Frederick J. Damerau (December 25, 1931 – January 27, 2009) was a pioneer of research on natural language processing and data mining. After earning his B.A. from Cornell University in 1953, he spent most of his career at IBM, in the Thomas J. Watson Research Center. He holds a PhD from Yale University. One of his most influential and ground-breaking papers was "A technique for computer detection and correction of spelling errors" published in 1964. He also developed and patented for IBM the first algorithm for placing hyphens automatically in words. In 1971 he published the book "Markov Models and Linguistic Theory : An Experimental Study of a Model for English." After being active in research for over four decades, Fred Damerau died on January 27, 2009.
Kdan Mobile
Kdan Mobile Software Limited is a software application development company based in Tainan City, Taiwan. Kdan also has branches in Taipei, Changsha, Irvine, California, Japan, and South Korea. The company was founded in 2009 by Kenny Su, the company's CEO. == History == Kdan Mobile was founded in 2009 by Kenny Su (蘇柏州) and develops an application for PDF documents. Su previously worked at the Industrial Technology Research Institute (ITRI) . In 2018, the company completed its Series B round of fundraising, in which it raised 16 million USD in total. Four global firms, Dattoz Partners (South Korea), WI Harper Group (U.S.), Taiwania Capital (Taiwan), and Golden Asia Fund Mitsubishi UFJ Capital (Japan), made up the Series B investment. Kdan previously raised 5 million USD in its Series A round in 2018.
Apprenticeship learning
In artificial intelligence, apprenticeship learning (or learning from demonstration or imitation learning) is the process of learning by observing an expert. It can be viewed as a form of supervised learning, where the training dataset consists of task executions by a demonstration teacher. == Mapping function approach == Mapping methods try to mimic the expert by forming a direct mapping either from states to actions, or from states to reward values. For example, in 2002 researchers used such an approach to teach an AIBO robot basic soccer skills. === Inverse reinforcement learning approach === Inverse reinforcement learning (IRL) is the process of deriving a reward function from observed behavior. While ordinary "reinforcement learning" involves using rewards and punishments to learn behavior, in IRL the direction is reversed, and a robot observes a person's behavior to figure out what goal that behavior seems to be trying to achieve. The IRL problem can be defined as: Given 1) measurements of an agent's behaviour over time, in a variety of circumstances; 2) measurements of the sensory inputs to that agent; 3) a model of the physical environment (including the agent's body): Determine the reward function that the agent is optimizing. IRL researcher Stuart J. Russell proposes that IRL might be used to observe humans and attempt to codify their complex "ethical values", in an effort to create "ethical robots" that might someday know "not to cook your cat" without needing to be explicitly told. The scenario can be modeled as a "cooperative inverse reinforcement learning game", where a "person" player and a "robot" player cooperate to secure the person's implicit goals, despite these goals not being explicitly known by either the person nor the robot. In 2017, OpenAI and DeepMind applied deep learning to the cooperative inverse reinforcement learning in simple domains such as Atari games and straightforward robot tasks such as backflips. The human role was limited to answering queries from the robot as to which of two different actions were preferred. The researchers found evidence that the techniques may be economically scalable to modern systems. Apprenticeship via inverse reinforcement learning (AIRP) was developed by in 2004 Pieter Abbeel, Professor in Berkeley's EECS department, and Andrew Ng, Associate Professor in Stanford University's Computer Science Department. AIRP deals with "Markov decision process where we are not explicitly given a reward function, but where instead we can observe an expert demonstrating the task that we want to learn to perform". AIRP has been used to model reward functions of highly dynamic scenarios where there is no obvious reward function intuitively. Take the task of driving for example, there are many different objectives working simultaneously - such as maintaining safe following distance, a good speed, not changing lanes too often, etc. This task, may seem easy at first glance, but a trivial reward function may not converge to the policy wanted. One domain where AIRP has been used extensively is helicopter control. While simple trajectories can be intuitively derived, complicated tasks like aerobatics for shows has been successful. These include aerobatic maneuvers like - in-place flips, in-place rolls, loops, hurricanes and even auto-rotation landings. This work was developed by Pieter Abbeel, Adam Coates, and Andrew Ng - "Autonomous Helicopter Aerobatics through Apprenticeship Learning" === System model approach === System models try to mimic the expert by modeling world dynamics. == Plan approach == The system