Variable-order Bayesian network (VOBN) models provide an important extension of both the Bayesian network models and the variable-order Markov models. VOBN models are used in machine learning in general and have shown great potential in bioinformatics applications. These models extend the widely used position weight matrix (PWM) models, Markov models, and Bayesian network (BN) models. In contrast to the BN models, where each random variable depends on a fixed subset of random variables, in VOBN models these subsets may vary based on the specific realization of observed variables. The observed realizations are often called the context and, hence, VOBN models are also known as context-specific Bayesian networks. The flexibility in the definition of conditioning subsets of variables turns out to be a real advantage in classification and analysis applications, as the statistical dependencies between random variables in a sequence of variables (not necessarily adjacent) may be taken into account efficiently, and in a position-specific and context-specific manner.
Tesla Dojo
Tesla Dojo is a series of supercomputers designed and built by Tesla for computer vision video processing and recognition. It was used for training Tesla's machine learning models to improve its Full Self-Driving (FSD) advanced driver-assistance system. It went into production in July 2023. Dojo's goal was to efficiently process millions of terabytes of video data captured from real-life driving situations from Tesla's 4+ million cars. This goal led to a considerably different architecture than conventional supercomputer designs. In August 2025, Bloomberg News reported that the Dojo project had been disbanded, though it was restarted in January 2026. == History == Tesla operates several massively parallel computing clusters for developing its Autopilot advanced driver assistance system. Its primary unnamed cluster using 5,760 Nvidia A100 graphics processing units (GPUs) was touted by Andrej Karpathy in 2021 at the fourth International Joint Conference on Computer Vision and Pattern Recognition (CCVPR 2021) to be "roughly the number five supercomputer in the world" at approximately 81.6 petaflops, based on scaling the performance of the Nvidia Selene supercomputer, which uses similar components. However, the performance of the primary Tesla GPU cluster has been disputed, as it was not clear if this was measured using single-precision or double-precision floating point numbers (FP32 or FP64). Tesla also operates a second 4,032 GPU cluster for training and a third 1,752 GPU cluster for automatic labeling of objects. The primary unnamed Tesla GPU cluster has been used for processing one million video clips, each ten seconds long, taken from Tesla Autopilot cameras operating in Tesla cars in the real world, running at 36 frames per second. Collectively, these video clips contained six billion object labels, with depth and velocity data; the total size of the data set was 1.5 petabytes. This data set was used for training a neural network intended to help Autopilot computers in Tesla cars understand roads. By August 2022, Tesla had upgraded the primary GPU cluster to 7,360 GPUs. Dojo was first mentioned by Elon Musk in April 2019 during Tesla's "Autonomy Investor Day". In August 2020, Musk stated it was "about a year away" due to power and thermal issues. Dojo was officially announced at Tesla's Artificial Intelligence (AI) Day on August 19, 2021. Tesla revealed details of the D1 chip and its plans for "Project Dojo", a datacenter that would house 3,000 D1 chips; the first "Training Tile" had been completed and delivered the week before. In October 2021, Tesla released a "Dojo Technology" whitepaper describing the Configurable Float8 (CFloat8) and Configurable Float16 (CFloat16) floating point formats and arithmetic operations as an extension of Institute of Electrical and Electronics Engineers (IEEE) standard 754. At the follow-up AI Day in September 2022, Tesla announced it had built several System Trays and one Cabinet. During a test, the company stated that Project Dojo drew 2.3 megawatts (MW) of power before tripping a local San Jose, California power substation. At the time, Tesla was assembling one Training Tile per day. In August 2023, Tesla powered on Dojo for production use as well as a new training cluster configured with 10,000 Nvidia H100 GPUs. In January 2024, Musk described Dojo as "a long shot worth taking because the payoff is potentially very high. But it's not something that is a high probability." In June 2024, Musk explained that ongoing construction work at Gigafactory Texas is for a computing cluster claiming that it is planned to comprise an even mix of "Tesla AI" and Nvidia/other hardware with a total thermal design power of at first 130 MW and eventually exceeding 500 MW. In August 2025, Bloomberg News reported that the Dojo project was disbanded, though Musk announced it would be restarted in January 2026 with a new chip iteration. == Technical architecture == The fundamental unit of the Dojo supercomputer is the D1 chip, designed by a team at Tesla led by ex-AMD CPU designer Ganesh Venkataramanan, including Emil Talpes, Debjit Das Sarma, Douglas Williams, Bill Chang, and Rajiv Kurian. The D1 chip is manufactured by the Taiwan