The Shapiro—Senapathy algorithm (S&S) is a computational method for identifying splice sites in eukaryotic genes. The algorithm employs a Position Weight Matrix (PWM) scoring formula to predict donor and acceptor splice sites in any given gene. This methodology has been used to discover splice sites and disease-causing splice site mutations in the human genome, and has become a standard tool in clinical genomics. The S&S algorithm has been cited in thousands of clinical studies, according to Google Scholar. It has also formed the basis of widely used software, including Human Splicing Finder, SROOGLE, and Alamut, which identify splice sites and splice site mutations that cause disease. The algorithm has uncovered splicing mutations in diseases ranging from cancers to inherited disorders, and predicted the deleterious effects of these mutations including exon skipping, intron retention, and cryptic splice site activation. == The algorithm == A splice site defines the boundary between a coding exon and a non-coding intron in eukaryotic genes. The S&S algorithm employs a sliding window, corresponding to the length of the splice site motif, to scan a gene sequence and detect potential splice sites. For each sliding window, the algorithm calculates a score by comparing the nucleotide sequence to a Position Weight Matrix (PWM) derived from known splice sites. This formula generates a percentile score, indicating the likelihood that a given sequence functions as a donor or acceptor splice site. The majority of disease-causing mutations in the human genome are located in splice sites. Clinical genomics studies analyze the splice site scores generated by the S&S algorithm to predict the consequences of splice site mutations including exon skipping and intron retention. The algorithm's sensitivity to single-nucleotide changes allows it to determine mutations that may impact RNA splicing and contribute to disease. In addition to identifying real splice sites, the S&S algorithm has been used to discover cryptic splice sites — alternative splice sites activated by mutations — which may disrupt normal splicing. The algorithm detects mutations that lead to the activation of cryptic splice sites, which may be located proximal to real splice sites or deep within non-coding introns. It has thus been used to determine the causes of numerous diseases that are due to cryptic splicing. == Cancer gene discovery using S&S == The S&S algorithm has been used to identify splice-site mutations in genes associated with several cancers. For example, genes causing commonly occurring cancers including breast cancer, ovarian cancer, colorectal cancer, leukemia, head and neck cancers, prostate cancer, retinoblastoma, squamous cell carcinoma, gastrointestinal cancer, melanoma, liver cancer, Lynch syndrome, skin cancer, and neurofibromatosis have been found. In addition, splicing mutations in genes causing less commonly known cancers including gastric cancer, gangliogliomas, Li-Fraumeni syndrome, Loeys–Dietz syndrome, Osteochondromas (bone tumor), Nevoid basal cell carcinoma syndrome, and Pheochromocytomas have been identified. Specific mutations in different splice sites in various genes causing breast cancer (e.g., BRCA1, PALB2), ovarian cancer (e.g., SLC9A3R1, COL7A1, HSD17B7), colon cancer (e.g., APC, MLH1, DPYD), colorectal cancer (e.g., COL3A1, APC, HLA-A), skin cancer (e.g., COL17A1, XPA, POLH), and Fanconi anemia (e.g., FANC, FANA) have been uncovered. The mutations in the donor and acceptor splice sites in different genes causing a variety of cancers that have been identified by S&S are shown in Table 1. == Discovery of genes causing inherited disorders using S&S == Specific mutations in different splice sites in various genes that cause inherited disorders, including, for example, Type 1 diabetes (e.g., PTPN22, TCF1 (HCF-1A)), hypertension (e.g., LDL, LDLR, LPL), Marfan syndrome (e.g., FBN1, TGFBR2, FBN2), cardiac diseases (e.g., COL1A2, MYBPC3, ACTC1), eye disorders (e.g., EVC, VSX1) have been uncovered. A few example mutations in the donor and acceptor splice sites in different genes causing a variety of inherited disorders identified using S&S are shown in Table 2. == Genes causing immune system disorders == More than 100 immune system disorders affect humans, including inflammatory bowel diseases, multiple sclerosis, systemic lupus erythematosus, bloom syndrome, familial cold autoinflammatory syndrome, and dyskeratosis congenita. The Shapiro–Senapathy algorithm has been used to discover genes and mutations involved in many immune disorder diseases, including Ataxia