Small language models or compact language models are artificial intelligence language models designed for human natural language processing including language and text generation. They are smaller in scale and scope than large language models. A large language model typically contains hundreds of billions of training parameters, with some models exceeding a trillion parameters. This substantial parameter count enables the model to encode vast amounts of information, thereby improving the generalizability and accuracy of its outputs. However, training such models demands enormous computational resources, rendering it infeasible for an individual to do so using a single computer and graphics processing unit. Small language models, on the other hand, use far fewer parameters, typically ranging from a few thousand to a few hundred million. This make them more feasible to train and host in resource-constrained environments such as a single computer or even a mobile device. Most contemporary (2020s) small language models use the same architecture as a large language model, but with a smaller parameter count and sometimes lower arithmetic precision. Parameter count is reduced by a combination of knowledge distillation and pruning. Precision can be reduced by quantization. Work on large language models mostly translate to small language models: pruning and quantization are also widely used to speed up large language models. == Models == Some notable models are: Below 1B parameters: Llama-Prompt-Guard-2-22M (detects prompt injection and jailbreaking, based on DeBERTa-xsmall), SmolLM2-135M, SmolLM2-360M 1–4B parameters: Llama3.2-1B, Qwen2.5-1.5B, DeepSeek-R1-1.5B, SmolLM2-1.7B, SmolVLM-2.25B, Phi-3.5-Mini-3.8B, Phi-4-Mini-3.8B, Gemma3-4B; closed-weights ones include Gemini Nano 4–14B parameters: Mistral 7B, Gemma 9B, Phi-4 14B. Phi-4 14B is marginally "small" at best, but Microsoft does market it as a small model. == Language model with small pre-training dataset == Traditional AI language systems need enormous computers and vast amounts of data. Pre-training matters, even tiny models show significant performance improvements when pre-trained performance increases with larger pre-training datasets. Classification accuracy improves when pre-training and test datasets share similar tokens. Shallow architectures can replicate deep model performance through collaborative learning.
Baby Bundle (app)
Baby Bundle is a parenting mobile app for iPhone and iPad. It was designed to help new parents through pregnancy and the first two years of parenthood. Developed in collaboration with medical experts, it helps track and record the child's development and growth, offers parental advice, manages vaccinations and health check-ups, stores photos and provides baby monitoring services. == History == Baby Bundle was founded in the United Kingdom by brothers, Nick and Anthony von Christierson. Each worked in investment banking prior to developing Baby Bundle, Nick at Greenhill & Co., and Anthony at Goldman Sachs. The idea for the app came when a friend's wife voiced her frustration over having multiple parenting apps on her smartphone. Nick and Anthony left their jobs to create a single app that would include all those features. They conducted market research by interviewing more than 500 parents in the UK and US. It took them a year to build the app, which was named by their mother. Looking for endorsement, they first went to the US in 2013 and partnered with parenting expert and pediatrician Dr. Jennifer Trachtenberg. Baby Bundle was launched in the US and Canadian App Stores in April 2014. In the same month, it became the #1 parenting app in iTunes and was featured by Apple as the #1 Editor's pick across all categories. Mashable called it one of the "Top 5 Can’t Miss Apps." Baby Bundle raised $1.8m seed round in March 2015 to fund development. The money came from a range of angel investors from across the US, UK and Asia. The von Christierson brothers have signed a deal to co-brand the app in the Middle East and expect to launch in Europe and Africa. == Features == Baby Bundle is an app for both the iPhone or iPad and provides smart monitoring tools and trackers for pregnancy and child development. It acts as a growth and daily activity tracker and offers parental advice, manages vaccinations and health check-ups. It has a parenting guide with tips and advice on what to expect when the baby arrives. An interactive forum also lets parents ask questions from others in the community. The app is free and also include paid premium features like the ability to turn two iPhones running into a baby monitor, a cloud service to share the child's data with a spouse and the ability to store data on more than one baby.