learns rules to associate preconditions and postconditions with each action. In one 1994 demonstration, a humanoid learns a generalized plan from only two demonstrations of a repetitive ball collection task. == Example == Learning from demonstration is often explained from a perspective that the working Robot-control-system is available and the human-demonstrator is using it. And indeed, if the software works, the Human operator takes the robot-arm, makes a move with it, and the robot will reproduce the action later. For example, he teaches the robot-arm how to put a cup under a coffeemaker and press the start-button. In the replay phase, the robot is imitating this behavior 1:1. But that is not how the system works internally; it is only what the audience can observe. In reality, Learning from demonstration is much more complex. One of the first works on learning by robot apprentices (anthropomorphic robots learning by imitation) was Adrian Stoica's PhD thesis in 1995. In 1997, robotics expert Stefan Schaal was working on the Sarcos robot-arm. The goal was simple: solve the pendulum swingup task. The robot itself can execute a movement, and as a result, the pendulum is moving. The problem is, that it is unclear what actions will result into which movement. It is an Optimal control-problem which can be described with mathematical formulas but is hard to solve. The idea from Schaal was, not to use a Brute-force solver but record the movements of a human-demonstration. The angle of the pendulum is logged over three seconds at the y-axis. This results into a diagram which produces a pattern. In computer animation, the principle is called spline animation. That means, on the x-axis the time is given, for example 0.5 seconds, 1.0 seconds, 1.5 seconds, while on the y-axis is the variable given. In most cases it's the position of an object. In the inverted pendulum it is the angle. The overall task consists of two parts: recording the angle over time and reproducing the recorded motion. The reproducing step is surprisingly simple. As an input we know, in which time step which angle the pendulum must have. Bringing the system to a state is called “Tracking control” or PID control. That means, we have a trajectory over time, and must find control actions to map the system to this trajectory. Other authors call the principle “steering behavior”, because the aim is to bring a robot to a given line.
Algorithm selection
Algorithm selection (sometimes also called per-instance algorithm selection or offline algorithm selection) is a meta-algorithmic technique to choose an algorithm from a portfolio on an instance-by-instance basis. It is motivated by the observation that on many practical problems, different algorithms have different performance characteristics. That is, while one algorithm performs well in some scenarios, it performs poorly in others and vice versa for another algorithm. If we can identify when to use which algorithm, we can optimize for each scenario and improve overall performance. This is what algorithm selection aims to do. The only prerequisite for applying algorithm selection techniques is that there exists (or that there can be constructed) a set of complementary algorithms. == Definition == Given a portfolio P {\displaystyle {\mathcal {P}}} of algorithms A ∈ P {\displaystyle {\mathcal {A}}\in {\mathcal {P}}} , a set of instances i ∈ I {\displaystyle i\in {\mathcal {I}}} and a cost metric m : P × I → R {\displaystyle m:{\mathcal {P}}\times {\mathcal {I}}\to \mathbb {R} } , the algorithm selection problem consists of finding a mapping s : I → P {\displaystyle s:{\mathcal {I}}\to {\mathcal {P}}} from instances I {\displaystyle {\mathcal {I}}} to algorithms P {\displaystyle {\mathcal {P}}} such that the cost ∑ i ∈ I m ( s ( i ) , i ) {\displaystyle \sum _{i\in {\mathcal {I}}}m(s(i),i)} across all instances is optimized. == Examples == === Boolean satisfiability problem (and other hard combinatorial problems) === A well-known application of algorithm selection is the Boolean satisfiability problem. Here, the portfolio of algorithms is a set of (complementary) SAT solvers, the instances are Boolean formulas, the cost metric is for example average runtime or number of unsolved instances. So, the goal is to select a well-performing SAT solver for each individual instance. In the same way, algorithm selection can be applied to many other N P {\displaystyle {\mathcal {NP}}} -hard problems (such as mixed integer programming, CSP, AI planning, TSP, MAXSAT, QBF and answer set programming). Competition-winning systems in SAT are SATzilla, 3S and CSHC === Machine learning === In machine learning, algorithm selection is better known as meta-learning. The portfolio of algorithms consists of machine learning algorithms (e.g., Random Forest, SVM, DNN), the instances are data sets and the cost metric is for example the error rate. So, the goal is to predict which machine learning algorithm will have a small error on each data set. == Instance features == The algorithm selection problem is mainly solved with machine learning techniques. By representing the problem instances by numerical features f {\displaystyle f} , algorithm selection can