Semiconductor Manufacturing Company (TSMC) using 7 nanometer (nm) semiconductor nodes, has 50 billion transistors and a large die size of 645 mm2 (1.0 square inch). Updating at Artificial Intelligence (AI) Day in 2022, Tesla announced that Dojo would scale by deploying multiple ExaPODs, in which there would be: 10 Cabinets per ExaPOD (1,062,000 cores, 3,000 D1 chips) 2 System Trays per Cabinet (106,200 cores, 300 D1 chips) 6 Training Tiles per System Tray (53,100 cores, along with host interface hardware) 25 D1 chips per Training Tile (8,850 cores) 354 computing cores per D1 chip According to Venkataramanan, Tesla's senior director of Autopilot hardware, Dojo will have more than an exaflop (a million teraflops) of computing power. For comparison, according to Nvidia, in August 2021, the (pre-Dojo) Tesla AI-training center used 720 nodes, each with eight Nvidia A100 Tensor Core GPUs for 5,760 GPUs in total, providing up to 1.8 exaflops of performance. === D1 chip === Each node (computing core) of the D1 processing chip is a general purpose 64-bit CPU with a superscalar core. It supports internal instruction-level parallelism, and includes simultaneous multithreading (SMT). It doesn't support virtual memory and uses limited memory protection mechanisms. Dojo software/applications manage chip resources. The D1 instruction set supports both 64-bit scalar and 64-byte single instruction, multiple data (SIMD) vector instructions. The integer unit mixes reduced instruction set computer (RISC-V) and custom instructions, supporting 8, 16, 32, or 64 bit integers. The custom vector math unit is optimized for machine learning kernels and supports multiple data formats, with a mix of precisions and numerical ranges, many of which are compiler composable. Up to 16 vector formats can be used simultaneously. ==== Node ==== Each D1 node uses a 32-byte fetch window holding up to eight instructions. These instructions are fed to an eight-wide decoder which supports two threads per cycle, followed by a four-wide, four-way SMT scalar scheduler that has two integer units, two address units, and one register file per thread. Vector instructions are passed further down the pipeline to a dedicated vector scheduler with two-way SMT, which feeds either a 64-byte SIMD unit or four 8×8×4 matrix multiplication units. The network on-chip (NOC) router links cores into a two-dimensional mesh network. It can send one packet in and one packet out in all four directions to/from each neighbor node, along with one 64-byte read and one 64-byte write to local SRAM per clock cycle. Hardware native operations transfer data, semaphores and barrier constraints across memories and CPUs. System-wide double data rate 4 (DDR4) synchronous dynamic random-access memory (SDRAM) memory works like bulk storage. ==== Memory ==== Each core has a 1.25 megabytes (MB) of SRAM main memory. Load and store speeds reach 400 gigabytes (GB) per second and 270 GB/sec, respectively. The chip has explicit core-to-core data transfer instructions. Each SRAM has a unique list parser that feeds a pair of decoders and a gather engine that feeds the vector register file, which together can directly transfer information across nodes. ==== Die ==== Twelve nodes (cores) are grouped into a local block. Nodes are arranged in an 18×20 array on a single die, of which 354 cores are available for applications. The die runs at 2 gigahertz (GHz) and totals 440 MB of SRAM (360 cores × 1.25 MB/core). It reaches 376 teraflops using 16-bit brain floating point (BF16) numbers or using configurable 8-bit floating point (CFloat8) numbers, which is a Tesla proposal, and 22 teraflops at FP32. Each die comprises 576 bi-directional serializer/deserializer (SerDes) channels along the perimeter to link to other dies, and moves 8 TB/sec across all four die edges. Each D1 chip has a thermal design power of approximately 400 watts. === Training Tile === The water-cooled Training Tile packages 25 D1 chips into a 5×5 array. Each tile supports 36 TB/sec of aggregate bandwidth via 40 input/output (I/O) chips - half the bandwidth of the chip mesh network. Each tile supports 10 TB/sec of on-tile bandwidth. Each tile has 11 GB of SRAM memory (25 D1 chips × 360 cores/D1 × 1.25 MB/core). Each tile achieves 9 petaflops at BF16/CFloat8 precision (25 D1 chips × 376 TFLOP/D1). Each tile consumes 15 kilowatts; 288 amperes at 52 volts. === System Tray === Six tiles are aggregated into a System Tray, which is integrated with a host interface. Each host interface includes 512 x86 cores, providing a Linux-based user environment. Previously, the Dojo System Tray was known as the Training Matrix, which includes six Training Tiles, 20 Dojo Interface Processor cards across four host servers, and Ethernet-l
STIT logic