Verbot
The Verbot (short for Verbal-Robot) was a chatbot program and artificial intelligence software development kit (SDK) designed for Windows and web platforms. == Early beginning == The origin of verbot traces back to Michael Mauldin's research during his time as a graduate student and post-doctoral fellow at Carnegie Mellon University. The creative foundation also stems from Peter Plantec's work in personality psychology and art direction. === Historic outline === In 1994, Michael Loren Mauldin, founder of Lycos, Inc., developed a prototype chatbot, Julia, which competed in the internationally known Turing test, for the coveted Loebner Prize. The Turing test matches computer scientist judges against machines to see if they can distinguish a computer from a real human. Julia was refined and developed, and in 1997, Dr. Mauldin and Peter Plantec, a clinical psychologist and animator, formed Virtual Personalities, Inc. (now Conversive, Inc.) in order to create a virtual human interface that would incorporate real-time animation as well as speech and natural language processing. The initial release, a stand-alone virtual person called Sylvie, was beta-tested to the public. This release was well received, and finally, after several versions, the production release (deemed version 3) of the Verbally Enhanced Software Robot, or Verbot, was deployed in fall 2000. The grandfather of all Verbots is Rog-O-Matic, which, although it could not talk, could and did explore a virtual world. Julia has been active on the internet in one form or another since 1989. A close cousin of Julia is Lycos, a robot that explores the World Wide Web and answers questions about it. Sylvie was the first Verbot with a face and a voice. Sylvie was the first Virtual Human with advanced, flexible interfacing capability. === Beginnings === The Virtual Personalities story goes back to 1978, where Mauldin was attending Rice University. Fascinated by the idea of ELIZA, he proceeded to write a program called "PET" for his 8 kilobyte Commodore PET Computer. PET included simple induction as a way to post new information, for example: Subject: I like my friend (later) Subject: I like food. PET: I have heard that food is your friend. Meanwhile, Plantec was separately designing a personality for "Entity", a theoretical virtual human that would interact comfortably with humans without pretending to be one. At that time the technology was not advanced enough to realize Entity. Mauldin got so involved with this that he majored in Computer Science and minored in Linguistics. === Rogue === In the late seventies and early eighties, a popular computer game at universities was Rogue, an implementation of Dungeons and Dragons where the player would descend 26 levels in a randomly created dungeon, fighting monsters, gathering treasure, and searching for the elusive "Amulet of Yendor". Mauldin was one of four grad students who devoted a large amount of time to building a program called "Rog-O-Matic" capable of retrieving the amulet and emerging victorious from the dungeon. === TinyMUD === In 1989, when James Aspnes at Carnegie Mellon created the first TinyMUD (a descendant of MUD and AberMUD), Mauldin was one of the first to create a computer player that would explore the text-based world of TinyMUD. But his first robot, Gloria, gradually accreted more and more linguistic ability, to the point that it could pass the "unsuspecting" Turing test. In this version of the test, the human has no reason to suspect that one of the other occupants of the room is controlled by a computer, and so is more polite and asks fewer probing questions. The second generation of Mauldin's TinyMUD robots was Julia, created on Jan. 8, 1990. Julia slowly developed into a more and more capable conversational agent, and assumed useful duties in the TinyMUD world, including tour guide, information assistant, note-taker, and message-relayer. She could even play the card game hearts along with the other human players. In 1991, Julia attended the first Loebner Prize contest in Boston, Massachusetts. Although she only finished third, she was ranked by one judge as more human than one of the human confederates, winning a coveted certificate of humanness in the world's first restricted Turing test. Julia continued to log in to various TinyMUD's and TinyMucks for the next seven years, and chatted with hundreds of people a month over the internet. === Lycos === Julia's job was to explore a virtual world consisting of pages of textual descriptions, with links between them, and to construct an internal map of that world and answer questions about it (including path information such as the shortest route from one room to another, and matching information, such as which rooms contained a certain kind of object or textual description). It was therefore only a very short cognitive leap from Julia to Lycos, another robotic agent that explores a virtual world made of hyperlinked pages of text, and which answers questions about those pages. Sylvie was born and her abilities were expanded greatly to include interfacing with computers and control systems via her serial ports. === Sylvie === Sylvie was the first intelligent animated virtual human. She was designed both as a conversation agent and as a virtual human interface that would form a bridge between the two. She became more popular as a conversation agent, but her designers believe she serves as a prototype for future virtual human interface design that will help us all cope with the increasing complexity of technology. As an aside, Plantec noticed that a large number of Sylvies have been sold in Southeast Asia. Upon investigation, he found out that students had discovered a "test" mode that would allow them to type in English sentences that Sylvie would pronounce in her somewhat stylized English. == Ownership == In 1997, Dr. Mauldin and Peter Plantec formed Virtual Personalities, Inc. to create Natural Language Processing solutions for companies. In 2001 Virtual Personalities, Inc. became Conversive, Inc. to reflect the focus on providing Customer Service and Marketing to the Enterprise Market. In late 2012 Avaya, Inc. acquired Conversive's assets including Verbots. == Verbot versions == The Verbot 4 version was created and released in 2004. In 2005 Version 4.1 of the Verbot Software was released with many feature enhancements and bug fixes, including built-in support for embedding C# code in outputs and conditionals. In early 2006 Conversive launched Verbots Online allowing Verbot 4 users to upload their knowledge and show off their bots to the world. In 2009 Version 5 was released, completely free and fully featured. In early 2012 the last version of Verbot, 5.0.1.2, was released to the general public with support for Windows 7. Later in 2012 Verbots Online completely shut down. == Verbots today == Verbots.com, its community of users, and its forums no longer exist, but the software and users can still be found. There has been no active development since the early 2012 release of Verbot 5.0.1.2.
