AI/ML News & Innovations Hub

Datasets input at model.fit produce unexpected results of training loss vs validation loss

Im trying to train a neural network (VAE) using tensorflow and Im getting different results based on the type of input in the model.fit. When I input arrays I get normal difference between the validation loss and the total loss. When I input a dataset based on the same input I get a normal total loss and a really small validation loss. I havent changed the model. The only things that changes is the input format. The code for when I input an array. train slices is (2627,138,138,1) and define the batch size in the model.fit train_slices = preprocess_data(CropTumor, file_array[train_dataset]) val_slices = preprocess_data(CropTumor, file_array[val_dataset]) # reset model weights before training VAE.set_weights(initial_weights) # fit model fit_results = VAE.fit(train_slices,train_slices, epochs=1000, validation_data=(val_slices,val_slices), callbacks=[early_stopping_kfold, tensorboard_callback], batch_size=batch_sz, verbose=2 ) The output Epoch 1/1000 2022-08-01 11:56:35.683852: I tensorflow/stream_executor/cuda/cuda_dnn.cc:384] Loaded cuDNN version 8401 2022-08-01 11:56:36.371780: I tensorflow/core/platform/default/subprocess.cc:304] Start cannot spawn child process: No such file or directory 2022-08-01 11:56:36.461054: I tensorflow/stream_executor/cuda/cuda_blas.cc:1786] TensorFloat-32 will be used for the matrix multiplication. This will only be logged once. 672/672 - 7s - loss: 537.2896 - val_loss: 213.7070 - 7s/epoch - 11ms/step Epoch 2/1000 672/672 - 5s - loss: 248.5211 - v…

How to create KM curves with time-dependent covariate

Some patient received an treatment as treated group and some did not as control group. In order to consider "immortal time bias" , I coded the treatment as a time-dependent covariate. Could you let me know it is right or not? Thanks. For example, there are two patients: P1 received the treatment 30 days after the start of this study and died 30 days after receiving the treatment. P2 did not received the treatment and survived at the end of this study (80 days). Here I code the data as follow: ID Treatment Time Death P1 0 30 0 P1 1 30 1 P2 0 80 0 Then I just used SAS code to create KM curves /*Survival curve*/ proc lifetest data=treatment; time Time*Dearg(0); strata Treatment; run;

Statistical analysis of a bioassay

I am analyzing data on some biossays and I need an advice on which statistical analysis to use and how to use ggplot2 R package to better visualize the results. As you can see in the above image I have infected bean plants with a pathogen, a Pseudomonas bacterium, and then tested the treatment with bacteriophages applied at the same time or 1h before the inoculation. In the scatter plot on the right you can see two repetitions of the four I have. In each plot on the y axis there is the disease index of the plants (from healthy plants designed by 0 to deceased plants, 5). I wanted to test if there is a significant difference between the four cases (positive and negative control and the two phage applications) and between all the possible pairs. I tried with Kruskal-Wallis test and Wilcoxon paired test and I found out that there are significant differences only in two of the four repetition (the two represented). Now the doubts I have: 1) Are there better suited statistical tests I can conduct? 2) There is a better way to visualize dthe data (I applied the letters directly on the image with power point, but I would prefer to do all by an R script) I add here the script I used to analyze the data. # load libraries library(tidyverse) library(patchwork) library(broom) library(agricolae) library(ggplot2) library (reshape2) #Clear the Worskspace rm(list=ls()) # Read data DF % filter(date=="23/11/2020") DF2 % filter(date=="18/02/2022") DF3 % filter(date=="22/03/2022") DF4 % filter(d…

AAAI Announces New President-Elect and New Executive Council Members for 2022

AAAI is pleased to announce that Stephen Smith has been elected to serve as the next President-Elect. He will serve in that role for two years, and will then serve as President for two years. AAAI is also pleased to announce four incoming Councilors elected to three-year terms. The post AAAI Announces New President-Elect and New Executive Council Members for 2022 appeared first on AAAI .

Correlation linear and non-linear

Correlation between two variables can be linear or non-linear. Accordingly different methods are used to find correlation coefficients. My query is if the correlation can be linear or non-linear, why in most places/articles the definition of the correlation assumes that the association is linear?

Three versions of the independent two sample t-test (and R)

