In machine learning and statistics, the learning rate is a tuning parameter in an optimization algorithm that determines the step size at each iteration while moving toward a minimum of a loss function. Since it influences to what extent newly acquired information overrides old information, it metaphorically represents the speed at which a machine learning model "learns". In the adaptive control literature, the learning rate is commonly referred to as gain. In setting a learning rate, there is a trade-off between the rate of convergence and overshooting. While the descent direction is usually determined from the gradient of the loss function, the learning rate determines how big a step is taken in that direction. Too high a learning rate will make the learning jump over minima, but too low a learning rate will either take too long to converge or get stuck in an undesirable local minimum. In order to achieve faster convergence, prevent oscillations and getting stuck in undesirable local minima the learning rate is often varied during training either in accordance to a learning rate schedule or by using an adaptive learning rate. The learning rate and its adjustments may also differ per parameter, in which case it is a diagonal matrix that can be interpreted as an approximation to the inverse of the Hessian matrix in Newton's method. The learning rate is related to the step length determined by inexact line search in quasi-Newton methods and related optimization algorithms. == Learning rate schedule == Initial rate can be left as system default or can be selected using a range of techniques. A learning rate schedule changes the learning rate during learning and is most often changed between epochs/iterations. This is mainly done with two parameters: decay and momentum. There are many different learning rate schedules but the most common are time-based, step-based and exponential. Decay serves to settle the learning in a nice place and avoid oscillations, a situation that may arise when too high a constant learning rate makes the learning jump back and forth over a minimum, and is controlled by a hyperparameter. Momentum is analogous to a ball rolling down a hill; we want the ball to settle at the lowest point of the hill (corresponding to the lowest error). Momentum both speeds up the learning (increasing the learning rate) when the error cost gradient is heading in the same direction for a long time and also avoids local minima by 'rolling over' small bumps. Momentum is controlled by a hyperparameter analogous to a ball's mass which must be chosen manually—too high and the ball will roll over minima which we wish to find, too low and it will not fulfil its purpose. The formula for factoring in the momentum is more complex than for decay but is most often built in with deep learning libraries such as Keras. Time-based learning schedules alter the learning rate depending on the learning rate of the previous time iteration. Factoring in the decay the mathematical formula for the learning rate is: η n + 1 = η 0 1 + d n {\displaystyle \eta _{n+1}={\frac {\eta _{0}}{1+dn}}} where η {\displaystyle \eta } is the learning rate, η 0 {\displaystyle \eta _{0}} is the original learning rate, d {\displaystyle d} is a decay parameter and n {\displaystyle n} is the iteration step. Step-based learning schedules changes the learning rate according to some predefined steps. The decay application formula is here defined as: η n = η 0 d ⌊ 1 + n r ⌋ {\displaystyle \eta _{n}=\eta _{0}d^{\left\lfloor {\frac {1+n}{r}}\right\rfloor }} where η n {\displaystyle \eta _{n}} is the learning rate at iteration n {\displaystyle n} , η 0 {\displaystyle \eta _{0}} is the initial learning rate, d {\displaystyle d} is how much the learning rate should change at each drop (0.5 corresponds to a halving) and r {\displaystyle r} corresponds to the drop rate, or how often the rate should be dropped (10 corresponds to a drop every 10 iterations). The floor function ( ⌊ … ⌋ {\displaystyle \lfloor \dots \rfloor } ) here drops the value of its input to 0 for all values smaller than 1. Exponential learning schedules are similar to step-based, but instead of steps, a decreasing exponential function is used. The mathematical formula for factoring in the decay is: η n = η 0 e − d n {\displaystyle \eta _{n}=\eta _{0}e^{-dn}} where d {\displaystyle d} is a decay parameter. == Adaptive learning rate == The issue with learning rate schedules is that they all depend on hyperparameters that must be manually chosen for each given learning session and may vary greatly depending on the problem at hand or the model used. To combat this, there are many different types of adaptive gradient descent algorithms such as Adagrad, Adadelta, RMSprop, and Adam which are generally built into deep learning libraries such as Keras.