telangiectasia, B-cell defects, epidermolysis bullosa, and X-linked agammaglobulinemia. Xeroderma pigmentosum, an autosomal recessive disorder is caused by faulty proteins formed due to new preferred splice donor site identified using S&S algorithm and resulted in defective nucleotide excision repair. Type I Bartter syndrome (BS) is caused by mutations in the gene SLC12A1. S&S algorithm helped in disclosing the presence of two novel heterozygous mutations c.724 + 4A > G in intron 5 and c.2095delG in intron 16 leading to complete exon 5 skipping. Mutations in the MYH gene, which is responsible for removing the oxidatively damaged DNA lesion are cancer-susceptible in the individuals. The IVS1+5C plays a causative role in the activation of a cryptic splice donor site and the alternative splicing in intron 1, S&S algorithm shows, guanine (G) at the position of IVS+5 is well conserved (at the frequency of 84%) among primates. This also supported the fact that the G/C SNP in the conserved splice junction of the MYH gene causes the alternative splicing of intron 1 of the β type transcript. Splice site scores were calculated according to S&S to find EBV infection in X-linked lymphoproliferative disease. Identification of Familial tumoral calcinosis (FTC) is an autosomal recessive disorder characterized by ectopic calcifications and elevated serum phosphate levels and it is because of aberrant splicing. == Application of S&S in hospitals for clinical practice and research == The Shapiro–Senapathy (S&S) algorithm has played a significant role in advancing the diagnosis and treatment of human diseases through its application in modern clinical genomics. With the widespread adoption of next-generation sequencing (NGS) technologies, the S&S algorithm is now routinely integrated into clinical practice by geneticists and diagnostic laboratories. It is implemented in various computational tools such as Human Splicing Finder (HSF), Splice Site Finder (SSF), and Alamut Visual, which assist in interpreting the functional impact of genetic variants on RNA splicing. The algorithm is particularly useful in identifying pathogenic splice site mutations in cases where the clinical presentation is unclear or where conventional diagnostic methods have failed to identify a causative gene. Its utility has been demonstrated across diverse patient cohorts, including individuals from different ethnic backgrounds with various cancers and inherited genetic disorders. The following are selected examples illustrating its application in clinical research. === Cancers === === Inherited disorders === == S&S - Algorithm for identifying splice sites, exons and split genes == The Shapiro–Senapathy algorithm (SSA) was developed to identify splice sites in uncharacterized genomic sequences, with early applications in the Human Genome Project. The method introduced a Position Weight Matrix (PWM)-based approach to analyze splicing sequences across eukaryotic organisms, marking the first computational framework to systematically define splice sites using probabilistic scoring. Key innovations of the algorithm included: Exon Detection – Exons were defined as sequences bounded by acceptor and donor splice sites with S&S scores above a threshold, requiring an open reading frame (ORF) for validation. Gene Prediction – The method enabled the identification of complete genes by assembling predicted exons, forming a basis for later gene-finding tools. Mutation Analysis – The algorithm distinguishes deleterious splice-site mutations (which disrupt protein function by lowering S&S scores) from neutral variations. This capability allowed researchers to study disease-linked cryptic splice sites in humans, animals, and plants. SSA's PWM-based framework influenced subsequent computational methods, including machine learning and neural network approaches, for splice-site prediction and alternative splicing research. It remains a foundational tool in genomics and disease studies. == Discovering the mechanisms of aberrant splicing in diseases == The Shapiro–Senapathy algorithm has been used to determine the various aberrant splicing mechanisms in genes due to deleterious mutations in the splice sites, which cause numerous diseases. Deleterious splice site mutations impair the normal splicing of the gene transcripts, and thereby make the encoded protei
Programming tool