This post concerns three versions of the independent two sample t-test: Student's t-test uses a pooled standard deviation in the denominator (all equations are shamelessly copied from Wikipedia - Student's t-test ): $$ {\displaystyle t={\frac {{\bar {X}}_{1}-{\bar {X}}_{2}}{s_{p}\cdot {\sqrt {{\frac {1}{n_{1}}}+{\frac {1}{n_{2}}}}}}}} $$ where: $$ {\displaystyle s_{p}={\sqrt {\frac {\left(n_{1}-1\right)s_{X_{1}}^{2}+\left(n_{2}-1\right)s_{X_{2}}^{2}}{n_{1}+n_{2}-2}}}} $$ and: $$d.f. = n_1 + n_2 -2 $$ It is well known that Student's t-test can be seriously biased for unequal sample sizes and variances (Welch 1938 demonstrated this analytically, before computers and the many published simulation studies on this subject). Next is "an alternative criterion that has often been employed," which is less biased for unequal sample sizes and variances than is Student's t (Welch 1938). This version uses the standard error of the difference in the denominator: $$ {\displaystyle t={\frac {{\bar {X}}_{1}-{\bar {X}}_{2}}{s_{\bar {\Delta }}}}} $$ where: $$ {\displaystyle s_{\bar {\Delta }}={\sqrt {{\frac {s_{1}^{2}}{n_{1}}}+{\frac {s_{2}^{2}}{n_{2}}}}}.} $$ Note that if $ n_1 = n_2 $ or $ s_{X_{1}}^{2} = s_{X_{2}}^{2} $ then Student's t-test is identical to this alternative (I'll let the reader do the algebra). Welch (1938) presented an approximate degrees of freedom for the above statistic that further reduces bias (see Wikipedia - Welch's t-test for the formula). Edit: I replaced the exam…

Under what additional condition is "sum of coefficients less than unity in magnitude" also sufficient for stationarity?

It is well established that a necessary condition for stationarity of both $AR$ and $ARMA$ processes is that the coefficients of the autoregressive components sum to less than unity in magnitude. Do we know any conditions that if taken in addition to the above suffice for guaranteeing stationarity? Of course, only non-trivial conditions are of interest such that sufficiency only holds under the combination of both. If there are any such conditions, which of them is the least restrictive?

Cross Validation and hyperparameter selection correct procedure

I am trying to run a regression supervised learning problem. The dataset is not very large and I wanted to do some cross-validation to avoid overfitting. As I have read it's important to do a sensitivity analysis to determine the value of k. Also, I would like to do some hyperparameter grid search for the algorithm (i.e. random forests). What would be the correct procedure? First take a random value of k and perform the hyperparameter grid search and with the correct hyperparameters do the sensitivity test for k or vice versa? Thanks in advance!

How to handle anomaly detections with multiple different timeseries' from network traffic?

I would like to implement an anomaly detection algorithm on multiple timeseries' from different network users. Since each user has different behavior and network traffic usage, my question is how can I implement an anomaly detection algorithm for this case? If possible I would like to have the model to be trained on online data, meaning when new data arrives it should be able to use that data, so that I dont need to train it over and over again. When dealing with new users, it should consider other users as reference and not immediately trigger an anomaly. I was thinking about training an ensemble model of LSTMs with different temporal properties such as sequences of minutes, hours, days, weeks, months in order to predict successfully short-term anomalys and long-term occurences. Does anyone else had the same problem in the past?

Calculating absolute risk from odds ratio, relative risk, or hazard ratio

Say I have a known cumulative probability distribution for the probability (absolute risk) of developing breast cancer vs. age in a group representing the general population. I also know the odds ratio, relative risk, and hazard ratio for the development of breast cancer in a population that carries a cancer susceptibility gene mutation (ATM, for example). How would I calculate the cumulative probability distribution for the development of breast cancer vs. age in ATM mutation carriers? I appreciate your help!

compute Dirichlet distribution parameter from known mean distribution

For a particular Bayesian study I am going to apply Dirichlet distribution as my proposal random number generator. I am going to update the distribution parameter every trial based on a given calculated mean of distribution. So, I would like to calculate the parameter, alpha for a multivariate Dirichlet distribution given that I have the mean of marginal distributions. consider a three dimensional Dirichlet distribution with the following means: $mu = (mu_1, mu_2, mu_3)$ given that: $$\mu_i= \frac{a_i}{\sum{a_j}}$$ for a three dimensional distribution, I will have: $$\alpha_1(1-\mu_1)-\mu_1\alpha_2-\mu_1\alpha_3=0$$ $$\alpha_2(1-\mu_2)-\mu_2\alpha_1-\mu_2\alpha_3=0$$ $$\alpha_3(1-\mu_3)-\mu_3\alpha_1-\mu_3\alpha_2=0$$ $$\begin{pmatrix}\alpha_1 \\\ \alpha_2 \\\ \alpha_3\end{pmatrix} . \begin{pmatrix}\ 1-\mu_1 & -\mu_1 & -\mu_1 \\\ 1-\mu_2 & -\mu_2 & -\mu_2 \\\ 1-\mu_3 & -\mu_3 & -\mu_3 \end{pmatrix} = 0$$ Obviously, the coefficients matrix, here the matrix containing $\mu_i$ , is singular. Any help to address this problem is highly appreciated. Thanks Rezgar

Why are $R^2$ values not available in mixed effects meta regressions

I'm currently conducting a meta-regression for my predictors from a meta-analysis of proportion. In calculating the mixed effects meta-regressions, I wanted to identify how much heterogeneity is accounted for when a predictor is considered in the model. To do this, I used the meta package with a metareg function in R: m.prop.reg The output provides out $I^2$ and $H^2$ values as well as the significance of our moderators, however, it does not provide an $R^2$ value for the amount of heterogeneity accounted for. My question is how can I get $R^2$ for my predictors.