be seen as a multi-class classification problem by learning a mapping f i ↦ A {\displaystyle f_{i}\mapsto {\mathcal {A}}} for a given instance i {\displaystyle i} . Instance features are numerical representations of instances. For example, we can count the number of variables, clauses, average clause length for Boolean formulas, or number of samples, features, class balance for ML data sets to get an impression about their characteristics. === Static vs. probing features === We distinguish between two kinds of features: Static features are in most cases some counts and statistics (e.g., clauses-to-variables ratio in SAT). These features ranges from very cheap features (e.g. number of variables) to very complex features (e.g., statistics about variable-clause graphs). Probing features (sometimes also called landmarking features) are computed by running some analysis of algorithm behavior on an instance (e.g., accuracy of a cheap decision tree algorithm on an ML data set, or running for a short time a stochastic local search solver on a Boolean formula). These feature often cost more than simple static features. === Feature costs === Depending on the used performance metric m {\displaystyle m} , feature computation can be associated with costs. For example, if we use running time as performance metric, we include the time to compute our instance features into the performance of an algorithm selection system. SAT solving is a concrete example, where such feature costs cannot be neglected, since instance features for CNF formulas can be either very cheap (e.g., to get the number of variables can be done in constant time for CNFs in the DIMACs format) or very expensive (e.g., graph features which can cost tens or hundreds of seconds). It is important to take the overhead of feature computation into account in practice in such scenarios; otherwise a misleading impression of the performance of the algorithm selection approach is created. For example, if the decision which algorithm to choose can be made with perfect accuracy, but the features are the running time of the portfolio algorithms, there is no benefit to the portfolio approach. This would not be obvious if feature costs were omitted. == Approaches == === Regression approach === One of the first successful algorithm selection approaches predicted the performance of each algorithm m ^ A : I → R {\displaystyle {\hat {m}}_{\mathcal {A}}:{\mathcal {I}}\to \mathbb {R} } and selected the algorithm with the best predicted performance a r g min A ∈ P m ^ A ( i ) {\displaystyle arg\min _{{\mathcal {A}}\in {\mathcal {P}}}{\hat {m}}_{\mathcal {A}}(i)} for an instance i {\displaystyle i} . === Clustering approach === A common assumption is that the given set of instances I {\displaystyle {\mathcal {I}}} can be clustered into homogeneous subsets and for each of these subsets, there is one well-performing algorithm for all instances in there. So, the training consists of identifying the homogeneous clusters via an unsupervised clustering approach and associating an algorithm with each cluster. A new instance is assigned to a cluster and the associated algorithm selected. A more modern approach is cost-sensitive hierarchical clustering using supervised learning to identify the homogeneous instance subsets. === Pairwise cost-sensitive classification approach === A common approach for multi-class classification is to learn pairwise models between every pair of classes (here algorithms) and choose the class that was predicted most often by the pairwise models. We can weight the instances of the pairwise prediction problem by the performance difference between the two algorithms. This is motivated by the fact that we care most about getting predictions with large differences correct, but the penalty for an incorrect prediction is small if there is almost no performance difference. Therefore, each instance i {\displaystyle i} for training a classification model A 1 {\displaystyle {\mathcal {A}}_{1}} vs A 2 {\displaystyle {\mathcal {A}}_{2}} is associated with a cost | m ( A 1 , i ) − m ( A 2 , i ) | {\displaystyle |m({\mathcal {A}}_{1},i)-m({\mathcal {A}}_{2},i)|} . == Requirements == The algorithm selection problem can be effectively applied under the following assumptions: The portfolio P {\displaystyle {\mathcal {P}}} of algorithms is complementary with respect to the instance set I {\displaystyle {\mathcal {I}}} , i.e., there is no single algorithm A ∈ P {\displaystyle {\mathcal {A}}\in {\mathcal {P}}} that dominates the performance of all other algorithms over I {\displaystyle {\mathcal {I}}} (see figures to the right for examples on complementary analysis). In some application, the computation of instance features is associated with a cost. For example, if the cost metric is running time, we have also to consider the time to compute the instance features. In such cases, the cost to compute features should not be larger than the performance gain through algorithm selection. == Application domains == Algorithm selection is not limited to single domains but can be applied to any kind of algorithm if the above requirements are satisfied. Application domains include: hard combinatorial problems: SAT, Mixed Integer Programming, CSP, AI Planning, TSP, MAXSAT, QBF and Answer Set Programming combinatorial auctions in machine learning, the problem is known as meta-learning software design black-box optimization multi-agent systems numerical optimization linear algebra, differential equations evolutionary algorithms vehicle routing problem power systems For an extensive list of literature about algorithm selection, we refer to a literature overview. == Variants of algorithm selection == === Online selection === Online algorithm selection refers to switching between different algorithms during the solving process. This is useful as a hyper-heuristic. In contrast, offline algorithm selection selects an algorithm for a given instance only once and before the solving process. === Computation of schedules === An extension of algorithm selection is the per-instance algorithm scheduling problem, in which we do not select only one solver, but we select a time budget for each algorithm
Compute (machine learning)
In machine learning and deep learning, compute is the amount of computing power or computational resources required to train machine learning models and large language models. More broadly, compute is the computational power or resources necessary for a computer or computer program to function. == Definition == Compute is commonly defined as the amount of computing power or computational resources required to train machine learning and large language models. The term "compute" has also been more broadly applied to cloud computing, referencing processing power, memory, networking, storage, and other resources required for the computation of any program. Compute is measured in petaflop/s-days and is used to document AI training. A petaflop/s-day (pfs-day) consists of performing 1015 neural net operations per second for one day, or a total of about 1020 operations. The compute-time product serves as a mental convenience, similar to kilowatt-hour for energy. An amount of compute is meant to give an idea of the number of actual operations performed. == History == In a 2018 analysis titled "AI and compute", artificial intelligence company OpenAI introduced the concept of compute. OpenAI identified two eras of training AI systems in terms of compute-usage. From 1959 to 2012, compute roughly followed Moore’s law. Between 2012 and 2018, the amount of compute used in the largest AI training runs increased exponentially, growing by more than 300,000 times — roughly doubling every 3.4 months. By comparison, Moore’s Law doubled every two years over the same period. One of the largest models, released in 2020, used 600,000 times more computing power than the 2012 model. After 2020, compute growth began to slow down, with the compute needed for the largest AI models continuing to slow down in 2023. The notion of compute has become increasingly used from the mid-2020s onwards. == Compute growth and AI progress == Larger AI models trained on more data and using more computational resources, tend to perform better. This happens even if the algorithms themselves remain unchanged. As early as 2018, OpenAI noted the exponential increase in compute to be have a key role in AI progress. OpenAI considers three factors drive the advance of AI: algorithmic innovation, data, and the amount of compute available for training. AI models with more compute not only improve in the tasks they were trained on but can develop emergent abilities. Incremental improvements can lead to more abrupt leaps in capabilities. AI provider SpaceXAI said in 2026 that their AI progress is driven by compute and used it a key metric in the AI training of its supercomputer Colossus, the which contains 1 million GPUs. Anthropic has a contract of $1.25 billion per month with SpaceXAI to buy all the compute capacity at Colossus 1 data center. === Criticism and policy === Increasing, promoting or constraining progress in artificial intelligence has often be done via controlling the amount of compute. Policymarkers have enacted policies and provided support to make compute resources more accessible to domestic AI researchers. In a January 2022 report, the Center for Security and Emerging Technology (CSET) suggested to institutions that increasingly powerful and generalizable AI (AGI) will likely require other strategies than maximizing compute. Some AI researchers are also concerned that government might exclusively focus on scaling compute instead of other strategies. The CSET has reported on the various bottlenecks which could explain why deep learning needs for compute have slow down: training is expensive and training extremely large models generates traffic jams across many processors that are difficult to manage. there is a limited supply of AI chips (see AI chip memory shortage). CSET advances that the main resource is human capital, specifically talented researchers — according to a 2023 published survey of more than 400 AI researchers, academic and private sector workers. The survey found that AI researchers are not primarily or exclusively constrained by compute access. However, both academic and industry AI researchers equally report concerns that insufficient compute could