A programming tool or software development tool is a computer program that is used to develop another computer program, usually by helping the developer manage computer files. For example, a programmer may use a tool called a source code editor to edit source code files, and then a compiler to convert the source code into machine code files. They may also use build tools that automatically package executable program and data files into shareable packages or install kits. A set of tools that are run one after another, with each tool feeding its output to the next one, is called a toolchain. An integrated development environment (IDE) integrates the function of several tools into a single program. Usually, an IDE provides a source code editor as well as other built-in or plug-in tools that help with compiling, debugging, and testing. Whether a program is considered a development tool can be subjective. Some programs, such as the GNU compiler collection, are used exclusively for software development while others, such as Notepad, are not meant specifically for development but are nevertheless often used for programming. == Categories == Notable categories of development tools: Assembler – Converts assembly language into machine code Bug tracking system – Software application that records software bugs Build automation – Building software via an unattended fashion Code review software – Activity where one or more people check a program's code Compiler – Software that translates code from one programming language to another Compiler-compiler – Program that generates parsers or compilers, a.k.a. parser generator Debugger – Software for debugging a computer program Decompiler – Program translating executable to source code Disassembler – Computer program to translate machine language into assembly language Documentation generator – Automation technology for creating software documentation Graphical user interface builder – Software development tool Linker – Program that combines intermediate build files into an executable file Loader – Loads executable files into memory and prepares them for execution by the CPU. Memory debugger – Software memory problem finder Minifier – Removal of unnecessary characters in code without changing its functionality Pretty-printer – Formatting to make code or markup easier to readPages displaying short descriptions of redirect targets Performance profiler – Measuring the time or resources used by a section of a computer program Static code analyzer – Analysis of computer programs without executing themPages displaying short descriptions of redirect targets Source code editor – Text editor specializing in software codePages displaying short descriptions of redirect targets Source code generation – Type of computer programmingPages displaying short descriptions of redirect targets Version control system – Stores and tracks versions of files
Structured sparsity regularization
Structured sparsity regularization is a class of methods, and an area of research in statistical learning theory, that extend and generalize sparsity regularization learning methods. Both sparsity and structured sparsity regularization methods seek to exploit the assumption that the output variable Y {\displaystyle Y} (i.e., response, or dependent variable) to be learned can be described by a reduced number of variables in the input space X {\displaystyle X} (i.e., the domain, space of features or explanatory variables). Sparsity regularization methods focus on selecting the input variables that best describe the output. Structured sparsity regularization methods generalize and extend sparsity regularization methods, by allowing for optimal selection over structures like groups or networks of input variables in X {\displaystyle X} . Common motivation for the use of structured sparsity methods are model interpretability, high-dimensional learning (where dimensionality of X {\displaystyle X} may be higher than the number of observations n {\displaystyle n} ), and reduction of computational complexity. Moreover, structured sparsity methods allow to incorporate prior assumptions on the structure of the input variables, such as overlapping groups, non-overlapping groups, and acyclic graphs. Examples of uses of structured sparsity methods include face recognition, magnetic resonance image (MRI) processing, socio-linguistic analysis in natural language processing, and analysis of genetic expression in breast cancer. == Definition and related concepts == === Sparsity regularization === Consider the linear kernel regularized empirical risk minimization problem with a loss function V ( y i , f ( x ) ) {\displaystyle V(y_{i},f(x))} and the ℓ 0 {\displaystyle \ell _{0}} "norm" as the regularization penalty: min w ∈ R d 1 n ∑ i = 1 n V ( y i , ⟨ w , x i ⟩ ) + λ ‖ w ‖ 0 , {\displaystyle \min _{w\in \mathbb {R} ^{d}}{\frac {1}{n}}\sum _{i=1}^{n}V(y_{i},\langle w,x_{i}\rangle )+\lambda \|w\|_{0},} where x , w ∈ R d {\displaystyle x,w\in \mathbb {R^{d}} } , and ‖ w ‖ 0 {\displaystyle \|w\|_{0}} denotes the ℓ 0 {\displaystyle \ell _{0}} "norm", defined as the number of nonzero entries of the vector w {\displaystyle w} . f ( x ) = ⟨ w , x i ⟩ {\displaystyle f(x)=\langle w,x_{i}\rangle } is said to be sparse if ‖ w ‖ 0 = s < d {\displaystyle \|w\|_{0}=s