Mathematics books for reinforcement learning

This question is not about the math prerequisites of reinforcement learning, but about the textbooks of mathematics that are enough to understand the literature on reinforcement learning. What are the mathematics books that are recommended to study in order to understand the majority of the reinforcement learning literature?

Best point forecast for MAPE

I have a data that can have different distribution. For example every time I get a set of different data : 1, 2, 3, 4. I use metric MAPE and try to find best point to minimize my MAPE. Is there a universal formula that can take my set of data everytime and after that give me best point for prediction for each set of data? If it would be MSE I would take mean (2,5 for example set) of data for each set, if i would use MAE I would take median of each set but what should I take for each set if I use MAPE ?

Comparing efficiency between estimators

Suppose that $\hat \theta_1, \hat \theta_2$ are two estimators of $\theta$ . Furthermore, assume that \begin{align} \sqrt{n}(\hat \theta_1-\theta)\overset{d}{\to}N(0,V_1)\\ \sqrt{n}(\hat \theta_2-\theta-B)\overset{d}{\to}N(0,V_2), \end{align} where $V_{1}=\lim_{n\to\infty} V_{1,n}$ , $V_{2}=\lim_{n\to\infty} V_{2,n}$ and $B:=B_n$ are all known (including $V_{1,n},V_{2,n}$ ). Does it make sense to talk about relative efficiency between $\hat \theta_1$ and $\hat \theta_2$ by comparing $V_1$ and $V_2$ (or even $V_{1,n}$ and $V_{2,n}$ ) in the presence of term $B$ ? For example, if I show that $V_{1,n}/n-V_{2,n}/n-B , then can I say that $\hat\theta_1$ is relatively more efficient that $\hat\theta_2$ ? *Based on the answer and a comment, I clarify that the term $B$ is relevant in the expression for $\hat \theta_2$ . Hence it depends on $n$ . $B$ itself is $o(1), n\to\infty$ .

How to measure the "correlation" between two sequences of objects ordered by different criteria, considering higher positions as higher weights?

I suppose this is a pretty common problem since when we use ranks we are often more interested in how similar they are at the top (or at the bottom, or both extremes) than in how they compare in the middle. Example of two ordered objects according to two different external criteria (I am writing them as adjacent letters for simplicity, which doesn't imply any previous relation like a The pair of sequences abcdefghijk - abcdefkjihg would be closer than abcdefghijk - defghijkabc because the first 3 objects have more weight than the last 5. The pair of sequences abcdefghijk - defghijkabc would be closer than abcdefghijk - defghicbajk because despite b and c being closer to the top, cba is the oppposite of abc . This second example might show a goal that is not feasible to conciliate with the first example in an objective way, so each example probably belong to two different measures, which would not be a problem in my case. Kendall's tau seems adequate for a non-weighted comparison, please correct if I am wrong and would need another dissimilarity or rank correlation measure for the non-weighted case. Besides that, here I am specifically interested in giving the highest weight to the first position and the lowest for the last position, and respective intermediate weights for the other positions. This answer presents a review with several measures, but I am not sure they provide the kind of measure I need. ps. I am looking for something I can reference, i.e. some well stablished…

Bootstrapping as a way to deal with multilevel data (similar to Bayesian hierarchical models)

Suppose you want to estimate the linear relationship between $x$ & $y$ among people. You can't afford to ask N = 100 people to measure the variables of interest and estimate the relationship, so instead you ask N = 20 people and you take repeated measurements from each individual (100 observation per person which makes a total of 2000 data points). But now your observations are not independent and are grouped with respect to individuals. You can run a linear regression for each individual separately and the relationship can be expressed as: $$y = intercept_{j} + \beta_{j} \times x$$ $$j = 1,...,20$$ So you will have 20 different relationships ( $\beta s$ ) which are individual specific but you are interested in the $\beta_{population}$ or the relationship in the population. Having a set of $\beta$ s ( $[\beta_{1}, ...,\beta_{20}]$ ) I would take bootstrap samples from them (let's say 100 bootstrapped sample of the same size) and calculate their means to estimate the distribution of the $\beta_{population}$ (assuming $\beta_{j} \sim N(\beta_{population}, \sigma_{pop}^2)$ ) with its confidence intervals. Would this approach be reasonable or its not valid in ways that I can't see?

What should be taken as random variables in the distributions of datasets?

