AI/ML News & Innovations Hub

Do all deep learning architectures learn features hierarchically?

CNNs are a famous example of hierarchical processing. Lower-level features such as edges are detected in the earlier layers while higher-level features such as the presence of a face are detected deeper in the network. All deep learning architectures share a common thing (excerpt taken from the Deep Learning book ): The modern term “deep learning” goes beyond the neuroscientific perspective on the current breed of machine learning models. It appeals to a more general principle of learning multiple levels of composition , which can be applied in machine learning frameworks that are not necessarily neurally inspired. My question is the following: Is learning multiple levels of composition equivalent to hierarchical processing/hierarchical feature extraction? If yes, what kind of hierarchies are extracted in fields other than computer vision? For example, what kind of features are extracted hierarchically from a transformer-based architecture or a graph neural network?

Congratulations to Stanford AI Lab PhD student Dora Zhao for an ICML 2024 Best Paper Award!

Congratulations to Stanford AI Lab PhD student Dora Zhao for an ICML 2024 Best Paper Award for a paper from her work at Sony AI on: Measure Dataset Diversity, Don’t Just Claim It

Congratulations to Aaron Lou, Chenlin Meng, and Stefano Ermon for an ICML 2024 Best Paper Award!

Congratulations to Aaron Lou, Chenlin Meng, and Stefano Ermon for an ICML 2024 Best Paper Award: Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution

Congratulations to Marco Pavone for a Robotics: Science and Systems Conference Best Paper Award!

Congratulations to Marco Pavone for winning the best paper award at the Robotics: Science and Systems Conference on AI Safety for autonomous systems.

AI existential risk probabilities are too unreliable to inform policy

How speculation gets laundered through pseudo-quantification

Building A Generative AI Platform

After studying how companies deploy generative AI applications, I noticed many similarities in their platforms. This post outlines the common components of a generative AI platform, what they do, and how they are implemented. I try my best to keep the architecture general, but certain applications might deviate. This is what the overall architecture looks like. This is a pretty complex system. This post will start from the simplest architecture and progressively add more components. In its simplest form, your application receives a query and sends it to the model. The model generates a response, which is returned to the user. There are no guardrails, no augmented context, and no optimization. The Model API box refers to both third-party APIs (e.g., OpenAI, Google, Anthropic) and self-hosted APIs. From this, you can add more components as needs arise. The order discussed in this post is common, though you don’t need to follow the exact same order. A component can be skipped if your system works well without it. Evaluation is necessary at every step of the development process. Enhance context input into a model by giving the model access to external data sources and tools for information gathering. Put in guardrails to protect your system and your users. Add model router and gateway to support complex pipelines and add more security. Optimize for latency and costs with cache. Add complex logic and write actions to maximize your system’s capabilities. Observability, which allow…

​​Reinventing Multi-Modal Search with Anyscale and MongoDB

Writing Kullback–Leibler divergence in terms of the score function

Assume we have two density functions $p(x)$ and $p'(x)$ for $x\in R^d$ . I would find a connection between Kullback–Leibler divergence between two densities in terms of the difference between the gradient of the log densities: $\nabla_{x} \log p(x)$ and $\nabla_{x} \log p'(x)$ . Something like $$\int p(x)\log\frac{p(x)}{p'(x)}dx\le \frac{1}{2}E_{x}||\nabla_{x} \log p(x)-\nabla_{x} \log p'(x)||_2^2.$$ Can anyone help me please. I have already tried starting the definition of the KL and then integrating by part, but I stuck at some points then.

Improve Utilization and Simplify Cluster Management with Anyscale Job Queues

Fair comparison method for a biased physics-based model and its ML-correction version

I'm working with two prediction models: A calibrated physics-based model that consistently overestimates and has a fixed bias. An XGBoost model that predicts the error of the physics model to create corrected predictions. (The dataset contains the physics model's predictions and the target variable. We then predict the error between the target variable and the physical models prediction) I'm trying to fairly compare these models, but I'm unsure about the best approach. Simply using MAPE for both doesn't seem meaningful due to the known bias in the physics model. Here's what I'm considering: For the physics model: Calculate errors: errors = physics_model_prediction - actual_value Compute mean error: mean_error = errors.mean() Adjust errors: adjusted_errors = errors - mean_error Calculate MAPE: MAPE(adjusted_errors, actual_value) For the corrected model: Calculate MAPE(corrected_prediction, actual_value) Is this approach valid? The residuals' distribution in the correction model is slightly skewed but close to normal, tending to overestimate at certain values. But I'm wondering if this method is statistically sound. If there are better ways to compare the improvements I would appreciate any insights!

Are there any non-transformer LLMs?

Almost all LLMs are based on the transformer architecture, but are there any examples of ones that don't use transformers?

Why using mutual information is allowed for feature selection if depends on the "scale" of entropies?