STIT logic (from seeing to it that) is a family of modal and branching-time logics for reasoning about agency and choice. A typical STIT operator has the form [ i s t i t : φ ] {\displaystyle [i\ {\mathsf {stit}}:\varphi ]} , usually read as "agent i {\displaystyle i} sees to it that φ {\displaystyle \varphi } ", and is interpreted in models where agents choose between alternative possible futures. STIT logics are used in action theory, deontic logic, epistemic logic, and the theory of intelligent agents to formalise notions such as "could have done otherwise", responsibility, joint action, and strategic ability in an indeterministic world. == Etymology == The acronym STIT comes from the English phrase "seeing to it that", introduced in influential work by Nuel Belnap and Michael Perloff on the logical analysis of agentive expressions. In this tradition, "to see to it that φ {\displaystyle \varphi } " is treated as a primitive agency operator, rather than being reduced to ordinary modal necessity. == History == Modern STIT logic arose in the 1980s in the context of branching-time semantics and formal theories of agency. Belnap and Perloff's article "Seeing to it that: A canonical form for agentives" introduced the idea of treating expressions of the form "agent i sees to it that φ" as a primitive modal operator, and analysed such sentences using a branching tree of moments and histories. This approach was further developed in a series of papers on indeterminism and agency and provided the conceptual core for later STIT formalisms. In the 1990s the basic formal systems of STIT logic were worked out. Horty and Belnap's influential paper on the deliberative STIT operator distinguished between a "Chellas" STIT that merely records the result of an agent's present choice and a "deliberative" STIT that requires the agent's choice to make a difference, and connected STIT with issues of action, omission, ability and obligation. Around the same time, Ming Xu proved completeness and decidability results for basic STIT systems, including a single-agent logic with Kripke-style semantics and axiomatizations for multi-agent deliberative STIT, thereby establishing STIT as a well-behaved normal modal framework. This early work was systematised in Belnap, Perloff and Xu's monograph Facing the Future: Agents and Choices in Our Indeterminist World, which presents a general branching-time semantics for individual and group STIT operators, discusses independence-of-agents conditions and articulates the metaphysical picture of an indeterministic "tree" of moments. At roughly the same time, Horty's book Agency and Deontic Logic developed deontic STIT logics in which obligations are tied to agents' available choices rather than to static states of affairs, and used the resulting systems to analyse "ought implies can", contrary-to-duty obligations and deontic paradoxes. These works helped to position STIT at the intersection of action theory, temporal logic and deontic logic. From the late 1990s and 2000s onward, STIT logics were combined with epistemic, temporal and strategic modalities. Broersen introduced complete STIT logics for knowledge and action and deontic-epistemic STIT systems that distinguish different modes of mens rea, with applications to responsibility and the specification of multi-agent systems. Work on group and coalitional agency investigated axiomatisations and complexity results for group STIT logics, and related STIT-based analyses of agency to coalition logic and alternating-time temporal logic (ATL) by exhibiting formal embeddings between the frameworks. Explicit temporal operators were added to STIT in so-called temporal STIT logics. Lorini proposed a temporal STIT with "next" and "until" operators along histories and showed how it can be applied to normative reasoning about ongoing behaviour and commitments. Ciuni and Lorini compared different semantics for temporal STIT, clarifying the relationships between branching-time, game-based and epistemic approaches, while Boudou and Lorini gave a semantics for temporal STIT based on concurrent game structures, thus strengthening links with standard models of multi-agent interaction used for ATL and strategy logic. In parallel, complexity-theoretic work by Balbiani, Herzig and Troquard and by Schwarzentruber and co-authors investigated the satisfiability and model-checking problems for various STIT fragments, showing for instance that many expressive group STIT logics are undecidable or of high computational complexity. In the 2010s, STIT ideas were combined with justification logic, imagination operators and refined deontic notions. Justification STIT logics, developed by Olkhovikov and others, merge explicit justifications with STIT-style agency so that producing a proof can itself be treated as an action that brings about knowledge, and they come with completeness and decidability results. Olkhovikov and Wansing introduced STIT imagination logics, together with axiomatic systems and tableau calculi, to model acts of voluntary imagining and their role in doxastic control. Other authors have proposed STIT-based logics of responsibility, blameworthiness and intentionality for use in philosophical and AI settings. Xu's survey article "Combinations of STIT with Ought and Know" (2015) reviews many of these developments and emphasises the interplay between deontic and epistemic STIT logics. Current research on STIT focuses on proof theory, automated reasoning and richer expressive resources. Lyon and van Berkel, building on earlier work on labelled calculi for STIT, have developed cut-free sequent systems and proof-search algorithms that yield syntactic decision procedures for a range of deontic and non-deontic multi-agent STIT logics and support applications such as duty checking and compliance