TIMIT
TIMIT is a corpus of phonemically and lexically transcribed speech of American English speakers of different sexes and dialects. Each transcribed element has been delineated in time. TIMIT was designed to further acoustic-phonetic knowledge and automatic speech recognition systems. It was commissioned by DARPA and corpus design was a joint effort between the Massachusetts Institute of Technology, SRI International, and Texas Instruments (TI). The speech was recorded at TI, transcribed at MIT, and verified and prepared for publishing by the National Institute of Standards and Technology (NIST). There is also a telephone bandwidth version called NTIMIT (Network TIMIT). TIMIT and NTIMIT are not freely available — either membership of the Linguistic Data Consortium, or a monetary payment, is required for access to the dataset. == Data == TIMIT contains ~5 hours of speech, of 10 sentences spoken by each of 630 speakers. The sentences were randomly sampled from a corpus of 2342 sentences. The speakers were native speakers of American English, classified under 8 major dialect regions: New England, Northern, North Midland, South Midland, Southern, New York City, Western, Army Brat (moved around). The speakers were 70% male and 30% female. Recordings were made in a noise-isolated recording booth at Texas Instrument, using a semi-automatic computer system (STEROIDS) to control the presentation of prompts to the speaker and the recording. Two-channel recordings were made using a Sennheiser HMD 414 headset-mounted microphone and a Brüel & Kjær 1/2" far-field pressure microphone (#4165). The speech was digitized at a sample rate of 20 kHz then and downsampled to 16 kHz. == History == The TIMIT telephone corpus was an early attempt to create a database with speech samples. It was published in the year 1988 on CD-ROM and consists of only 10 sentences per speaker. Two 'dialect' sentences were read by each speaker, as well as another 8 sentences selected from a larger set Each sentence averages 3 seconds long and is spoken by 630 different speakers. It was the first notable attempt in creating and distributing a speech corpus and the overall project has produced costs of 1.5 million US$. An update was released in October 1990. It included full 630-speaker corpus; checked and corrected transcriptions; word-alignment transcriptions; NIST SPHERE-headered waveform files and header manipulation software; phonemic dictionary; new test and training subsets balanced for dialectal and phonetic coverage; more extensive documentation. The full name of the project is DARPA-TIMIT Acoustic-Phonetic Continuous Speech Corpus and the acronym TIMIT stands for Texas Instruments/Massachusetts Institute of Technology. The main reason why a corpus of telephone speech was created was to train speech recognition software. In the Blizzard challenge, different software has the obligation to convert audio recordings into textual data and the TIMIT corpus was used as a standardized baseline.
Azuqua
Azuqua is an American cloud-based integration and automation company headquartered in Seattle, Washington. As such, they integrate SaaS applications and create automations that are designed to eliminate manual work. Azuqua's platform has the ability to set up workflows between multiple applications so disparate teams can stay in the loop. Azuqua's customers include companies such as Charles Schwab, General Electric, General Motors, HubSpot, and Airbnb. == History == Nikhil Hasija and Craig Unger founded Azuqua in 2011. In 2013, the team participated in Techstars Microsoft's Windows Azure Accelerator, a Seattle-based incubator that helps entrepreneurs gain traction through deep mentor engagement and rapid iteration cycles. Azuqua announced in 2014 that they have received their Series A funding from Ignition Partners which amounted to $5 million. 2017 included a 65% growth in new customers, a doubling of new SaaS connectors, and a 50% growth in overall employee headcount. Azuqua also received their Series B funding which totaled to $10.8 million. This funding was led by Insight Ventures Partners, with DFJ and Ignition Partners also joining the round In March 2018, Azuqua hired Todd Owens as CEO. Owens was previously CEO of Appuri, a customer data platform. Hasija has transitioned to the role of Chief Product Officer. Azuqua also hired on Dan Kogan who has taken on the role of Chief Marketing Officer. Kogan previously worked at Tableau, a BI and analytics company, as a Senior Director of Product Marketing. Okta acquired Azuqua in 2019. == Product Description/Features == Logic Library: Logic functions that can be used for data processing, branching logic, and business rules Drag and Drop Visual Designer: No-code visual designer Use of API's for each cloud service a business is using to allow the various apps to communicate and share data API Publishing: Integrations and automations can be made available as secure endpoints, webhooks, or open services Connector Builder: Build a connector to an application Connector Library: Pre-built connectors to SaaS applications Error Handling: Automations that execute when an error is detected
G.9963
Recommendation G.9963 is a home networking standard under development at the International Telecommunication Union standards sector, the ITU-T. It was begun in 2010 by ITU-T to add multiple-input and multiple-output (known as MIMO) capabilities to the G.hn standard originally defined in Recommendation G.9960. The standard is also known as "G.hn-mimo". As part of the family of G.hn standards, G.9963 was endorsed by the HomeGrid Forum.