Learning curve (machine learning)
In machine learning (ML), a learning curve (or training curve) is a graphical representation that shows how a model's performance on a training set (and usually a validation set) changes with the number of training iterations (epochs) or the amount of training data. Typically, the number of training epochs or training set size is plotted on the x-axis, and the value of the loss function (and possibly some other metric such as the cross-validation score) on the y-axis. Synonyms include error curve, experience curve, improvement curve and generalization curve. More abstractly, learning curves plot the difference between learning effort and predictive performance, where "learning effort" usually means the number of training samples, and "predictive performance" means accuracy on testing samples. Learning curves have many useful purposes in ML, including: choosing model parameters during design, adjusting optimization to improve convergence, and diagnosing problems such as overfitting (or underfitting). Learning curves can also be tools for determining how much a model benefits from adding more training data, and whether the model suffers more from a variance error or a bias error. If both the validation score and the training score converge to a certain value, then the model will no longer significantly benefit from more training data. == Formal definition == When creating a function to approximate the distribution of some data, it is necessary to define a loss function L ( f θ ( X ) , Y ) {\displaystyle L(f_{\theta }(X),Y)} to measure how good the model output is (e.g., accuracy for classification tasks or mean squared error for regression). We then define an optimization process which finds model parameters θ {\displaystyle \theta } such that L ( f θ ( X ) , Y ) {\displaystyle L(f_{\theta }(X),Y)} is minimized, referred to as θ ∗ {\displaystyle \theta ^{}} . === Training curve for amount of data === If the training data is { x 1 , x 2 , … , x n } , { y 1 , y 2 , … y n } {\displaystyle \{x_{1},x_{2},\dots ,x_{n}\},\{y_{1},y_{2},\dots y_{n}\}} and the validation data is { x 1 ′ , x 2 ′ , … x m ′ } , { y 1 ′ , y 2 ′ , … y m ′ } {\displaystyle \{x_{1}',x_{2}',\dots x_{m}'\},\{y_{1}',y_{2}',\dots y_{m}'\}} , a learning curve is the plot of the two curves i ↦ L ( f θ ∗ ( X i , Y i ) ( X i ) , Y i ) {\displaystyle i\mapsto L(f_{\theta ^{}(X_{i},Y_{i})}(X_{i}),Y_{i})} i ↦ L ( f θ ∗ ( X i , Y i ) ( X i ′ ) , Y i ′ ) {\displaystyle i\mapsto L(f_{\theta ^{}(X_{i},Y_{i})}(X_{i}'),Y_{i}')} where X i = { x 1 , x 2 , … x i } {\displaystyle X_{i}=\{x_{1},x_{2},\dots x_{i}\}} === Training curve for number of iterations === Many optimization algorithms are iterative, repeating the same step (such as backpropagation) until the process converges to an optimal value. Gradient descent is one such algorithm. If θ i ∗ {\displaystyle \theta _{i}^{}} is the approximation of the optimal θ {\displaystyle \theta } after i {\displaystyle i} steps, a learning curve is the plot of i ↦ L ( f θ i ∗ ( X , Y ) ( X ) , Y ) {\displaystyle i\mapsto L(f_{\theta _{i}^{}(X,Y)}(X),Y)} i ↦ L ( f θ i ∗ ( X , Y ) ( X ′ ) , Y ′ ) {\displaystyle i\mapsto L(f_{\theta _{i}^{}(X,Y)}(X'),Y')}
Video Super Resolution
RTX Video Super Resolution (RTX VSR) is a video scaling feature by Nvidia. It was released on February 28, 2023. == History == The feature was first unveiled during CES 2023 as RTX Video Super Resolution. It uses the on-board Tensor Cores to upscale browser video content in real time. Video Super Resolution was initially only available on RTX 30 and 40 series GPUs, while support for 20 series GPUs was added afterwards; it is now available on all Nvidia RTX-branded GPUs. The feature supports input resolutions from 360p to 1440p and a max output of 4K and comes without support for HDR content although that could be likely added in the future. Nvidia released RTX Video Super Resolution 1.5 with improved video quality and RTX 20 series support on October 17, 2023. == Reception == According to ComputerBase, although "the algorithm is not yet working flawlessly", the feature is "overall recommendable".