Deep image prior
Deep image prior is a type of convolutional neural network used to enhance a given image with no prior training data other than the image itself. A neural network is randomly initialized and used as prior to solve inverse problems such as noise reduction, super-resolution, and inpainting. Image statistics are captured by the structure of a convolutional image generator rather than by any previously learned capabilities. == Method == === Background === Inverse problems such as noise reduction, super-resolution, and inpainting can be formulated as the optimization task x ∗ = m i n x E ( x ; x 0 ) + R ( x ) {\displaystyle x^{}=min_{x}E(x;x_{0})+R(x)} , where x {\displaystyle x} is an image, x 0 {\displaystyle x_{0}} a corrupted representation of that image, E ( x ; x 0 ) {\displaystyle E(x;x_{0})} is a task-dependent data term, and R(x) is the regularizer. Deep neural networks learn a generator/decoder x = f θ ( z ) {\displaystyle x=f_{\theta }(z)} which maps a random code vector z {\displaystyle z} to an image x {\displaystyle x} . The image corruption method used to generate x 0 {\displaystyle x_{0}} is selected for the specific application. === Specifics === In this approach, the R ( x ) {\displaystyle R(x)} prior is replaced with the implicit prior captured by the neural network (where R ( x ) = 0 {\displaystyle R(x)=0} for images that can be produced by a deep neural networks and R ( x ) = + ∞ {\displaystyle R(x)=+\infty } otherwise). This yields the equation for the minimizer θ ∗ = a r g m i n θ E ( f θ ( z ) ; x 0 ) {\displaystyle \theta ^{}=argmin_{\theta }E(f_{\theta }(z);x_{0})} and the result of the optimization process x ∗ = f θ ∗ ( z ) {\displaystyle x^{}=f_{\theta ^{}}(z)} . The minimizer θ ∗ {\displaystyle \theta ^{}} (typically a gradient descent) starts from a randomly initialized parameters and descends into a local best result to yield the x ∗ {\displaystyle x^{}} restoration function. ==== Overfitting ==== A parameter θ may be used to recover any image, including its noise. However, the network is reluctant to pick up noise because it contains high impedance while useful signal offers low impedance. This results in the θ parameter approaching a good-looking local optimum so long as the number of iterations in the optimization process remains low enough not to overfit data. === Deep Neural Network Model === Typically, the deep neural network model for deep image prior uses a U-Net like model without the skip connections that connect the encoder blocks with the decoder blocks. The authors in their paper mention that "Our findings here (and in other similar comparisons) seem to suggest that having deeper architecture is beneficial, and that having skip-connections that work so well for recognition tasks (such as semantic segmentation) is highly detrimental." == Applications == === Denoising === The principle of denoising is to recover an image x {\displaystyle x} from a noisy observation x 0 {\displaystyle x_{0}} , where x 0 = x + ϵ {\displaystyle x_{0}=x+\epsilon } . The distribution ϵ {\displaystyle \epsilon } is sometimes known (e.g.: profiling sensor and photon noise) and may optionally be incorporated into the model, though this process works well in blind denoising. The quadratic energy function E ( x , x 0 ) = | | x − x 0 | | 2 {\displaystyle E(x,x_{0})=||x-x_{0}||^{2}} is used as the data term, plugging it into the equation for θ ∗ {\displaystyle \theta ^{}} yields the optimization problem m i n θ | | f θ ( z ) − x 0 | | 2 {\displaystyle min_{\theta }||f_{\theta }(z)-x_{0}||^{2}} . === Super-resolution === Super-resolution is used to generate a higher resolution version of image x. The data term is set to E ( x ; x 0 ) = | | d ( x ) − x 0 | | 2 {\displaystyle E(x;x_{0})=||d(x)-x_{0}||^{2}} where d(·) is a downsampling operator such as Lanczos that decimates the image by a factor t. === Inpainting === Inpainting is used to reconstruct a missing area in an image x 0 {\displaystyle x_{0}} . These missing pixels are defined as the binary mask m ∈ { 0 , 1 } H × V {\displaystyle m\in \{0,1\}^{H\times V}} . The data term is defined as E ( x ; x 0 ) = | | ( x − x 0 ) ⊙ m | | 2 {\displaystyle E(x;x_{0})=||(x-x_{0})\odot m||^{2}} (where ⊙ {\displaystyle \odot } is