Data storage
Data storage is the recording (storing) of information (data) in a storage medium. Handwriting, phonographic recording, magnetic tape, and optical discs are all examples of storage media. Biological molecules such as RNA and DNA are considered by some as data storage. Recording may be accomplished with virtually any form of energy. Electronic data storage requires electrical power to store and retrieve data. Data stored in a digital, machine-readable medium is called digital data. Computer data storage is one of the core functions of a general-purpose computer. Electronic documents can be stored in much less space than paper documents. Barcodes and magnetic ink character recognition (MICR) are two ways of recording machine-readable data on paper. == Recording media == A recording medium is physical material that holds information. Newly created information is distributed and can be stored in four storage media–print, film, magnetic, and optical–and seen or heard in four information flows–telephone, radio, TV, and the Internet as well as being observed directly. Digital information is stored on electronic media in many different recording formats. With electronic media, the data and the recording media are sometimes referred to as "software" despite the more common use of the word to describe computer software. With (traditional art) static media, art materials such as crayons may be considered both equipment and medium as the wax, charcoal or chalk material from the equipment becomes part of the surface of the medium. Some recording media may be temporary, either by design or by nature. Volatile organic compounds may be used to purposely make data expire over time or to reduce environmental impact. Data such as smoke signals or skywriting are temporary by nature. Depending on the volatility, a gas (e.g., atmosphere, smoke) or a liquid surface such as a lake would be considered a temporary recording medium, if it could be considered a recording medium at all. == Global capacity, digitization, and trends == A 2003 UC Berkeley report estimated that about five exabytes of new information were produced in 2002 and that 92% of this data was stored on magnetic media (primarily hard disk drives). This was about twice the data produced in 1999. The amount of data transmitted over telecommunications systems in 2002 was nearly 18 exabytes—three and a half times more than was recorded on non-volatile storage. Telephone calls constituted 98% of the telecommunicated information in 2002. The researchers' highest estimate for the growth rate of newly stored information (uncompressed) was more than 30% per year. In a more limited study, the International Data Corporation estimated that the total amount of digital data in 2007 was 281 exabytes and that the total amount of digital data produced exceeded the global storage capacity for the first time. A 2011 article in Science estimated that the year 2002 was the beginning of the digital age for information storage: an age in which more information is stored on digital storage devices than on analog storage devices. In 1986, approximately 1% of the world's capacity to store information was in digital format; this grew to 3% by 1993, to 25% by 2000, and to 94% by 2007. These figures correspond to less than three compressed exabytes in 1986, and 295 compressed exabytes in 2007. The quantity of digital storage doubled roughly every three to four years. It is estimated that around 120 zettabytes of data will be generated in 2023, an increase of 60x from 2010, and that it will increase to 181 zettabytes generated in 2025. == Mass storage ==
Object co-segmentation