Bag-of-words model
The bag-of-words (BoW) model is a model of text which uses an unordered collection (a "bag") of words. It is used in natural language processing and information retrieval (IR). It disregards word order (and thus most of syntax or grammar) but captures multiplicity. The bag-of-words model is commonly used in methods of document classification where, for example, the (frequency of) occurrence of each word is used as a feature for training a classifier. It has also been used for computer vision. An early reference to "bag of words" in a linguistic context can be found in Zellig Harris's 1954 article on Distributional Structure. == Definition == The following models a text document using bag-of-words. Here are two simple text documents: Based on these two text documents, a list is constructed as follows for each document: Representing each bag-of-words as a JSON object, and attributing to the respective JavaScript variable: Each key is the word, and each value is the number of occurrences of that word in the given text document. The order of elements is free, so, for example {"too":1,"Mary":1,"movies":2,"John":1,"watch":1,"likes":2,"to":1} is also equivalent to BoW1. It is also what we expect from a strict JSON object representation. Note: if another document is like a union of these two, its JavaScript representation will be: So, as we see in the bag algebra, the "union" of two documents in the bags-of-words representation is, formally, the disjoint union, summing the multiplicities of each element. === Word order === The BoW representation of a text removes all word ordering. For example, the BoW representation of "man bites dog" and "dog bites man" are the same, so any algorithm that operates with a BoW representation of text must treat them in the same way. Despite this lack of syntax or grammar, BoW representation is fast and may be sufficient for simple tasks that do not require word order. For instance, for document classification, if the words "stocks" "trade" "investors" appears multiple times, then the text is likely a financial report, even though it would be insufficient to distinguish between Yesterday, investors were rallying, but today, they are retreating.andYesterday, investors were retreating, but today, they are rallying.and so the BoW representation would be insufficient to determine the detailed meaning of the document. == Implementations == Implementations of the bag-of-words model might involve using frequencies of words in a document to represent its contents. The frequencies can be "normalized" by the inverse of document frequency, or tf–idf. Additionally, for the specific purpose of classification, supervised alternatives have been developed to account for the class label of a document. Lastly, binary (presence/absence or 1/0) weighting is used in place of frequencies for some problems (e.g., this option is implemented in the WEKA machine learning software system). == Hashing trick == A common alternative to using dictionaries is the hashing trick, where words are mapped directly to indices with a hash function. When using a hash function, no memory is required to store a dictionary. In practice, hashing simplifies the implementation of bag-of-words models and improves scalability. Collisions can occur when two words are hashed to the same index, but this happens infrequently and may function as a form of regularization.