prevent them from contributing meaningfully to AI research in the future. High compute users are more concerned about compute access. When asked about which resource provided by the government would be the most useful to them, some AI researchers select compute, other prefer grant funding. For this goal, CSET advised policymakers to ensure that even researchers with smaller budgets could effectively contribute to AI research. Other proposed strategies include using contemporary AI algorithms, managing modern AI infrastructure or focusing on interdisciplinary work between the AI field and other fields of computer science. A 2024 study on compute access found that academic-only AI research teams often have less compute intensive research topics, especially foundation models, compared to industry AI labs. As a consequence, academia is likely to play a smaller role in advancing such techniques. The researchers suggest nationally-sponsored computing infrastructure as well as open science initiatives to boost academic compute access. === Data === A 2022 study found that current large language models are significantly under-trained, a consequence of focusing on scaling language models whilst keeping the amount of training data constant. By training over 400 language models of various parameter and token size, they found that "for compute-optimal training", the model size and the number of training tokens should ideally be scaled equally: for every doubling of model size the number of training tokens should also be doubled.
International Speech Communication Association
The International Speech Communication Association (ISCA) is a non-profit organization and one of the two main professional associations for speech communication science and technology, the other association being the IEEE Signal Processing Society. == Purpose == The purpose of the International Speech Communication Association (ISCA) is to promote the study and application of automatic speech processing, including speech recognition and synthesis, as well as related areas such as speaker recognition and speech compression. The association's activities cover all aspects of speech processing, including computational, linguistic, and theoretical aspects. The primary goal of the International Speech Communication Association (ISCA) is to advance the field of automatic speech processing and communication technology through research, education, and collaboration. By promoting the study and application of speech technologies such as speech recognition, speech synthesis, speaker recognition, and speech compression, ISCA aims to foster innovation and development in the areas of human-computer interaction, telecommunications, and multimedia applications. ISCA serves as a platform for researchers, academics, industry professionals, and students to exchange knowledge, share best practices, and foster interdisciplinary dialogue in the field of speech communication science. Through conferences, workshops, publications, and educational initiatives, ISCA seeks to enhance the understanding of speech processing mechanisms, improve the accuracy and efficiency of speech technologies, and explore new frontiers in the realm of human language communication. Furthermore, ISCA plays a crucial role in promoting international collaboration and networking among professionals in the speech communication community. By facilitating partnerships and cooperation between individuals and organizations worldwide, ISCA seeks to drive global progress in speech technology research and application, ultimately contributing to the advancement of communication systems, accessibility tools, and interactive interfaces that benefit society as a whole. == Conferences == ISCA organizes yearly the Interspeech conference. Most recent Interspeech: 2013 Lyon, France 2014 Singapore 2015 Dresden, Germany 2016 San Francisco, US 2017 Stockholm, Sweden 2018 Hyderabad, India 2019 Graz, Austria 2020 Shanghai, China (fully virtual) 2021 Brno, Czechia (hybrid) 2022 Incheon, South Korea 2023 Dublin, Ireland 2023 Kos Island, Greece Forthcoming Interspeech: 2025 Rotterdam, the Netherlands == ISCA board == The ISCA president for 2023-2025 is Odette Scharenborg. The vice president is Bhuvana Ramabhadran and the other members are professionals in the field. == History of ISCA == The precursor to Interspeech was a conference called Eurospeech, first held in 1989 and organised by Jean-Pierre Tubach. It was the conference of the European Speech Communication Association (ESCA), itself the precursor of the International Speech Communication Association (ISCA). A year later another conference on speech science and technology was started: the International Conference on Spoken Language Processing (ICSLP), which was founded in 1990 by Hiroya Fujisaki. The first ISCA (vs. ESCA) event was the merging of Eurospeech and ICSLP to create ICSLP-Interspeech, held in Beijing, China in 2000. This was followed by Eurospeech-Interspeech, which was held in Aalborg, Denmark in 2001. In 2007, the Eurospeech and ICSLP parts of the conference names were dropped and Interspeech became the name of the yearly conference (first Interspeech location: Antwerp, Belgium).