In computer vision, object co-segmentation is a special case of image segmentation, which is defined as jointly segmenting semantically similar objects in multiple images or video frames. == Challenges == It is often challenging to extract segmentation masks of a target/object from a noisy collection of images or video frames, which involves object discovery coupled with segmentation. A noisy collection implies that the object/target is present sporadically in a set of images or the object/target disappears intermittently throughout the video of interest. Early methods typically involve mid-level representations such as object proposals. == Dynamic Markov networks-based methods == A joint object discover and co-segmentation method based on coupled dynamic Markov networks has been proposed recently, which claims significant improvements in robustness against irrelevant/noisy video frames. Unlike previous efforts which conveniently assumes the consistent presence of the target objects throughout the input video, this coupled dual dynamic Markov network based algorithm simultaneously carries out both the detection and segmentation tasks with two respective Markov networks jointly updated via belief propagation. Specifically, the Markov network responsible for segmentation is initialized with superpixels and provides information for its Markov counterpart responsible for the object detection task. Conversely, the Markov network responsible for detection builds the object proposal graph with inputs including the spatio-temporal segmentation tubes. == Graph cut-based methods == Graph cut optimization is a popular tool in computer vision, especially in earlier image segmentation applications. As an extension of regular graph cuts, multi-level hypergraph cut is proposed to account for more complex high order correspondences among video groups beyond typical pairwise correlations. With such hypergraph extension, multiple modalities of correspondences, including low-level appearance, saliency, coherent motion and high level features such as object regions, could be seamlessly incorporated in the hyperedge computation. In addition, as a core advantage over co-occurrence based approach, hypergraph implicitly retains more complex correspondences among its vertices, with the hyperedge weights conveniently computed by eigenvalue decomposition of Laplacian matrices. == CNN/LSTM-based methods == In action localization applications, object co-segmentation is also implemented as the segment-tube spatio-temporal detector. Inspired by the recent spatio-temporal action localization efforts with tubelets (sequences of bounding boxes), Le et al. present a new spatio-temporal action localization detector Segment-tube, which consists of sequences of per-frame segmentation masks. This Segment-tube detector can temporally pinpoint the starting/ending frame of each action category in the presence of preceding/subsequent interference actions in untrimmed videos. Simultaneously, the Segment-tube detector produces per-frame segmentation masks instead of bounding boxes, offering superior spatial accuracy to tubelets. This is achieved by alternating iterative optimization between temporal action localization and spatial action segmentation. The proposed segment-tube detector is illustrated in the flowchart on the right. The sample input is an untrimmed video containing all frames in a pair figure skating video, with only a portion of these frames belonging to a relevant category (e.g., the DeathSpirals). Initialized with saliency based image segmentation on individual frames, this method first performs temporal action localization step with a cascaded 3D CNN and LSTM, and pinpoints the starting frame and the ending frame of a target action with a coarse-to-fine strategy. Subsequently, the segment-tube detector refines per-frame spatial segmentation with graph cut by focusing on relevant frames identified by the temporal action localization step. The optimization alternates between the temporal action localization and spatial action segmentation in an iterative manner. Upon practical convergence, the final spatio-temporal action localization results are obtained in the format of a sequence of per-frame segmentation masks (bottom row in the flowchart) with precise starting/ending frames.