Consider the following two paragraphs taken from the paper titles Generative Adversarial Nets by Ian J. Goodfellow et.al #1: Abstract We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution , and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1....... #2: Excerpt from Introduction The promise of deep learning is to discover rich, hierarchical models that represent probability distributions over the kinds of data encountered in artificial intelligence applications, such as natural images, audio waveforms containing speech, and symbols in natural language corpora. here we saw the word distribution thrice. It is very common to encounter the phrase data distribution in machine learning papers. For the types of data we use in artificial intelligence, there can be infinite random samples. We collect some instances and form a dataset, based on which we try to get the data distribution. Even most of the literature uses the phrases data distribution or probability distribution . I personally never came across papers t…

Different MSE for each observation for SARIMA-model

I am doing a time series forecast with the SARIMA-model in STATA. I wanted to calculate the MSE to set up confidence intervals for my prediction. I have done that using the following command: predict MSE, mse dynamic(tm(2022m5)) STATA then generates a new variable of course with the MSE in it. Now I am wondering why the MSE is not similar for each observation. It is about roughly the same value for all observations of one year with some exceptions. It heavily differs for different years though. For me (if i have not fundamentally misunderstood the concept of MSE) it just does not make sense to have different values for the MSE for each observation. I would apreciate any help.

Does the P value need to be "back-transformed" after logging one variable in regression?

I have a data series, sediment concentration in water versus time. The data is not normal but I want to use regression and so have logged the sediment concentration ( $\log_{10}(x)$ ). The residual and fit to model is now much better. I want to use the p value and r squared values from the logged regression output. Do these need to be transformed back using $e^x$ ? This is being carried out in Excel using the data analysis tool:

What does "All store and access operations (for S(t) , A(t), and R(t)) can take their index mod n + 1" mean?

It's from the book Introduction to Reinforcement Learning. Second edition, chapter7: n-step Bootstrapping, page 147, n-step Sarsa. I made the algo work, but I still don't understand the phrase. Preferably explained in Python terms. The introductory part for the algo: Initialize Q(s, a) arbitrarily, for all s $\in$ S, a $\in$ A Initialize $\pi$ to be $\epsilon$ -greedy with respect to Q, or to a fixed given policy Algorithm parameters: step size $\alpha$ $\in$ (0, 1], small $\epsilon$ > 0, a positive integer n All store and access operations (for S(t), A(t), and R(t)) can take their index mod n + 1

How to predict time signal based on multi-input signals?

I would like to approximate the following relation by a neural network $y = \mathcal{f}(x_1(t),x_2(t))$ Here, I have only one output variable that is a function of 2 other variables which vary in time. Now, I want to be able to predict $y$ , given any shape of the 2 independent variables in time. For this reason, I have a training set corresponding to different input and output signals. However, I don't know how to make the neural network understand the concept of time which is very important since I expect the solution at time $t_k$ to be influenced by the previous instants in time. For this reason, I added as input variable the time derivative as $y = \mathcal{f}\left(x_1(t),x_2(t), \dfrac{\partial x_1 (t)}{\partial t}\right)$ This solution seems to work quite well for the fully connected neural network that I'm using. However, I would like to know if there are other ways to treat such problems where the time history is important.

#LearnAI Stories – Barnabas Chua

In-batch negative training Improves the results

I have read Dense passage retrieval for Open Domain Question Answering , and in page 6 it talks about in-batch negative training, it states the following: We find that using a similar configuration (7 gold negative passages), in-batch negative training improves the results substantially. The key difference between the two is whether the gold negative passages come from the same batch or from the whole training set. Effectively, in-batch negative training is an easy and memory-efficient way to reuse the negative examples already in the batch rather than creating new ones. It produces more pairs and thus increases the number of train- ing examples, which might contribute to the good model performance. As a result, accuracy consis- tently improves as the batch size grows. I know that it is more efficient and hence we can increase the number of negatives by increasing the batch size (or we can use other techniques if we cannot do that), but in the paper they have the following table: How having the same number of negatives (3rd & 4th rows) gave different results? Is that something realted to the gradient or something, since we sample from the same examples that contribute to calculating the gradient in one step?

Time Series clustering: Changing warping window for Dynamic time warping

I'm working on the same type of data and i want to classify the times series to find clear pattern of use. My data is collected from clients of a telecom company, and we want to detect pattern of the amount of data consumed by clients with their wifi box. So each client have a time series of how much data he consumed each 6minutes (I resampled it to hours). I also applied DTW with KmeansTimeSeries using tslearn : km = TimeSeriesKMeans(n_clusters = cluster_count, metric="dtw", verbose=1) labels = km.fit_predict(mySeries) My question is i want to change the warping window for DTW in python, and i'm pretty sure it's easy think to do, but i just coudn't find a way to do it. I also want to find the best window parameter for my case.

R: Johansen test for two variables

I am trying to replicate work from this paper , specifically examination of the rationality of inflation and inflation expectations when both series are non-stationary I(1). I need to apply Johansen cointegration method to test whether series of inflation and inflation expectation are cointegrated with a vector (1,-1). However, bh5lrtest does not work no matter how I set 'r' argument (I believe it should be equal to 1). Where is the problem? Is there any alternative regarding the code? data When r = 1, I get following error: Error in h(simpleError(msg, call)) : error in evaluating the argument 'object' in selecting a method for function 'summary': Row number of 'H' is unequal to VAR order. However, for any r greater or equal to 2, I get this: Error in h(simpleError(msg, call)) : error in evaluating the argument 'object' in selecting a method for function 'summary': Count of cointegrating relationships is out of allowable range.