Attention (machine learning)
In machine learning, attention is a method that determines the importance of each component in a sequence relative to the other components in that sequence. In natural language processing, importance is represented by "soft" weights assigned to each word in a sentence. More generally, attention encodes vectors called token embeddings across a fixed-width sequence that can range from tens to millions of tokens in size. Unlike "hard" weights, which are computed during the backwards training pass, "soft" weights exist only in the forward pass and therefore change with every step of the input. Earlier designs implemented the attention mechanism in a serial recurrent neural network (RNN) language translation system, but a more recent design, namely the transformer, removed the slower sequential RNN and relied more heavily on the faster parallel attention scheme. Inspired by ideas about attention in humans, the attention mechanism was developed to address the weaknesses of using information from the hidden layers of recurrent neural networks. Recurrent neural networks favor information contained in words at the end of a sentence and thus deemed more recent, thereby tending to attenuate the significance and associated predictive weight assigned to information earlier in the sentence. Attention allows a token equal access to any part of a sentence directly, rather than only through the previous state. == History == Additional surveys of the attention mechanism in deep learning are provided by Niu et al. and Soydaner. The major breakthrough came with self-attention, where each element in the input sequence attends to all others, enabling the model to capture global dependencies. This idea was central to the Transformer architecture, which replaced recurrence with attention mechanisms. As a result, Transformers became the foundation for models like BERT, T5 and generative pre-trained transformers (GPT). == Overview == The modern era of machine attention was revitalized by grafting an attention mechanism (Fig 1. orange) to an Encoder-Decoder. Figure 2 shows the internal step-by-step operation of the attention block (A) in Fig 1. === Interpreting attention weights === In translating between languages, alignment is the process of matching words from the source sentence to words of the translated sentence. Networks that perform verbatim translation without regard to word order would show the highest scores along the (dominant) diagonal of the matrix. The off-diagonal dominance shows that the attention mechanism is more nuanced. Consider an example of translating I love you to French. On the first pass through the decoder, 94% of the attention weight is on the first English word I, so the network offers the word je. On the second pass of the decoder, 88% of the attention weight is on the third English word you, so it offers t'. On the last pass, 95% of the attention weight is on the second English word love, so it offers aime. In the I love you example, the second word love is aligned with the third word aime. Stacking soft row vectors together for je, t', and aime yields an alignment matrix: Sometimes, alignment can be multiple-to-multiple. For example, the English phrase look it up corresponds to cherchez-le. Thus, "soft" attention weights work better than "hard" attention weights (setting one attention weight to 1, and the others to 0), as we would like the model to make a context vector consisting of a weighted sum of the hidden vectors, rather than "the best one", as there may not be a best hidden vector. == Variants == Many variants of attention implement soft weights, such as fast weight programmers, or fast weight controllers (1992). A "slow" neural network outputs the "fast" weights of another neural network through outer products. The slow network learns by gradient descent. It was later renamed as "linearized self-attention". Bahdanau-style attention, also referred to as additive attention, Luong-style attention, which is known as multiplicative attention, Early attention mechanisms similar to modern self-attention were proposed using recurrent neural networks. However, the highly parallelizable self-attention was introduced in 2017 and successfully used in the Transformer model, positional attention and factorized positional attention. For convolutional neural networks, attention mechanisms can be distinguished by the dimension on which they operate, namely: spatial attention, channel attention, or