checking in autonomous systems. Sawasaki has proposed first-order cstit-based STIT logics that can distinguish de re and de dicto readings of agency statements and has proved strong completeness results for Hilbert systems over finite models, moving the STIT programme beyond the purely propositional level. Further work investigates interpreted-system and computationally grounded semantics for STIT and its extensions in order to model the behaviour of autonomous agents in multi-agent settings, and proposes STIT-based semantics for epistemic notions based on patterns of information disclosure in interactive systems. == Branching-time semantics == STIT logics are usually interpreted over branching-time models. A standard STIT frame consists of: a non-empty set of moments T {\displaystyle T} , partially ordered by < {\displaystyle <} so that ( T , < ) {\displaystyle (T,<)} forms a tree (every pair of moments with a common predecessor has a greatest lower bound); a set of histories, each history being a maximal linearly ordered subset of T {\displaystyle T} ; a non-empty set of agents A g {\displaystyle Ag} ; for each agent i ∈ A g {\displaystyle i\in Ag} and moment m {\displaystyle m} , a choice function c h o i c e i m {\displaystyle {\mathsf {choice}}_{i}^{m}} that partitions the set of histories passing through m {\displaystyle m} into choice cells. The idea is that a moment represents a time at which choices are made, and histories represent complete possible future courses of events. At each moment, each agent's choice corresponds to selecting one of the available cells of histories determined by their choice function. Formulas are evaluated at pairs ( m , h ) {\displaystyle (m,h)} of a moment and a history through that moment (sometimes written m / h {\displaystyle m/h} ). A valuation assigns truth-values to atomic propositions at such indices; Boolean connectives are interpreted pointwise as in Kripke-style modal logic. == Chellas and deliberative STIT operators == Several STIT operators have been distinguished in the literature. A common approach uses two closely related operators, often called Chellas STIT and deliberative STIT. Let H m {\displaystyle H_{m}} be the set of histories passing through a moment m {\displaystyle m} , and write H m {\displaystyle H_{m}} ⟦ φ ⟧ m = { h ∈ H m ∣ M , m / h ⊨ φ } {\displaystyle {\text{⟦}}\varphi {\text{⟧}}_{m}=\{h\in H_{m}\mid M,m/h\models \varphi \}} for the set of histories at m {\displaystyle m} where φ {\displaystyle \varphi } holds. The Chellas STIT operator, often written [ i c s t i t : φ ] {\displaystyle [i\ {\mathsf {cstit}}:\varphi ]} , is given by M , m / h ⊨ [ i c s t i t : φ ] iff c h o i c e i m ( h ) ⊆ ⟦ φ ⟧ m . {\displaystyle M,m/h\models [i\ {\mathsf {cstit}}:\varphi ]\quad {\text{iff}}\quad {\mathsf {choice}}_{i}^{m}(h)\subseteq {\text{⟦}}\varphi {\text{⟧}}_{m}.} Intuitively, agent i {\displaystyle i} sees to it that φ {\displaystyle \varphi } if φ {\displaystyle \varphi } holds at all histories compatible with their present choice. The deliberative STIT operator, [ i d s t i t : φ ] {\displaystyle [i\ {\mathsf {dstit}}:\varphi ]} , adds
Kernel embedding of distributions
In machine learning, the kernel embedding of distributions (also called the kernel mean or mean map) comprises a class of nonparametric methods in which a probability distribution is represented as an element of a reproducing kernel Hilbert space (RKHS). A generalization of the individual data-point feature mapping done in classical kernel methods, the embedding of distributions into infinite-dimensional feature spaces can preserve all of the statistical features of arbitrary distributions, while allowing one to compare and manipulate distributions using Hilbert space operations such as inner products, distances, projections, linear transformations, and spectral analysis. This learning framework is very general and can be applied to distributions over any space Ω {\displaystyle \Omega } on which a sensible kernel function (measuring similarity between elements of Ω {\displaystyle \Omega } ) may be defined. For example, various kernels have been proposed for learning from data which are: vectors in R d {\displaystyle \mathbb {R} ^{d}} , discrete classes/categories, strings, graphs/networks, images, time series, manifolds, dynamical systems, and other structured objects. The theory behind kernel embeddings of distributions has been primarily developed by Alex Smola, Le Song, Arthur Gretton, and Bernhard Schölkopf. A review of recent works on kernel embedding of distributions can be found in. The analysis of distributions is fundamental in machine learning and statistics, and many algorithms in these fields rely on information theoretic approaches such as entropy, mutual information, or Kullback–Leibler divergence. However, to estimate these quantities, one must first either perform density estimation, or employ sophisticated space-partitioning/bias-correction strategies which are typically infeasible for high-dimensional data. Commonly, methods for modeling complex distributions rely on parametric assumptions that may be unfounded or computationally challenging (e.g. Gaussian mixture models), while nonparametric methods like kernel density estimation (Note: the smoothing kernels in this context