What consequence would a polynomial time algorithm for SAT have on AGI?

$P$ vs $NP$ is a famous problem. We generally believe $P\neq NP$ . However suppose there is a polynomial time algorithm of order say $O((n+m)^2)$ or $O((n+m)^3)$ (a low degree polynomial complexity with small hidden constants) for $n$ variable SAT problem in $m$ clauses, then what consequence would it have on $AI$ and machine learning? Would $AGI$ be any closer?

Determine if high dimensional data is multimodal

I have p-dimensional data and I need to determine if that data has significant modes or if it’s clustered in any way. Here p=50, (dense embedding), we have n samples and p What are some ways to do that? I can think of creating sample covarince matrix of 50*50. Not sure how to analyze that full matrix (eigenvalues, etc) for conclusions. I can also try to do kmeans and see if it converges or do GMM but I needs to try on many number of clusters. The data could have many clusters in theory, possibly residing on lower dim manifolds. Any thoughts? Thanks.

Mechanical Relationship between two variables

Recently I attended a conference in which one of the authors was talking about the mechanical relationship between variables "X" and "Y" (I don't remember exactly the variables) and possible sources of endogeneity. I searched a lot on the net for the meaning of the mechanical relationship between two variables, but I found nothing. Can anybody help me out in this regard? It is much appreciated.

Correctly evaluating (different) time series models

Currently, I am going through Hyndman's book to learn more about time series (I use the 2nd version with the forecast package). I have trouble understanding the "correct" way of evaluating multiple (let's say 5 different) time series models. I can think of 3 different ways, and I am not sure which is the correct one: Split data into train/test => Fit all models on the training set and evaluate the models on the test set. Even though this would give us an indication of which model is the best, it will estimate the test error too small since we made a model choice based on it. Use time series cross validation as described in that part . This helps us finding the best model based on the lowest CV error. But after retraining the "best" model on the whole training data we don't have a test set left. Split the data into train/test => Fit all models on train and use time series CV on train => After finding the best model, retrain on the best model on train data set and evaluate it on the test data. In my opinion, the last approach would be the correct one if our goal is to find the best model and also get a reasonable error on a test set. Is my intuition correct, or am I missing something here? I don't know if I think too complicated? I have a second question that is closely related to this one: Is it possible to use the tsCV function from forecast on a combination of forecast models/hybrid model (e.g. the average forecast of an ARIMA, ETS, other models). In other words, the same a…

bimodal outcome - non normally distributed residuals

I have an outcome variable that is bimodal, this is because in about half the sample is measured from 0 to 5, and half the time from 0 to 7. Because of the different scales, I have decided to normalize this variable, but the normalization is particular; I discuss it below. In my study, there are two people who are assessed by three people, so there are six assessments; I rank these six assessments (1st, 2nd, 3rd..., 6th), subtract 1 so that the rank is 0th, 1st, ...5th, and divide by 5. In this way, this transformed variable goes from 0 to 1. The sample is composed of a few hundred of such six assessments (i.e. this gives about 1k assessments in total). As mentioned above about half of the times the three assessors asses the two assessed people on a scale from 0 to 5, while other times they assess them on a scale from 0 to 7. Within the triplet of assessors there is no scale variation, and what scale is going to be used within a triplet of assessors is random. I run an OLS with this transformed outcome on multiple explanatory variables, and cluster the standard errors at group and assessor level. If I run an OLS on the original variable, without normalization, the residuals are not normally distributed. Moreover, any other transformation of the outcome (i.e. natural logarithm; standardization at group level; unitization at group level, with zero minimum), with clustered standard errors at group and assessor level, do not solve the non-normality of the residuals. Have you hea…

How do you estimate the magnitude of seasonality in time-series data?

I have some time-series data showing the monthly counts of hospital admissions. It has both a long-term trend (increasing) and seasonality (highest in summer). I am trying to measure the magnitude of the seasonality (i.e. how much greater is the rate of admission in summer than in winter?). Here is some sample data (this is R code): ts We can fit Poisson models with and without seasonality, and plot the predicted values: model1 $pred1 pred2 Clearly there is strong evidence of seasonality: anova(model1, model2, test = 'LRT') Analysis of Deviance Table Model 1: N ~ period Model 2: N ~ period + month Resid. Df Resid. Dev Df Deviance Pr(>Chi) 1 58 202.753 2 47 62.334 11 140.42 However, how do we measure the magnitude of this seasonality (rather than the statistical evidence for seasonality or goodness-of-fit of the model)? My first thought was to calculate the extent to which residuals are reduced by adding a term for seasonality into the model, in comparison to the data. # calculate residuals: ts $resid1 pred1 - ts $N ts$ resid2 $pred2 - ts$ N # calculate the reduction in residuals in comparison to the data (sum(abs(ts $resid1)) - sum(abs(ts$ resid2))) / sum(ts$N) This suggests there is 2.7% seasonality in this data. Does this make sense? Is there a standard approach to this problem?