the Hadamard product). The intuition behind this is that the loss is computed only on the known pixels in the image, and the network is going to learn enough about the image to fill in unknown parts of the image even though the computed loss doesn't include those pixels. This strategy is used to remove image watermarks by treating the watermark as missing pixels in the image. === Flash–no-flash reconstruction === This approach may be extended to multiple images. A straightforward example mentioned by the author is the reconstruction of an image to obtain natural light and clarity from a flash–no-flash pair. Video reconstruction is possible but it requires optimizations to take into account the spatial differences. == Implementations == A reference implementation rewritten in Python 3.6 with the PyTorch 0.4.0 library was released by the author under the Apache 2.0 license: deep-image-prior A TensorFlow-based implementation written in Python 2 and released under the CC-SA 3.0 license: deep-image-prior-tensorflow A Keras-based implementation written in Python 2 and released under the GPLv3: machine_learning_denoising == Example == See Astronomy Picture of the Day (APOD) of 2024-02-18
Randomized Hough transform
Hough transforms are techniques for object detection, a critical step in many implementations of computer vision, or data mining from images. Specifically, the Randomized Hough transform is a probabilistic variant to the classical Hough transform, and is commonly used to detect curves (straight line, circle, ellipse, etc.) The basic idea of Hough transform (HT) is to implement a voting procedure for all potential curves in the image, and at the termination of the algorithm, curves that do exist in the image will have relatively high voting scores. Randomized Hough transform (RHT) is different from HT in that it tries to avoid conducting the computationally expensive voting process for every nonzero pixel in the image by taking advantage of the geometric properties of analytical curves, and thus improve the time efficiency and reduce the storage requirement of the original algorithm. == Motivation == Although Hough transform (HT) has been widely used in curve detection, it has two major drawbacks: First, for each nonzero pixel in the image, the parameters for the existing curve and redundant ones are both accumulated during the voting procedure. Second, the accumulator array (or Hough space) is predefined in a heuristic way. The more accuracy needed, the higher parameter resolution should be defined. These two needs usually result in a large storage requirement and low speed for real applications. Therefore, RHT was brought up to tackle this problem. == Implementation == In comparison with HT, RHT takes advantage of the fact that some analytical curves can be fully determined by a certain number of points on the curve. For example, a straight line can be determined by two points, and an ellipse (or a circle) can be determined by three points. The case of ellipse detection can be used to illustrate the basic idea of RHT. The whole process generally consists of three steps: Fit ellipses with randomly selected points. Update the accumulator array and corresponding scores. Output the ellipses with scores higher than some predefined threshold. === Ellipse fitting === One general equation for defining ellipses is: a ( x − p ) 2 + 2 b ( x − p ) ( y − q ) + c ( y − q ) 2 = 1 {\displaystyle a(x-p)^{2}+2b(x-p)(y-q)+c(y-q)^{2}=1} with restriction: a c − b 2 > 0 {\displaystyle ac-b^{2}>0} However, an ellipse can be fully determined if one knows three points on it and the tangents in these points. RHT starts by randomly selecting three points on the ellipse. Let them be X 1 {\displaystyle X_{1}} , X 2 {\displaystyle X_{2}} and X 3 {\displaystyle X_{3}} . The first step is to find the tangents of these three points. They can be found by fitting a straight line using least squares technique for a small window of neighboring pixels. The next step is to find the intersection points of the tangent lines. This can be easily done by solving the line equations found in the previous step. Then let the intersection points be T 12 {\displaystyle T_{12}} and T 23 {\displaystyle T_{23}} , the midpoints of line