Neural scaling law
In machine learning, a neural scaling law is an empirical scaling law that describes how neural network performance changes as key factors are scaled up or down. These factors typically include the number of parameters, training dataset size, and training cost. Some models also exhibit performance gains by scaling inference through increased test-time compute (TTC), extending neural scaling laws beyond training to the deployment phase. == Introduction == In general, a deep learning model can be characterized by four parameters: model size, training dataset size, training cost, and the post-training error rate (e.g., the test set error rate). Each of these variables can be defined as a real number, usually written as N , D , C , L {\displaystyle N,D,C,L} (respectively: parameter count, dataset size, computing cost, and loss). A neural scaling law is a theoretical or empirical statistical law between these parameters. There are also other parameters with other scaling laws. === Size of the model === In most cases, the model's size is simply the number of parameters. However, one complication arises with the use of sparse models, such as mixture-of-expert models. With sparse models, during inference, only a fraction of their parameters are used. In comparison, most other kinds of neural networks, such as transformer models, always use all their parameters during inference. === Size of the training dataset === The size of the training dataset is usually quantified by the number of data points within it. Larger training datasets are typically preferred, as they provide a richer and more diverse source of information from which the model can learn. This can lead to improved generalization performance when the model is applied to new, unseen data. However, increasing the size of the training dataset also increases the computational resources and time required for model training. With the "pretrain, then finetune" method used for most large language models, there are two kinds of training dataset: the pretraining dataset and the finetuning dataset. Their sizes have different effects on model performance. Generally, the finetuning dataset is less than 1% the size of pretraining dataset. In some cases, a small amount of high quality data suffices for finetuning, and more data does not necessarily improve performance. Many scaling laws, due to their inherent diminishing returns nature, value data based on a submodular set function which was shown in a paper on this topic. === Cost of training === Training cost is typically measured in terms of time (how long it takes to train the model) and computational resources (how much processing power and memory are required). It is important to note that the cost of training can be significantly reduced with efficient training algorithms, optimized software libraries, and parallel computing on specialized hardware such as GPUs or TPUs. The cost of training a neural network model is a function of several factors, including model size, training dataset size, the training algorithm complexity, and the computational resources available. In particular, doubling the training dataset size does not necessarily double the cost of training, because one may train the model for several times over the same dataset (each being an "epoch"). === Performance === The performance of a neural network model is evaluated based on its ability to accurately predict the output given some input data. Common metrics for evaluating model performance include: Negative log-likelihood per token (logarithm of perplexity) for language modeling; Accuracy, precision, recall, and F1 score for classification tasks; Mean squared error (MSE) or mean absolute error (MAE) for regression tasks; Elo rating in a competition against other models, such as gameplay or preference by a human judge. Performance can be improved by using more data, larger models, different training algorithms, regularizing the model to prevent overfitting, and early stopping using a validation set. When the performance is a number bounded within the range of [ 0 , 1 ] {\displaystyle [0,1]} , such as accuracy, precision, etc., it often scales as a sigmoid function of cost, as seen in the figures. == Examples == === (Hestness, Narang, et al, 2017) === The 2017 paper is a common reference point for neural scaling laws fitted by statistical analysis on experimental data. Previous works before the 2000s, as cited in the paper, were either theoretical