Survival analysis interaction interpretation: interaction is insignificant, but pairwise comparison does hold significance

I'm new with survival analyses, and have been breaking my head over this for a while, so I hope someone can explain this in a simple manner: Variable HR Gender(Female) 1.391* Treat (Yes) 0.544*** Gender*Treat 1.149 So, the interaction term is not significant, does this simply mean that the effect of treatment is not different for men and women? Or, does the interaction effect now show the effect of treatment for women, while Treat shows the effect for men, meaning that treat does nothing for women but does for men? Similarly, does the HR of gender now ONLY represent the effect of gender in the non-treatment group? I further get confused by the fact that the pairwise comparisons of these groups (malextreat, malexcontrol, femalextreat, femalexcontrol) dó hold significant differences. How does that work if the interaction is not significant?

Finding coefficients for a loss function with multiple parts and different scales

I'm working on reusing the paper Pixel2Mesh++: Multi-View 3D Mesh Generation via Deformation , which deforms a 3D mesh to make it fit some 2D images, in another scenario than the original one. The loss function is comprised of four terms : the chamfer loss, the cosine loss, a Laplacian regularization and an edge length regularization ( description here ). These 4 terms have very different scales and are somewhat unbound, meaning that I can't normalize all of them using a maximum value and that if I leave them as-is, some of them will be lost (lol) in the optimization. To add to the problem, not all parts of the loss are equally important : the chamfer loss, minimizing the distance between the expected and actual meshes, should for example have a higher weight in the optimization when compared to the regularization factors. In the github repository , some coefficients were found that may or may not be optimal for the problem at hand. I'm guessing they're not bad, considering that's what the authors decided to use. However, the method used to find these coefficients isn't explained anywhere (I even tried opening a github issue regarding this but I didn't get any answer). Now, I've read many questions here regarding optimization of the questions of multi-loss optimization and loss scaling, but I still don't know how to approach the problem of finding the optimal coefficients. I thought about using a hyperparameter optimizer like Hyperband to find the best coefficients, but what…

Dose McDiarmid's inequality holds in hilbert space,specially or in $L^2$ space?

Consider McDiarmid's inequality in Hilbert space, or can we extend McDiarmid's inequality to functional data analysis with the complete observed data?

Instrumental variable: Indirect effect of Z on Y

I have found an instrument variable ( $Z$ ) for my econometric model. The relevance constraint holds, however I still have a question about the instrument exogeneity and that is: It is said that it needs to be convincingly ruled out any direct effect of the instrument on the dependent variable or any effect running through omitted variables. But what if $Z$ influences another variable, $V,$ by which $Y$ is influenced ( $Z\to V\to Y$ )? Is then my exclusion restriction violated? I have data on $V,$ such that I could control for it.

Is there a state-of-the-art deep learning paper that uses center point regression instead of bounding box regression, for object tracking?

Almost all deep learning based object tracking methods perform bounding box regression. Siamese-based networks which are very popular for object tracking also perform bounding box regression most of the time, although SiamFC type exceptions exist. And some other networks use a detection + tracking mechanism using the center points of the objects, however, what I am looking for is whether any Siamese based networks that predict the center of the object rather than the bounding box exist. I searched the literature but couldn't find any, I would appreciate any guidance.

Inference Modeling for COVID Data

I am trying to build a model to analyze the relationship between COVID-19 mortality rate in each U.S. state or county (y) and independent variables (x) including: Vaccination rate: 1st, 2nd, booster COVID-19 policies: mask mandate, vaccine mandate, social distancing, quarantine mandate Amount of healthcare resources: number of doctor / 1k population, % GDP spent on health Need to control for: Demographics: sex, race, income, age Major health conditions that increase COVID-19 mortality: cardio-pulmonary diseases, diabetes, obesity, age I am struggling to select the appropriate model. I don't know how to account for the time component of the pandemic, i.e. the 4 COVID waves, in an inference model. I have thought of some modeling ideas, but they all seem to have some issues: Linear model Pick one point in time, e.g. peak of the 2nd wave, do a linear model between the dependent and independent variables. Observational unit: state or county. Can also pick one period of time (e.g. the entire 2nd COVID wave) and aggregate the predictors for each state/county. Issue: Linear models assume independence between data points but COVID mortality at a given time point is likely related to that from its preceding time point(s) Time series model Issue: Time series models are made for predictions, but I would like to make inferences. Causal Impact analysis: would enable me to estimate the impact of an intervention Issues: (1) I have multiple interventions/mandates that I want to assess, and (…

Can I use Sentence-Bert to embed event triples?

I extracted event triples from sentences using OpenIE. Can I concatenate the components in the event triple to make it a sentence and use Sentence-Bert to embed the event? It seems no one has done this way before so I am questioning my idea.