segments X 1 X 2 {\displaystyle X_{1}X_{2}} and X 2 X 3 {\displaystyle X_{2}X_{3}} be M 12 {\displaystyle M_{12}} and M 23 {\displaystyle M_{23}} . Then the center of the ellipse will lie in the intersection of T 12 M 12 {\displaystyle T_{12}M_{12}} and T 23 M 23 {\displaystyle T_{23}M_{23}} . Again, the coordinates of the intersected point can be determined by solving line equations and the detailed process is skipped here for conciseness. Let the coordinates of ellipse center found in previous step be ( x 0 , y 0 ) {\displaystyle (x_{0},y_{0})} . Then the center can be translated to the origin with x ′ = x − x 0 {\displaystyle x'=x-x_{0}} and y ′ = y − y 0 {\displaystyle y'=y-y_{0}} so that the ellipse equation can be simplified to: a x ′ 2 + 2 b x ′ y ′ + c y ′ 2 = 1 {\displaystyle ax'^{2}+2bx'y'+cy'^{2}=1} Now we can solve for the rest of ellipse parameters: a {\displaystyle a} , b {\displaystyle b} and c {\displaystyle c} by substituting the coordinates of X 1 {\displaystyle X_{1}} , X 2 {\displaystyle X_{2}} and X 3 {\displaystyle X_{3}} into the equation above. === Accumulating === With the ellipse parameters determined from previous stage, the accumulator array can be updated correspondingly. Different from classical Hough transform, RHT does not keep "grid of buckets" as the accumulator array. Rather, it first calculates the similarities between the newly detected ellipse and the ones already stored in accumulator array. Different metrics can be used to calculate the similarity. As long as the similarity exceeds some predefined threshold, replace the one in the accumulator with the average of both ellipses and add 1 to its score. Otherwise, initialize this ellipse to an empty position in the accumulator and assign a score of 1. === Termination === Once the score of one candidate ellipse exceeds the threshold, it is determined as existing in the image (in other words, this ellipse is detected), and should be removed from the image and accumulator array so that the algorithm can detect other potential ellipses faster. The algorithm terminates when the number of iterations reaches a maximum limit or all the ellipses have been detected. Pseudo code for RHT: while (we find ellipses AND not reached the maximum epoch) { for (a fixed number of iterations) { Find a potential ellipse. if (the ellipse is similar to an ellipse in the accumulator) then Replace the one in the accumulator with the average of two ellipses and add 1 to the score; else Insert the ellipse into an empty position in the accumulator with a score of 1; } Select the ellipse with the best score and save it in a best ellipse table; Eliminate the pixels of the best ellipse from the image; Empty the accumulator; }
Attensity
Attensity was an American company that provided social analytics and engagement applications for social customer relationship management (social CRM). Attensity's text analytics software applications extracted facts, relationships and sentiment from unstructured data. == History == Attensity was founded in 2000. An early investor in Attensity was In-Q-Tel, which funds technology to support the missions of the US Government and the broader DOD. InTTENSITY, an independent company that has combined Inxight with Attensity Software (the only joint development project that combines two InQTel funded software packages), was the exclusive distributor and outlet for Attensity in the Federal Market. In 2009, Attensity Corp., then based in Palo Alto, merged with Germany's Empolis and Living-e AG to form Attensity Group. In 2010, Attensity Group acquired Biz360, a provider of social media monitoring and market intelligence solutions. In early 2012, Attensity Group divested itself of the Empolis business unit via a management buyout; that unit currently conducts business under its pre-merger name. Attensity Group was a closely held private company. Its majority shareholder was Aeris Capital, a private Swiss investment office advising a high-net-worth individual and his charitable foundation. Foundation Capital, Granite Ventures, and Scale Venture Partners were among Biz360's investors and thus became shareholders in Attensity Group. In February 2016, Attensity's IP assets were acquired by InContact, and Attensity closed.