Multistage interconnection networks
Multistage interconnection networks (MINs) are a class of high-speed computer networks usually composed of processing elements (PEs) on one end of the network and memory elements (MEs) on the other end, connected by switching elements (SEs). The switching elements themselves are usually connected to each other in stages, hence the name. MINs are typically used in high-performance or parallel computing as a low-latency interconnection (as opposed to traditional packet switching networks), though they could be implemented on top of a packet switching network. Though the network is typically used for routing purposes, it could also be used as a co-processor to the actual processors for such uses as sorting; cyclic shifting, as in a perfect shuffle network; and bitonic sorting. == Background == Interconnection network are used to connect nodes, where nodes can be a single processor or group of processors, to other nodes. Interconnection networks can be categorized on the basis of their topology. Topology is the pattern in which one node is connected to other nodes. There are two main types of topology: static and dynamic. Static interconnect networks are hard-wired and cannot change their configurations. A regular static interconnect is mainly used in small networks made up of loosely couple nodes. The regular structure signifies that the nodes are arranged in specific shape and the shape is maintained throughout the networks. Some examples of static regular interconnections are: Completely connected network In a mesh network, multiple nodes are connected with each other. Each node in the network is connected to every other node in the network. This arrangement allows proper communication of the data between the nodes. But, there are a lot of communication overheads due to the increased number of node connections. Shared busThis network topology involves connection of the nodes with each other over a bus. Every node communicates with every other node using the bus. The bus utility ensures that no data is sent to the wrong node. But, the bus traffic is an important parameter which can affect the system. RingThis is one of the simplest ways of connecting nodes with each other. The nodes are connected with each other to form a ring. For a node to communicate with some other node, it has to send the messages to its neighbor. Therefore, the data message passes through a series of other nodes before reaching the destination. This involves increased latency in the system. TreeThis topology involves connection of the nodes to form a tree. The nodes are connected to form clusters and the clusters are in-turn connected to form the tree. This methodology causes increased complexity in the network. Hypercube This topology consists of connections of the nodes to form cubes. The nodes are also connected to the nodes on the other cubes. ButterflyThis is one of the most complex connections of the nodes. As the figure suggests, there are nodes which are connected and arranged in terms of their ranks. They are arranged in the form of a matrix. In dynamic interconnect networks, the nodes are interconnected via an array of simple switching elements. This interconnection can then be changed by use of routing algorithms, such that the path from one node to other nodes can be varied. Dynamic interconnections can be classified as: Single stage Interconnect Network Multistage interconnect Network Crossbar switch connections == Crossbar Switch Connections == In crossbar switch, there is a dedicated path from one processor to other processors. Thus, if there are n inputs and m outputs, we will need nm switches to realize a crossbar. As the number of outputs increases, the number of switches increases by factor of n. For large network this will be a problem. An alternative to this scheme is staged switching. == Single Stage Interconnect Network == In a single stage interconnect network, the input nodes are connected to output via a single stage of switches. The figure shows 88 single stage switch using shuffle exchange. As one can see, from a single shuffle, not all input can reach all output. Multiple shuffles are required for all inputs to be connected to all the outputs. == Multistage Interconnect Network == A multistage interconnect network is formed by cascading multiple single stage switches. The switches can then use their own routing algorithm, or be controlled by a centralized router, to form a completely interconnected network. Multistage Interconnect Network can be classified into three types: Non-blocking: A non-blocking network can connect any idle input to any idle output, regardless of the connections already established across the network. Crossbar is an example of this type of network. Rearrangeable non-blocking: This type of network can establish all possible connections between inputs and outputs by rearranging its existing connections. Blocking: This type of network cannot realize all possible connections between inputs and outputs. This is because a connection between one free input to another free output is blocked by an existing connection in the network. The number of switching elements required to realize a non-blocking network in highest, followed by rearrangeable non-blocking. Blocking network uses least switching elements. == Examples == Multiple types of multistage interconnection networks exist. === Omega network === An Omega network consists of multiple stages of 22 switching elements. Each input has a dedicated connection to an output. An NN omega network has log2(N) stages and N/2 switching elements in each stage for a perfect shuffle between stages. Thus the network has complexity of 0(N log(N)). Each switching element can employ its own switching algorithm. Consider an 88 omega network. There are 8! = 40320 1-to-1 mappings from input to output. There are 12 switching element for a total permutation of 2^12 = 4096. Thus, it is a blocking network. === Clos network === A Clos network uses 3 stages to switch from N inputs to N outputs. In the first stage, there are r= N/n crossbar switches and each switch is of size nm. In the second stage there are m switches of size rr and finally the last stage is a mirror of the first stage with r switches of size mn. A clos network will be completely non-blocking if m >= 2n-1. The number of connections, though more than omega network is much less than that of a crossbar network. === Beneš network === A Beneš network is a rearrangeably non-blocking network derived from the clos network by initializing n = m = 2. There are (2log2(N) - 1) stages, with each stage containing N/2 22 crossbar switches. An 88 Beneš network has 5 stages of switching elements, and each stage has 4 switching elements. The center three stages has two 44 benes network. The 44 Beneš network, can connect any input to any output recursively.