Generalized Visual Language Models

Processing images to generate text, such as image captioning and visual question-answering, has been studied for years. Traditionally such systems rely on an object detection network as a vision encoder to capture visual features and then produce text via a text decoder. Given a large amount of existing literature, in this post, I would like to only focus on one approach for solving vision language tasks, which is to extend pre-trained generalized language models to be capable of consuming visual signals .

Difference-in-differences (DiD) with 3 treatment states

I was wondering how I would adjust my model where there would be 3 treatment states. There is the untreated but the treated is split into longer treated or less treated. What I mean by this is that the treatment occurs in the 6th year and continued into the 7th year. But there is another type of treatment in which it was introduced in the 7th year. My original model was: This is showing that there were only two treatment states: treated and untreated. The data is for 7 years and of which the treatment occurs in the 6th and 7th years. I am wondering how I would change this so that there are now three treatment groups: untreated, longer treated, and less treated. Here, the treatment occurs in the 6th year and continues for some individuals. And, for other individuals, the treatment occurs in the 7th year. Currently, I wrote out: Where U is for the untreated, LO stands for being treated for longer, so in the 6th and 7th years, and then LE standard for being treated for less, so only in the 7th year. Would appreciate any help! Thank you

Embedding Quality of Transfer Learning model vs Contrastive learning model

I am working on Contrastive learning which is a technique to learn features based on the concept of learning from comparing two or more instances. The downstream task is a classification problem. Transfer Learning Due to limited data, I tried to use Transfer learning model trained on "Imagenet"(ResNet50 V2 "Deep Residual Learning for Image Recognition Kaiming He, et.al"). I used the embedding from the pretrained model and trained Linear SVM and achieved a F1 score of 0.84. Contrastive Learning I also trained a model for contrastive learning using Facenet technique("FaceNet: A Unified Embedding for Face Recognition and Clustering Florian Schroff, et.al") and further used the embedding for training a Linear SVM for classification problem. The achieved F1 score is 0.83. Problem Though the scores of both the concept are closeby, I tried to evaluate the quality of both the embeddings using Silhouette Coefficient. Overall Silhouette Coefficient: Transfer Learning Embedding = 0.05 Contrastive Learning Embedding = 0.49 I do not understand this behaviour of the system that even with lower Silhouette Coefficient, the transfer learning model is able to perform well. Kindly provide me with your views on it

LinkBERT: Improving Language Model Training with Document Link

Language Model Pretraining Language models (LMs), like BERT 1 and the GPT series 2 , achieve remarkable performance on many natural language processing (NLP) tasks. They are now the foundation of today’s NLP systems. 3 These models serve important roles in products and tools that we use every day, such as search engines like Google 4 and personal assistants like Alexa 5 . These LMs are powerful because they can be pretrained via self-supervised learning on massive amounts of text data on the web without the need for labels, after which the pretrained models can be quickly adapted to a wide range of new tasks without much task-specific finetuning. For instance, BERT is pretrained to predict randomly masked words in original text (masked language modeling), e.g. predicting the masked word “dog” from “My __ is fetching the ball”. GPTs are pretrained to predict the next word given a previous sequence of text (causal language modeling), e.g. predicting the next word “ball” from “My dog is fetching the”. In either cases, through pretraining, LMs learn to encode various knowledge from a text corpus that helps to perform downstream applications involving language understanding or generation. In particular, LMs can learn world knowledge (associations between concepts like “dog”, “fetch”, “ball”) from training text where the concepts appear together, and help for knowledge-intensive applications like question answering. 6 Challenges. A challenge with most common LM pretraining strateg…

Assess team win rate by combining individual win rates?

Suppose I have 5 players on each team in a game where each player selects a character (League of legends, Valorant, etc.). I am considering the overall win rates of the characters, and the unique win rate attached to the player is not considered (i.e some players focus on using X character, so their win rate may be higher with this specific character, I am ignoring this, I am using the win rate of a character across all games played in a given season). Each character has been played (in a game that has matches consisting of 2 teams formed by 5 players each, selected at random) anywhere from at least 4,000 to even 60,000 times in a given patch (lets just use season for simplicity). Upon entry, team A may have overall individual character win rates (calculated across all matches played from and up to a certain point) of [40%, 50%, 55%, 50%, 60%] and team B may have individual win rates of [50%, 45%, 60%, 40%, 65%] for each character (not player) respectively. Is it naïve to just average the win rates and consider that the team win rate, or is there a more robust method of doing this? Despite all the characters having 4000+ games, some may have only around 4000, some may have 8000, some may have even more depending on popularity, so I imagine there may need to be sort of weighting, but I am not sure if 4000+ games is enough to disregard the need for a weighting system (as opposed to a basketball player playing less than 20 games a season and then 70 games the next, clearly weig…