have a different interpretation than the kernels discussed here) or characteristic function representation (via the Fourier transform of the distribution) break down in high-dimensional settings. Methods based on the kernel embedding of distributions sidestep these problems and also possess the following advantages: Data may be modeled without restrictive assumptions about the form of the distributions and relationships between variables Intermediate density estimation is not needed Practitioners may specify the properties of a distribution most relevant for their problem (incorporating prior knowledge via choice of the kernel) If a characteristic kernel is used, then the embedding can uniquely preserve all information about a distribution, while thanks to the kernel trick, computations on the potentially infinite-dimensional RKHS can be implemented in practice as simple Gram matrix operations Dimensionality-independent rates of convergence for the empirical kernel mean (estimated using samples from the distribution) to the kernel embedding of the true underlying distribution can be proven. Learning algorithms based on this framework exhibit good generalization ability and finite sample convergence, while often being simpler and more effective than information theoretic methods Thus, learning via the kernel embedding of distributions offers a principled drop-in replacement for information theoretic approaches and is a framework which not only subsumes many popular methods in machine learning and statistics as special cases, but also can lead to entirely new learning algorithms. == Definitions == Let X {\displaystyle X} denote a random variable with domain Ω {\displaystyle \Omega } and distribution P {\displaystyle P} . Given a symmetric, positive-definite kernel k : Ω × Ω → R {\displaystyle k:\Omega \times \Omega \rightarrow \mathbb {R} } the Moore–Aronszajn theorem asserts the existence of a unique RKHS H {\displaystyle {\mathcal {H}}} on Ω {\displaystyle \Omega } (a Hilbert space of functions f : Ω → R {\displaystyle f:\Omega \to \mathbb {R} } equipped with an inner product ⟨ ⋅ , ⋅ ⟩ H {\displaystyle \langle \cdot ,\cdot \rangle _{\mathcal {H}}} and a norm ‖ ⋅ ‖ H {\displaystyle \|\cdot \|_{\mathcal {H}}} ) for which k {\displaystyle k} is a reproducing kernel, i.e., in which the element k ( x , ⋅ ) {\displaystyle k(x,\cdot )} satisfies the reproducing property ⟨ f , k ( x , ⋅ ) ⟩ H = f ( x ) ∀ f ∈ H , ∀ x ∈ Ω . {\displaystyle \langle f,k(x,\cdot )\rangle _{\mathcal {H}}=f(x)\qquad \forall f\in {\mathcal {H}},\quad \forall x\in \Omega .} One may alternatively consider x ↦ k ( x , ⋅ ) {\displaystyle x\mapsto k(x,\cdot )} as an implicit feature mapping φ : Ω → H {\displaystyle \varphi :\Omega \rightarrow {\mathcal {H}}} (which is therefore also called the feature space), so that k ( x , x ′ ) = ⟨ φ ( x ) , φ ( x ′ ) ⟩ H {\displaystyle k(x,x')=\langle \varphi (x),\varphi (x')\rangle _{\mathcal {H}}} can be viewed as a measure of similarity between points x , x ′ ∈ Ω . {\displaystyle x,x'\in \Omega .} While the similarity measure is linear in the feature space, it may be highly nonlinear in the original space depending on the choice of kernel. === Kernel embedding === The kernel embedding of the distribution P {\displaystyle P} in H {\displaystyle {\mathcal {H}}} (also called the kernel mean or mean map) is given by: μ X := E [ k ( X , ⋅ ) ] = E [ φ ( X ) ] = ∫ Ω φ ( x ) d P ( x ) {\displaystyle \mu _{X}:=\mathbb {E} [k(X,\cdot )]=\mathbb {E} [\varphi (X)]=\int _{\Omega }\varphi (x)\ \mathrm {d} P(x)} If P {\displaystyle P} allows a square integrable density p {\displaystyle p} , then μ X = E k p {\displaystyle \mu _{X}={\mathcal {E}}_{k}p} , where E k {\displaystyle {\mathcal {E}}_{k}} is the Hilbert–Schmidt integral operator. A kernel is characteristic if the mean embedding μ : { family of distributions over Ω } → H {\displaystyle \mu :\{{\text{family of distributions over }}\Omega \}\to {\mathcal {H}}} is injective. Each distribution can thus be uniquely represented in the RKHS and all statistical features of distributions are preserved by the kernel embedding if a characteristic kernel is used. === Empirical kernel embedding === Given n {\displaystyle n} training examples { x 1 , … , x n } {\displaystyle \{x_{1},\ldots ,x_{n}\}} drawn independently and identically distributed (i.i.d.) from P , {\displaystyle P,} the kernel embedding of P {\displaystyle P} can be empirically estimated as μ ^ X = 1 n ∑ i = 1 n φ ( x i ) {\displaystyle {\widehat {\mu }}_{X}={\frac {1}{n}}\sum _{i=1}^{n}\varphi (x_{i})} === Joint distribution embedding === If Y {\displaystyle Y} denotes another random variable (for simplicity, assume the co-domain of Y {\displaystyle Y} is also Ω {\displaystyle \Omega } with the same kernel k {\displaystyle k} which satisfies ⟨ φ ( x ) ⊗ φ ( y ) , φ ( x ′ ) ⊗ φ ( y ′ ) ⟩ = k ( x , x ′ ) k ( y , y ′ ) {\displaystyle \langle \varphi (x)\otimes \varphi (y),\varphi (x')\otimes \varphi (y')\rangle =k(x,x')k(y,y')} ), then the joint distribution P ( x , y ) ) {\displaystyle P(x,y))} can be mapped into a tensor product feature space H ⊗ H {\displaystyle {\mathcal {H}}\otimes {\mathcal {H}}} via C X Y = E [ φ ( X ) ⊗ φ ( Y ) ] = ∫ Ω × Ω φ ( x ) ⊗ φ ( y ) d P ( x , y ) {\displaystyle {\mathcal {C}}_{XY}=\mathbb {E} [\varphi (X)\otimes \varphi (Y)]=\int _{\Omega \times \Omega }\varphi (x)\otimes \varphi (y)\ \mathrm {d} P(x,y)} By the equivalence between a tensor and a linear map, this joint embedding may be interpreted as an uncentered cross-covariance operator C X Y : H → H {\displaystyle {\mathcal {C}}_{XY}:{\mathcal {H}}\to {\mathcal {H}}} from which the cross-covariance of functions f , g ∈ H {\displaystyle f,g\in {\mathcal {H}}} can be computed as Cov ( f ( X ) , g ( Y ) ) := E [ f ( X ) g ( Y ) ] − E [ f ( X ) ] E [ g ( Y ) ] = ⟨ f , C X Y g ⟩ H = ⟨ f ⊗ g , C X Y ⟩ H ⊗ H {\displaystyle \operatorname {Cov} (f(X),g(Y)):=\mathbb {E} [f(X)g(Y)]-\mathbb {E} [f(X)]\mathbb {E} [g(Y)]=\langle f,{\mathcal {C}}_{XY}g\rangle _{\mathcal {H}}=\langle f\otimes g,{\mathcal {C}}_{XY}\rangle _{{\mathcal {H}}\otimes {\mathcal {H}}}} Given n {\displaystyle n} pairs of training examples { ( x 1 , y 1 ) , … , ( x n , y n ) } {\displaystyle \{(x_{1},y_{1}),\dots ,(x_{n},y_{n})\}} drawn i.i.d. from P {\displaystyle P} , we can also empirically estimate the joint distribution kernel embedding via C ^ X Y = 1 n ∑ i = 1 n φ ( x i ) ⊗ φ ( y i ) {\displaystyle {\widehat {\mathcal {C}}}_{XY}={\frac {1}{n}}\sum _{i=1}^{n}\varphi (x_{i})\otimes \varphi (y_{i})} === Conditional distribution embedding === Given a conditional distribution P ( y ∣ x ) , {\displaystyle P(y\mid x),} one can define the corresponding RKHS embedding as μ Y ∣ x = E [ φ ( Y ) ∣ X ] = ∫ Ω φ ( y ) d P ( y ∣ x ) {\displaystyle \mu _{Y\mid x}=\mathbb {E} [\varphi (Y)\mid X]=\int _{\Omega
Matrix regularization
In the field of statistical learning theory, matrix regularization generalizes notions of vector regularization to cases where the object to be learned is a matrix. The purpose of regularization is to enforce conditions, for example sparsity or smoothness, that can produce stable predictive functions. For example, in the more common vector framework, Tikhonov regularization optimizes over min x ‖ A x − y ‖ 2 + λ ‖ x ‖ 2 {\displaystyle \min _{x}\left\|Ax-y\right\|^{2}+\lambda \left\|x\right\|^{2}} to find a vector x {\displaystyle x} that is a stable solution to the regression problem. When the system is described by a matrix rather than a vector, this problem can be written as min X ‖ A X − Y ‖ 2 + λ ‖ X ‖ 2 , {\displaystyle \min _{X}\left\|AX-Y\right\|^{2}+\lambda \left\|X\right\|^{2},} where the vector norm enforcing a regularization penalty on x {\displaystyle x} has been extended to a matrix norm on X {\displaystyle X} . Matrix regularization has applications in matrix completion, multivariate regression, and multi-task learning. Ideas of feature and group selection can also be extended to matrices, and these can be generalized to the nonparametric case of multiple kernel learning. == Basic definition == Consider a matrix W {\displaystyle W} to be learned from a set of examples, S = ( X i t , y i t ) {\displaystyle S=(X_{i}^{t},y_{i}^{t})} , where i {\displaystyle i} goes from 1 {\displaystyle 1} to n {\displaystyle n} , and t {\displaystyle t} goes from 1 {\displaystyle 1} to T {\displaystyle T} . Let each input matrix X i {\displaystyle X_{i}} be ∈ R D T {\displaystyle \in \mathbb {R} ^{DT}} , and let W {\displaystyle W} be of size D × T {\displaystyle D\times T} . A general model for the output y {\displaystyle y} can be posed as y i t = ⟨ W , X i t ⟩ F , {\displaystyle y_{i}^{t}=\left\langle W,X_{i}^{t}\right\rangle _{F},} where the inner product is the Frobenius inner product. For different applications the matrices X i {\displaystyle X_{i}} will have different forms, but for each of these the optimization problem to infer W {\displaystyle W} can be written as min W ∈ H E ( W ) + R ( W ) , {\displaystyle \min _{W\in {\mathcal {H}}}E(W)+R(W),} where E {\displaystyle E} defines the empirical error for a given W {\displaystyle W} , and R ( W ) {\displaystyle R(W)} is a matrix regularization penalty. The function R ( W ) {\displaystyle R(W)} is typically chosen to be convex and is often selected to enforce sparsity (using ℓ 1 {\displaystyle \ell ^{1}} -norms) and/or smoothness (using ℓ 2 {\displaystyle \ell ^{2}} -norms). Finally, W {\displaystyle W} is in the space of matrices H {\displaystyle {\mathcal {H}}} with Frobenius inner product ⟨ … ⟩ F {\displaystyle \langle \dots \rangle _{F}} . == General applications == === Matrix completion === In the problem of matrix completion, the matrix X i t {\displaystyle X_{i}^{t}} takes the form X i t = e t ⊗ e i ′ , {\displaystyle X_{i}^{t}=e_{t}\otimes e_{i}',} where ( e t ) t {\displaystyle (e_{t})_{t}} and ( e i ′ ) i {\displaystyle (e_{i}')_{i}} are the canonical basis in R T {\displaystyle \mathbb {R} ^{T}} and R D {\displaystyle \mathbb {R} ^{D}} . In this case the role of the Frobenius inner product is to select individual elements w i t {\displaystyle w_{i}^{t}} from the matrix W {\displaystyle W} . Thus, the output y {\displaystyle y} is a sampling of entries from the matrix W {\displaystyle W} . The problem of reconstructing W {\displaystyle W} from a small set of sampled entries is possible only under certain restrictions on the matrix, and these restrictions can be enforced by a regularization function. For example, it might be assumed that W {\displaystyle W} is low-rank, in which case the regularization penalty