Neural field
In machine learning, a neural field (also known as implicit neural representation, neural implicit, or coordinate-based neural network), is a mathematical field that is fully or partially parametrized by a neural network. Initially developed to tackle visual computing tasks, such as rendering or reconstruction (e.g., neural radiance fields), neural fields emerged as a promising strategy to deal with a wider range of problems, including surrogate modelling of partial differential equations, such as in physics-informed neural networks. Differently from traditional machine learning algorithms, such as feed-forward neural networks, convolutional neural networks, or transformers, neural fields do not work with discrete data (e.g. sequences, images, tokens), but map continuous inputs (e.g., spatial coordinates, time) to continuous outputs (i.e., scalars, vectors, etc.). This makes neural fields not only discretization independent, but also easily differentiable. Moreover, dealing with continuous data allows for a significant reduction in space complexity, which translates to a much more lightweight network. == Formulation and training == According to the universal approximation theorem, provided adequate learning, sufficient number of hidden units, and the presence of a deterministic relationship between the input and the output, a neural network can approximate any function to any degree of accuracy. Hence, in mathematical terms, given a field y = Φ ( x ) {\textstyle {\boldsymbol {y}}=\Phi ({\boldsymbol {x}})} , with x ∈ R n {\displaystyle {\boldsymbol {x}}\in \mathbb {R} ^{n}} and y ∈ R m {\displaystyle {\boldsymbol {y}}\in \mathbb {R} ^{m}} , a neural field Ψ θ {\displaystyle \Psi _{\theta }} , with parameters θ {\displaystyle {\boldsymbol {\theta }}} , is such that: Ψ θ ( x ) = y ^ ≈ y {\displaystyle \Psi _{\theta }({\boldsymbol {x}})={\hat {\boldsymbol {y}}}\approx {\boldsymbol {y}}} === Training === For supervised tasks, given N {\displaystyle N} examples in the training dataset (i.e., ( x i , y i ) ∈ D t r a i n , i = 1 , … , N {\displaystyle ({\boldsymbol {x_{i}}},{\boldsymbol {y_{i}}})\in {\mathcal {D_{train}}},i=1,\dots ,N} ), the neural field parameters can be learned by minimizing a loss function L {\displaystyle {\mathcal {L}}} (e.g., mean squared error). The parameters θ ~ {\displaystyle {\tilde {\theta }}} that satisfy the optimization problem are found as: θ ~ = argmin θ 1 N ∑ ( x i , y i ) ∈ D t r a i n L ( Ψ θ ( x i ) , y i ) {\displaystyle {\tilde {\boldsymbol {\theta }}}={\underset {\boldsymbol {\theta }}{\text{argmin}}}\;{\frac {1}{N}}\sum _{({\boldsymbol {x_{i}}},{\boldsymbol {y_{i}}})\in {\mathcal {D_{train}}}}{\mathcal {L}}(\Psi _{\theta }({\boldsymbol {x}}_{i}),{\boldsymbol {y}}_{i})} Notably, it is not necessary to know the analytical expression of Φ {\displaystyle \Phi } , for the previously reported training procedure only requires input-output pairs. Indeed, a neural field is able to offer a continuous and differentiable surrogate of the true field, even from purely experimental data. Moreover, neural fields can be used in unsupervised settings, with training objectives that depend on the specific task. For example, physics-informed neural networks may be trained on just the residual. === Spectral bias === As for any artificial neural network, neural fields may be characterized by a spectral bias (i.e., the tendency to preferably learn the low frequency content of a field), possibly leading to a poor representation of the ground truth. In order to overcome this limitation, several strategies have been developed. For example, SIREN uses sinusoidal activations, while the Fourier-features approach embeds the input through sines and cosines. == Conditional neural fields == In many real-world cases, however, learning a single field is not enough. For example, when reconstructing 3D vehicle shapes from Lidar data, it is desirable to have a machine learning model that can work with arbitrary shapes (e.g., a car, a bicycle, a truck, etc.). The solution is to include additional parameters, the latent variables (or latent code) z ∈ R d {\displaystyle {\boldsymbol {z}}\in \mathbb {R} ^{d}} , to vary the field and adapt it to diverse tasks. === Latent code production === When dealing with conditional neural fields, the first design choice is represented by the way in which the latent code is produced. Specifically, two main strategies can be identified: Encoder: the latent code is the output of a