combinations. These variants recombine the encoder-side inputs to redistribute those effects to each target output. Often, a correlation-style matrix of dot products provides the re-weighting coefficients. In the figures below, W is the matrix of context attention weights, similar to the formula in Overview section above. == Optimizations == === Flash attention === The size of the attention matrix is proportional to the square of the number of input tokens. Therefore, when the input is long, calculating the attention matrix requires a lot of GPU memory. Flash attention is an implementation that reduces the memory needs and increases efficiency without sacrificing accuracy. It achieves this by partitioning the attention computation into smaller blocks that fit into the GPU's faster on-chip memory, reducing the need to store large intermediate matrices and thus lowering memory usage while increasing computational efficiency. === FlexAttention === FlexAttention is an attention kernel developed by Meta that allows users to modify attention scores prior to softmax and dynamically chooses the optimal attention algorithm. == Applications == Attention is widely used in natural language processing, computer vision, and speech recognition. In NLP, it improves context understanding in tasks like question answering and summarization. In vision, visual attention helps models focus on relevant image regions, enhancing object detection and image captioning. === Attention maps as explanations for vision transformers === From the original paper on vision transformers (ViT), visualizing attention scores as a heat map (called saliency maps or attention maps) has become an important and routine way to inspect the decision making process of ViT models. One can compute the attention maps with respect to any attention head at any layer, while the deeper layers tend to show more semantically meaningful visualization. Attention rollout is a recursive algorithm to combine attention scores across all layers, by computing the dot product of successive attention maps. Because vision transformers are typically trained in a self-supervised manner, attention maps are generally not class-sensitive. When a classification head is attached to the ViT backbone, class-discriminative attention maps (CDAM) combines attention maps and gradients with respect to the class [CLS] token. Some class-sensitive interpretability methods originally developed for convolutional neural networks can be also applied to ViT, such as GradCAM, which back-propagates the gradients to the outputs of the final attention layer. Using attention as basis of explanation for the transformers in language and vision is not without debate. While some pioneering papers analyzed and framed attention scores as explanations, higher attention scores do not always correlate with greater impact on model performances. == Mathematical representation == === Standard scaled dot-product attention === For matrices: Q ∈ R m × d k , K ∈ R n × d k {\displaystyle Q\in \mathbb {R} ^{m\times d_{k}},K\in \mathbb {R} ^{n\times d_{k}}} and V ∈ R n × d v {\displaystyle V\in \mathbb {R} ^{n\times d_{v}}} , the scaled dot-product, or QKV attention, is defined as: Attention ( Q , K , V ) = softmax ( Q K T d k ) V ∈ R m × d v {\displaystyle {\text{Attention}}(Q,K,V)={\text{softmax}}\left({\frac {QK^{T}}{\sqrt {d_{k}}}}\right)V\in \mathbb {R} ^{m\times d_{v}}} where T {\displaystyle {}^{T}} denotes transpose and the softmax function is applied independently to every row of its argument. The matrix Q {\displaystyle Q} contains m {\displaystyle m} queries, while matrices K , V {\displaystyle K,V} jointly contain an unordered set of n {\displaystyle n} key-value pairs. Value vectors in matrix V {\displaystyle V} are weighted using the weights resulting from the softmax operation, so that the rows of the m {\displaystyle m} -by- d v {\displaystyle d_{v}} output matrix are confined to the convex hull of the points in R d v {\displaystyle \mathbb {R} ^{d_{v}}} given by the rows of V {\displaystyle V} . To understand the permutation invariance and permutation equivariance properties of QKV attention, let A ∈ R m × m {\displaystyle A\in \mathbb {R} ^{m\times m}} and B ∈ R n × n {\displaystyle B\in \mathbb {R} ^{n\times n}} be permutation matrices; and D ∈ R m × n {\displaystyle D\in \mathbb {R} ^{m\times n}} an arbitrary matrix. The softmax function is permutation equivariant in the sense that: softmax ( A D B ) = A softmax ( D ) B {\displays