Posterior distribution with multiple parameters

The data are observations of, $x$ , the length of time in excess of $14$ days that it took for a random sample of $10$ grasshoppers infected with a fungal disease to die, where the times are measured from infection. It is believed that a gamma distribution may be a good model here, so that the p.d.f. of $x$ is $$ f(x) = \frac{\beta^\alpha}{(\alpha-1)!}x^{\alpha-1}e^{-x\beta}\quad x>0. $$ In this case the parameter $\beta$ could be any positive real number, but $\alpha$ can only take positive integer values. Show that a gamma prior for $\beta$ is conjugate. That is, show that if $\beta$ has a gamma prior, then the posterior for $\beta$ is also a gamma distribution. I looked at the solution and I do not understand how the did the very first line. The first line of the solution simply says $$ f(\beta|x)\propto f(x|\beta,\alpha)\pi(\beta).\tag{1} $$ I have tried to see why ( $1$ ) is true and I get nowhere. My first attempt was to do $$ \begin{align*} f(\beta|x)&=\int f(\alpha,\beta|x) d\alpha\\ &= \int \frac{f(x|\alpha,\beta)f(\alpha,\beta)}{f(x)} d\alpha\\ &\propto\int f(x|\alpha,\beta)f(\alpha,\beta) d\alpha\\ &=\int f(x|\alpha,\beta)f(\alpha)f(\beta) d\alpha\\ &=f(\beta)\int f(x|\alpha,\beta)f(\alpha) d\alpha. \end{align*} $$ This doesn't seem to help at all, so I am stuck on how we can get to $(1)$ . Any help would be much appreciated; I am very much struggling with this topic!

How to obtain the p-values ​of a gamlssMX model?

I am working with a dataset that includes a binary target variable (0 or 1). I have built a model with the gamlssMX() function included on the "gamlss.mx" package to explain a continuous target variable as a function of continuous and qualitative features, alongside with the target one. My goal is to find out if this pathway is significant, but when I build the gamlssMX() model, I get an error when applying the summary() function, which returns p-values associated with each of the features in other models, such as the linear model. If someone can help me I would appreciate it very much. Thank you very much!

Stanford AI Lab Papers and Talks at ACL 2022

The 60th Annual Meeting of the Association for Computational Linguistics (ACL) 2022 is taking place May 22nd - May 27th. We’re excited to share all the work from SAIL that’s being presented, and you’ll find links to papers, videos and blogs below. Feel free to reach out to the contact authors directly to learn more about the work that’s happening at Stanford! List of Accepted Papers LinkBERT: Pretraining Language Models with Document Links Authors : Michihiro Yasunaga, Jure Leskovec*, Percy Liang* Contact : myasu@cs.stanford.edu Links: Paper | Website Keywords : language model, pretraining, knowledge, hyperlink, bionlp When classifying grammatical role, BERT doesn’t care about word order… except when it matters Authors : Isabel Papadimitriou, Richard Futrell, Kyle Mahowald Contact : isabelvp@stanford.edu Links: Paper Keywords : large language models, analysis, word order, order invariance, grammatical role, syntax, semantics Problems with Cosine as a Measure of Embedding Similarity for High Frequency Words Authors : Kaitlyn Zhou, Kawin Ethayarajh, Dallas Card, Dan Jurafsky Contact : katezhou@stanford.edu Keywords : cosine similarity, training data frequency, model analysis Faithful or Extractive? On Mitigating the Faithfulness-Abstractiveness Trade-off in Abstractive Summarization Authors : Faisal Ladhak, Esin Durmus, He He, Claire Cardie, Kathleen McKeown Contact : esdurmus@stanford.edu Links: Paper Keywords : text summarization, text generation, evaluation, faithfulness Sp…

At which point, does the momentum based GD helps really in this figure?

Classical gradient descent algorithms sometimes overshoot and escape minima as they depend on the gradient only. You can see such a problem during the update from point 6. In classical GD algorithm, the update equation is $$\theta_{t+1} = \theta_{t} - \eta \times \triangledown_{\theta} \ell$$ In the momentum based GD algorithm, the update equations are $$v_0 = 0$$ $$v_{t+1} = \alpha v_t + \eta \times \triangledown_{\theta} \ell $$ $$\theta = \theta - \eta \times \triangledown_{\theta} \ell$$ I am writing all the equations concisely by removing the obvious variables used such as inputs to loss functions. In the lecture I'm listening to, the narrator says that momentum-based GD helps during the update at point 6 and the update will not lead to point 7 as shown in the figure and goes towards minima. But for me, it seems that even momentum-based GD will go to point 7 and the update at point 7 will be benefited from the momentum-based GD as it does not lead to point 8 and goes towards minima. Am I correct? If not, at which point does the momentum-based GD actually help?

Comment on Beyond multistakeholder tokenism: A provisional examination of participation in the IGF by babul

It is рerfect time to make somе plans for the futurе and it is time to be hapρy. I have learn tһiѕ submіt and if I may just I desire to suggest you few fascinating things or ѕuggestions. Mɑybe you could write next articles referring to this article. I want to learn more issues approximately it!

Making of AI4K Storybook, Daisy and her AI Friends