second neural network, acting as an encoder. During training, the loss function is the objective used to learn the parameters of both the neural field and the encoder. Auto-decoding: each training example has its own latent code, jointly trained with the neural field parameters. When the model has to process new examples (i.e., not originally present in the training dataset), a small optimization problem is solved, keeping the network parameters fixed and only learning the new latent variables. Since the latter strategy requires additional optimization steps at inference time, it sacrifices speed, but keeps the overall model smaller. Moreover, despite being simpler to implement, an encoder may harm the generalization capabilities of the model. For example, when dealing with a physical scalar field f : R 2 → R {\displaystyle f:\mathbb {R} ^{2}\rightarrow \mathbb {R} } (e.g., the pressure of a 2D fluid), an auto-decoder-based conditional neural field can map a single point to the corresponding value of the field, following a learned latent code z {\displaystyle {\boldsymbol {z}}} . However, if the latent variables were produced by an encoder, it would require access to the entire set of points and corresponding values (e.g. as a regular grid or a mesh graph), leading to a less robust model. === Global and local conditioning === In a neural field with global conditioning, the latent code does not depend on the input and, hence, it offers a global representation (e.g., the overall shape of a vehicle). However, depending on the task, it may be more useful to divide the domain of x {\displaystyle {\boldsymbol {x}}} in several subdomains, and learn different latent codes for each of them (e.g., splitting a large and complex scene in sub-scenes for a more efficient rendering). This is called local conditioning. === Conditioning strategies === There are several strategies to include the conditioning information in the neural field. In the general mathematical framework, conditioning the neural field with the latent variables is equivalent to mapping them to a subset θ ∗ {\displaystyle {\boldsymbol {\theta }}^{}} of the neural field parameters: θ ∗ = Γ ( z ) {\displaystyle {\boldsymbol {\theta }}^{}=\Gamma ({\boldsymbol {z}})} In practice, notable strategies are: Concatenation: the neural field receives, as input, the concatenation of the original input x {\displaystyle {\boldsymbol {x}}} with the latent codes z {\displaystyle {\boldsymbol {z}}} . For feed-forward neural networks, this is equivalent to setting θ ∗ {\displaystyle {\boldsymbol {\theta }}^{}} as the bias of the first layer and Γ ( z ) {\displaystyle \Gamma ({\boldsymbol {z}})} as an affine transformation. Hypernetworks: a hypernetwork is a neural network that outputs the parameters of another neural network. Specifically, it consists of approximating Γ ( z ) {\displaystyle \Gamma ({\boldsymbol {z}})} with a neural network Γ ^ γ ( z ) {\displaystyle {\hat {\Gamma }}_{\gamma }({\boldsymbol {z}})} , where γ {\displaystyle {\boldsymbol {\gamma }}} are the trainable parameters of the hypernetwork. This approach is the most general, as it allows to learn the optimal mapping from latent codes to neural field parameters. However, hypernetworks are associated to larger computational and memory complexity, due to the large number of trainable parameters. Hence, leaner approaches have been developed. For example, in the Feature-wise Linear Modulation (FiLM), the hypernetwork only produces scale and bias coefficients for the neural field layers. === Meta-learning === Instead of relying on the latent code to adapt the neural field to a specific task, it is also possible to exploit gradient-based meta-learning. In this case, the neural field is seen as the specialization of an underlying meta-neural-field, whose parameters are modified to fit the specific task, through a few steps of gradient descent. An extension of this meta-learning framework is the CAVIA algorithm, that splits the trainable parameters in context-specific and shared groups, improving parallelization and interpretability, while reducing meta-overfitting. This strategy is similar to the auto-decoding conditional neural field, but the training procedure is substantially different. == Applications == Thanks to the possibility of efficiently modelling diverse mathematical fields with neural networks, neural fields have been applied to a wide range of problems: 3D scene reconstruction: neural fields can be used to model t