI was still in control. Eva's status indicator was also solid purple, but was soon flashing green, suggesting the AI had disengaged but was ready to take control again. With all cars stalled, and Eva being off the track, the race was effectively red-flagged and suspended. Hailey, Gianna, and Constructor AI drove themselves back to their team's pits; Eva did not, it was towed to the main pits on a flatbed truck. Constructor was the first to arrive at the pits, followed by Gianna and Hailey, in that order. This incident, combined with loss of internet connection, led to Eva retiring - it did not finish the race. Eventually, it was decided to resume the race. With Eva retired, the restart order was Gianna first, Hailey second, and Constructor AI third. The race was also shortened - from eight laps to six. With lap 5 under full-course yellow, this meant all three remaining teams would effectively restart the race on the sixth and final lap. The trio left the pits at 22:25 Gulf Standard Time, and the race resumed two minutes later. At first, Gianna was winning with Hailey 2.6 seconds behind, but then Gianna stopped on turn 5, giving Hailey the lead. Constructor AI also overtook Gianna, but not without briefly stopping. Gianna remained stopped, its status indicator solid red - it did not finish either. With both Italian teams out of the picture, Hailey finished first and won A2RL 2024, with Constructor AI finishing second, 27.2 seconds behind. === Final race classification ===
Imagen (text-to-image model)
Imagen is a series of text-to-image models developed by Google DeepMind. They were developed by Google Brain until the company's merger with DeepMind in April 2023. Imagen is primarily used to generate images from text prompts, similar to Stability AI's Stable Diffusion, OpenAI's DALL-E, or Midjourney. The original version of the model was first discussed in a paper from May 2022. The tool produces high-quality images and is available to all users with a Google account through services including Gemini, ImageFX, and Vertex AI. == History == Imagen's original version was first presented in a paper published in May 2022. It featured the ability to generate high-fidelity images from natural language. The second version, Imagen 2 was released in December 2023. The standout feature was text and logo generation. Imagen 3 was released in August 2024. Google claims that the newest version provides better detail and lighting on generated images. On 20 May 2025 at Google I/O 2025 the company released an improved model, Imagen 4. == Technology == Imagen uses two key technologies. The first is the use of transformer-based large language models, notably T5, to understand text and subsequently encode text for image synthesis. The second is the use of cascaded diffusion models providing high-fidelity image generation. Imagen generates image in three stages, starting from a base of 64x64, then upsampled to 256x256 and 1024x1024. Imagen 4 generates image up to 2k. == Capabilities == Imagen can generate photorealistic images from text prompts. It can also create various styles, such as cinematic, 35mm film, illustration, and surreal. Like most text-to-image generative AI models, Imagen has difficulty rendering human fingers, text, ambigrams and other forms of typography. The model can generate images in five aspect ratios, namely 9:16, 3:4, 1:1, 4:3, and 16:9. Imagen can also refine already generated images by editing existing text prompts.
BBC Own It
The BBC Own It app was a British information site designed to protect and support children using the Internet. The app was launched in 2017 and retired in 2022, though the website retired in 2024 and has since moved to BBC Teach. As part of the BBC's partnership with Internet Matters, the not-for-profit contributed to content on the BBC Own It website. == History == In 2016, The Royal Foundation of The Duke and Duchess of Cambridge established The Royal Foundation Taskforce on the Prevention of Cyberbullying. Work began in 2017 by the BBC to create an app about cyberbullying and online safety (later titled Own It) in response to a call for action from the Taskforce. In December 2017, the BBC launched Own It. In November 2018, work on the BBC Own It App was announced by Prince William. In September 2019, the BBC Own It App was launched into the AppStore and Google Play. In 2022, the BBC discontinued the app, although the website was still active, however in 2024, the website was discontinued, and now any links to the website now redirect to a BBC Teach page. == Awards == UXUK award for Best Education or Learning Experience (2019) Banff World Media Festival Rockies Award for Children & Youth Interactive Content (2020) CogX Award for Best Innovation In Natural Language Processing (2020)
Argument mining
Argument mining, or argumentation mining, is a research area within the natural language processing field. The goal of argument mining is the automatic extraction and identification of argumentative structures from natural language text with the aid of computer programs. Such argumentative structures include the premise, conclusions, the argument scheme and the relationship between the main and subsidiary argument, or the main and counter-argument within discourse. The Argument Mining workshop series is the main research forum for argument mining related research. == Applications == Argument mining has been applied in many different genres including the qualitative assessment of social media content (e.g. Twitter, Facebook), where it provides a powerful tool for policy-makers and researchers in social and political sciences. Other domains include legal documents, product reviews, scientific articles, online debates, newspaper articles and dialogical domains. Transfer learning approaches have been successfully used to combine the different domains into a domain agnostic argumentation model. Argument mining has been used to provide students individual writing support by accessing and visualizing the argumentation discourse in their texts. The application of argument mining in a user-centered learning tool helped students to improve their argumentation skills significantly compared to traditional argumentation learning applications. == Challenges == Given the wide variety of text genres and the different research perspectives and approaches, it has been difficult to reach a common and objective evaluation scheme. Many annotated data sets have been proposed, with some gaining popularity, but a consensual data set is yet to be found. Annotating argumentative structures is a highly demanding task. There have been successful attempts to delegate such annotation tasks to the crowd but the process still requires a lot of effort and carries significant cost. Initial attempts to bypass this hurdle were made using the weak supervision approach.