can take the form of a nuclear norm. R ( W ) = λ ‖ W ‖ ∗ = λ ∑ i | σ i | , {\displaystyle R(W)=\lambda \left\|W\right\|_{}=\lambda \sum _{i}\left|\sigma _{i}\right|,} where σ i {\displaystyle \sigma _{i}} , with i {\displaystyle i} from 1 {\displaystyle 1} to min D , T {\displaystyle \min D,T} , are the singular values of W {\displaystyle W} . === Multivariate regression === Models used in multivariate regression are parameterized by a matrix of coefficients. In the Frobenius inner product above, each matrix X {\displaystyle X} is X i t = e t ⊗ x i {\displaystyle X_{i}^{t}=e_{t}\otimes x_{i}} such that the output of the inner product is the dot product of one row of the input with one column of the coefficient matrix. The familiar form of such models is Y = X W + b {\displaystyle Y=XW+b} Many of the vector norms used in single variable regression can be extended to the multivariate case. One example is the squared Frobenius norm, which can be viewed as an ℓ 2 {\displaystyle \ell ^{2}} -norm acting either entrywise, or on the singular values of the matrix: R ( W ) = λ ‖ W ‖ F 2 = λ ∑ i ∑ j | w i j | 2 = λ Tr ( W ∗ W ) = λ ∑ i σ i 2 . {\displaystyle R(W)=\lambda \left\|W\right\|_{F}^{2}=\lambda \sum _{i}\sum _{j}\left|w_{ij}\right|^{2}=\lambda \operatorname {Tr} \left(W^{}W\right)=\lambda \sum _{i}\sigma _{i}^{2}.} In the multivariate case the effect of regularizing with the Frobenius norm is the same as the vector case; very complex models will have larger norms, and, thus, will be penalized more. === Multi-task learning === The setup for multi-task learning is almost the same as the setup for multivariate regression. The primary difference is that the input variables are also indexed by task (columns of Y {\displaystyle Y} ). The representation with the Frobenius inner product is then X i t = e t ⊗ x i t . {\displaystyle X_{i}^{t}=e_{t}\otimes x_{i}^{t}.} The role of matrix regularization in this setting can be the same as in multivariate regression, but matrix norms can also be used to couple learning problems across tasks. In particular, note that for the optimization problem min W ‖ X W − Y ‖ 2 2 + λ ‖ W ‖ 2 2 {\displaystyle \min _{W}\left\|XW-Y\right\|_{2}^{2}+\lambda \left\|W\right\|_{2}^{2}} the solutions corresponding to each column of Y {\displaystyle Y} are decoupled. That is, the same solution can be found by solving the joint problem, or by solving an isolated regression problem for each column. The problems can be coupled by adding an additional regularization penalty on the covariance of solutions min W , Ω ‖ X W − Y ‖ 2 2 + λ 1 ‖ W ‖ 2 2 + λ 2 Tr ( W T Ω − 1 W ) {\displaystyle \min _{W,\Omega }\left\|XW-Y\right\|_{2}^{2}+\lambda _{1}\left\|W\right\|_{2}^{2}+\lambda _{2}\operatorname {Tr} \left(W^{T}\Omega ^{-1}W\right)} where Ω {\displaystyle \Omega } models the relationship between tasks. This scheme can be used to both enforce similarity of solutions across tasks, and to learn the specific structure of task similarity by alternating between optimizations of W {\displaystyle W} and Ω {\displaystyle \Omega } . When the relationship between tasks is known to lie on a graph, the Laplacian matrix of the graph can be used to couple the learning problems. == Spectral regularization == Regularization by spectral filtering has been used to find stable solutions to problems such as those discussed above by addressing ill-posed matrix inversions (see for example Filter function for Tikhonov regularization). In many cases the regularization function acts on the input (or kernel) to ensure a bounded inverse by eliminating small singular values, but it can also be useful to have spectral norms that act on the matrix that is to be learned. There are a number of matrix norms that act on the singular values of the matrix. Frequently used examples include the Schatten p-norms, with p = 1 or 2. For example, matrix regularization with a Schatten 1-norm, also called the nuclear norm, can be used to enforce sparsity in the spectrum of a matrix. This has been used in the context of matrix completion when the matrix in question is believed to have a restricted rank. In this case the optimization problem becomes: min ‖ W ‖ ∗ subject to W i , j = Y i j . {\displaystyle \min \left\|W\right\|_{}~~{\text{ subject to }}~~W_{i,j}=Y_{ij}.} Spectral Regularization is also used to enforce a reduced rank coefficient matrix in multivariate regression. In this setting, a reduced rank coefficient matrix can be found by keeping just the top n {\displaystyle n} singular values, but this can be extended to keep any reduced set of singular values and vectors. == Structured sparsity == Sparse optimization has become the focus of much research interest as a way to find solutions that depend on a small number of variables (see e.g. the Lasso method). In principle, entry-wise sparsity can be enforced by penalizing the entry-wise ℓ 0 {\displaystyle \ell ^{0}} -norm of the matrix, but the ℓ 0 {\displaystyle \ell ^{0}} -norm is not convex. In practice this can be implemented by convex relaxation to the ℓ 1 {\displaystyle \ell ^{1}} -norm. While entry-wise regularization with an ℓ 1 {\displaystyle \ell ^{1}} -norm will find solutions with a small number of nonzero elements, applying an ℓ 1 {
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.