or orders of magnitude smaller in scale. Whereas previous works generally found the scaling exponent to scale like L ∝ D − α {\displaystyle L\propto D^{-\alpha }} , with α ∈ { 0.5 , 1 , 2 } {\displaystyle \alpha \in \{0.5,1,2\}} , the paper found that α ∈ [ 0.07 , 0.35 ] {\displaystyle \alpha \in [0.07,0.35]} . Of the factors they varied, only task can change the exponent α {\displaystyle \alpha } . Changing the architecture optimizers, regularizers, and loss functions, would only change the proportionality factor, not the exponent. For example, for the same task, one architecture might have L = 1000 D − 0.3 {\displaystyle L=1000D^{-0.3}} while another might have L = 500 D − 0.3 {\displaystyle L=500D^{-0.3}} . They also found that for a given architecture, the number of parameters necessary to reach lowest levels of loss, given a fixed dataset size, grows like N ∝ D β {\displaystyle N\propto D^{\beta }} for another exponent β {\displaystyle \beta } . They studied machine translation with LSTM ( α ∼ 0.13 {\displaystyle \alpha \sim 0.13} ), generative language modelling with LSTM ( α ∈ [ 0.06 , 0.09 ] , β ≈ 0.7 {\displaystyle \alpha \in [0.06,0.09],\beta \approx 0.7} ), ImageNet classification with ResNet ( α ∈ [ 0.3 , 0.5 ] , β ≈ 0.6 {\displaystyle \alpha \in [0.3,0.5],\beta \approx 0.6} ), and speech recognition with two hybrid (LSTMs complemented by either CNNs or an attention decoder) architectures ( α ≈ 0.3 {\displaystyle \alpha \approx 0.3} ). === (Henighan, Kaplan, et al, 2020) === A 2020 analysis studied statistical relations between C , N , D , L {\displaystyle C,N,D,L} over a wide range of values and found similar scaling laws, over the range of N ∈ [ 10 3 , 10 9 ] {\displaystyle N\in [10^{3},10^{9}]} , C ∈ [ 10 12 , 10 21 ] {\displaystyle C\in [10^{12},10^{21}]} , and over multiple modalities (text, video, image, text to image, etc.). In particular, the scaling laws it found are (Table 1 of ): For each modality, they fixed one of the two C , N {\displaystyle C,N} , and varying the other one ( D {\displaystyle D} is varied along using D = C / 6 N {\displaystyle D=C/6N} ), the achievable test loss satisfies L = L 0 + ( x 0 x ) α {\displaystyle L=L_{0}+\left({\frac {x_{0}}{x}}\right)^{\alpha }} where x {\displaystyle x} is the varied variable, and L 0 , x 0 , α {\displaystyle L_{0},x_{0},\alpha } are parameters to be found by statistical fitting. The parameter α {\displaystyle \alpha } is the most important one. When N {\displaystyle N} is the varied variable, α {\displaystyle \alpha } ranges from 0.037 {\displaystyle 0.037} to 0.24 {\displaystyle 0.24} depending on the model modality. This corresponds to the α = 0.34 {\displaystyle \alpha =0.34} from the Chinchilla scaling paper. When C {\displaystyle C} is the varied variable, α {\displaystyle \alpha } ranges from 0.048 {\displaystyle 0.048} to 0.19 {\displaystyle 0.19} depending on the model modality. This corresponds to the β = 0.28 {\displaystyle \beta =0.28} from the Chinchilla scaling paper. Given fixed computing budget, optimal model parameter count is consistently around N o p t ( C ) = ( C 5 × 10 − 12 petaFLOP-day ) 0.7 = 9.0 × 10 − 7 C 0.7 {\displaystyle N_{opt}(C)=\left({\frac {C}{5\times 10^{-12}{\text{petaFLOP-day}}}}\right)^{0.7}=9.0\times 10^{-7}C^{0.7}} The parameter 9.0 × 10 − 7 {\displaystyle 9.0\times 10^{-7}} varies by a factor of up to 10 for different modalities. The exponent parameter 0.7 {\displaystyle 0.7} varies from 0.64 {\displaystyle 0.64} to 0.75 {\displaystyle 0.75} for different modalities. This exponent corresponds to the ≈ 0.5 {\displaystyle \approx 0.5} from the Chinchilla scaling paper. It's "strongly suggested" (but not statistically checked) that D o p t ( C ) ∝ N o p t ( C ) 0.4 ∝ C 0.28 {\displaystyle D_{opt}(C)\propto N_{opt}(C)^{0.4}\propto C^{0.28}} . This exponent corresponds to the ≈ 0.5 {\displaystyle \approx 0.5} from the Chinchilla scaling paper. The scaling law of L = L 0 + ( C 0 / C ) 0.048 {\displaystyle L=L_{0}+(C_{0}/C)^{0.048}} was confirmed during the training of GPT-3 (Figure 3.1 ). === Chinchilla scaling (Hoffmann, et al, 2022) === One particular scaling law ("Chinchilla scaling") states that, for a large language model (LLM) autoregressively trained for one epoch, with a cosine learning rate schedule, we have: { C = C 0 N D L = A N α + B D β + L 0 {\displaystyle {\begin{cases}C=C_{0}ND\\L={\frac {A}{N^{\alpha }}}+{\frac {B}{D^{\beta }}}+L_{0}\end{cases}}} where the variables are C {\displaystyle C} is the cost o