Semantic folding
Semantic folding theory describes a procedure for encoding the semantics of natural language text in a semantically grounded binary representation. This approach provides a framework for modelling how language data is processed by the neocortex. == Theory == Semantic folding theory draws inspiration from Douglas R. Hofstadter's Analogy as the Core of Cognition which suggests that the brain makes sense of the world by identifying and applying analogies. The theory hypothesises that semantic data must therefore be introduced to the neocortex in such a form as to allow the application of a similarity measure and offers, as a solution, the sparse binary vector employing a two-dimensional topographic semantic space as a distributional reference frame. The theory builds on the computational theory of the human cortex known as hierarchical temporal memory (HTM), and positions itself as a complementary theory for the representation of language semantics. A particular strength claimed by this approach is that the resulting binary representation enables complex semantic operations to be performed simply and efficiently at the most basic computational level. == Two-dimensional semantic space == Analogous to the structure of the neocortex, Semantic Folding theory posits the implementation of a semantic space as a two-dimensional grid. This grid is populated by context-vectors in such a way as to place similar context-vectors closer to each other, for instance, by using competitive learning principles. This vector space model is presented in the theory as an equivalence to the well known word space model described in the information retrieval literature. Given a semantic space (implemented as described above) a word-vector can be obtained for any given word Y by employing the following algorithm: For each position X in the semantic map (where X represents cartesian coordinates) if the word Y is contained in the context-vector at position X then add 1 to the corresponding position in the word-vector for Y else add 0 to the corresponding position in the word-vector for Y The result of this process will be a word-vector containing all the contexts in which the word Y appears and will therefore be representative of the semantics of that word in the semantic space. It can be seen that the resulting word-vector is also in a sparse distributed representation (SDR) format [Schütze, 1993] & [Sahlgreen, 2006]. Some properties of word-SDRs that are of particular interest with respect to computational semantics are: high noise resistance: As a result of similar contexts being placed closer together in the underlying map, word-SDRs are highly tolerant of false or shifted "bits". boolean logic: It is possible to manipulate word-SDRs in a meaningful way using boolean (OR, AND, exclusive-OR) and/or arithmetical (SUBtract) functions . sub-sampling: Word-SDRs can be sub-sampled to a high degree without any appreciable loss of semantic information. topological two-dimensional representation: The SDR representation maintains the topological distribution of the underlying map therefore words with similar meanings will have similar word-vectors. This suggests that a variety of measures can be applied to the calculation of semantic similarity, from a simple overlap of vector elements, to a range of distance measures such as: Euclidean distance, Hamming distance, Jaccard distance, cosine similarity, Levenshtein distance, Sørensen-Dice index, etc. == Semantic spaces == Semantic spaces in the natural language domain aim to create representations of natural language that are capable of capturing meaning. The original motivation for semantic spaces stems from two core challenges of natural language: Vocabulary mismatch (the fact that the same meaning can be expressed in many ways) and ambiguity of natural language (the fact that the same term can have several meanings). The application of semantic spaces in natural language processing (NLP) aims at overcoming limitations of rule-based or model-based approaches operating on the keyword level. The main drawback with these approaches is their brittleness, and the large manual effort required to create either rule-based NLP systems or training corpora for model learning. Rule-based and machine learning-based models are fixed on the keyword level and break down if the vocabulary differs from that defined in the rules or from the training material used for the statistical models. Research in semantic spaces dates back more than 20 years. In 1996, two papers were published that raised a lot of attention around the general idea of creating semantic spaces: latent semantic analysis from Microsoft and Hyperspace Analogue to Language from the University of California. However, their adoption was limited by the large computational effort required to construct and use those semantic spaces. A breakthrough with regard to the accuracy of modelling associative relations between words (e.g. "spider-web", "lighter-cigarette", as opposed to synonymous relations such as "whale-dolphin", "astronaut-driver") was achieved by explicit semantic analysis (ESA) in 2007. ESA was a novel (non-machine learning) based approach that represented words in the form of vectors with 100,000 dimensions (where each dimension represents an Article in Wikipedia). However practical applications of the approach are limited due to the large number of required dimensions in the vectors. More recently, advances in neural networking techniques in combination with other new approaches (tensors) led to a host of new recent developments: Word2vec from Google and GloVe from Stanford University. Semantic folding represents a novel, biologically inspired approach to semantic spaces where each word is represented as a sparse binary vector with 16,000 dimensions (a semantic fingerprint) in a 2D semantic map (the semantic universe). Sparse binary representation are advantageous in terms of computational efficiency, and allow for the storage of very large numbers of possible patterns. == Visualization == The topological distribution over a two-dimensional grid (outlined above) lends itself to a bitmap type visualization of the semantics of any word or text, where each active semantic feature can be displayed as e.g. a pixel. As can be seen in the images shown here, this representation allows for a direct visual comparison of the semantics of two (or more) linguistic items. Image 1 clearly demonstrates that the two disparate terms "dog" and "car" have, as expected, very obviously different semantics. Image 2 shows that only one of the meaning contexts of "jaguar", that of "Jaguar" the car, overlaps with the meaning of Porsche (indicating partial similarity). Other meaning contexts of "jaguar" e.g. "jaguar" the animal clearly have different non-overlapping contexts. The visualization of semantic similarity using Semantic Folding bears a strong resemblance to the fMRI images produced in a research study conducted by A.G. Huth et al., where it is claimed that words are grouped in the brain by meaning. voxels, little volume segments of the brain, were found to follow a pattern were semantic information is represented along the boundary of the visual cortex with visual and linguistic categories represented on posterior and anterior side respectively.