It is common to use mutual information as feature selection method. However, I fail to see why this is the case, since the mutual information $I(X, Y)$ depends on both entropies $H(X)$ and $H(Y)$ via the formula : $$ I(X, Y) = H(X) + H(Y) - H(X,Y)$$ meaning that comparing $I(X_i, Y)$ and $(X_j, Y)$ as a measure for selecting between $X_i$ and $X_j$ is not straightforward since the measure can be bloated by the marginal entropies. It is like selecting between $X_i$ and $X_j$ based on the covariance with $Y$ instead of correlation. The only way I can think that such a comparison is allowed is due to the equivalent formula: $$I(X, Y) = H(Y) - H(Y|X)$$ As the first term $H(Y)$ is the same for all $X_i$ then the ramking shouldn't depend on the "scale" of $H(Y)$ . Is that correct or am I missing something?

Introducing Anyscale’s Unified Log Viewer

Deploy Ray Serve with up to 50% fewer nodes using Anyscale Replica Compaction

Making Insights-Driven Decisions in an Ecosystem of Ecosystems

Our approach to securing our cloud environments

Les sciences du numérique à la conquête du ciel et de l’espace

Les sciences du numérique à la conquête du ciel et de l’espace mtestari mar, 07/09/2024 - 15:23 Après le lancement réussi d’Ariane 6 en juillet 2024 à Kourou, et alors que le nombre de voyageurs aériens atteint de nouveaux records en 2025, les sciences et technologies du numérique revêtent désormais une importance cruciale dans les domaines de l’aérospatial et l’aéronautique. © Pixabay/ Photo V. Stefanov Ces deux secteurs, très présents en Nouvelle-Aquitaine et en Occitanie, couvrent également des enjeux sociétaux et économiques significatifs. C’est tout un écosystème, opéré par Aerospace Valley , premier pôle de compétitivité européen, qui participe à l’étude, la conception, la fabrication et la commercialisation de ces technologies. Dans ce contexte, les équipes du Centre Inria de l’université de Bordeaux offrent à leurs partenaires industriels et académiques, les outils et les connaissances nécessaires pour renforcer la sécurité, la compétitivité et la décarbonation des systèmes, en s’appuyant sur leurs expertises telles que la modélisation, la simulation et la cryptographie . La conception des systèmes aéronautiques : un challenge scientifique pour chaque composante Il est primordial de développer des produits (aéronefs, avions, drones, lanceurs de satellites) les plus performants possible d’un point de vue du service rendu que de l’optimisation de la ressource exploitée. Grâce à la modélisation et à la simulation, Inria contribue à la création de modèles précis et réali…

Using conditional probability as an estimate in a loss function

I have a rather large ML framework that takes multiple conditional probability terms that are computed via classifiers/neural networks. This arbitrary loss function is computed via a function: loss_value = arbitrary_loss(probability1, probability2, ..., P(Y|Z)) I wish to have an end-to-end framework that computes everything and trains everything together. So I do not want to have an independently trained classifier. Say at some point I develop some intermediate values (embeddings) Z from the input samples X. I wish to model the conditional probability P(Y|Z) via an MLP softmax layer. This term P(Y|Z) is then estimated and plugged into the final loss which is the sum and product of other probabilities. P(Y|Z) = MLP(input_Z) #probability given input Z over labels My issue is that if I simply take the value of the softmax layer to estimate this probability and plug it in, at no point are the true labels taken into account for a supervised machine learning problem. How can I fix this without modifying the final loss function? TLDR: I need a probability term P(Y|Z) modeled via an MLP softmax layer to be used in a complex arbitrary loss function. How do i ensure this term is accurate via the true label values, so that it can be used in the final loss?

Understanding Empirical Bayes

I am trying to understand the basics of empirical bayes. I found myself struggling a lot to understand this so I tried to create a toy example involving the estimation of the success probabilities for a coin. 1) Traditional Bayesian Analysis (Beta-Binomial Conjugacy): Consider a sequence of $n$ coin tosses where we observe $X$ heads. The likelihood of observing $X$ heads given the true probability of heads $\theta$ is binomial: $$ P(X|\theta) = \binom{n}{X} \theta^X (1-\theta)^{n-X} $$ Chose a Beta prior distribution for the probability of heads $\theta$ : $$ P(\theta) = \text{Beta}(\alpha, \beta) = \frac{\theta^{\alpha-1} (1-\theta)^{\beta-1}}{B(\alpha, \beta)} $$ Using Bayes' theorem, the posterior distribution is proportional to the likelihood times the prior: $$ P(\theta|X) \propto P(X|\theta) P(\theta) $$ Substituting the binomial likelihood and the Beta prior, we get: $$ P(\theta|X) \propto \theta^X (1-\theta)^{n-X} \cdot \theta^{\alpha-1} (1-\theta)^{\beta-1} $$ $$ P(\theta|X) \propto \theta^{X+\alpha-1} (1-\theta)^{n-X+\beta-1} $$ Thus, the posterior distribution is also a Beta distribution: $$ P(\theta|X) = \text{Beta}(X + \alpha, n - X + \beta) $$ 2) Empirical Bayes: It seems like in Empirical Bayes, the parameters of the priors are estimated from the data instead of being chosen before hand. To me it makes more sense to use Method of Moments (instead of MLE) to estimate the parameters $\alpha$ and $\beta$ of the Beta prior. The sample mean $\hat{\theta}$ and sampl…

What's the relation between Generalized Policy Iteration (GPI), Actor-Critic, and Q-learning methods?

It seems to me that Generalized Policy Iteration (GPI) and Actor-Critic are the same, and Q-learning methods are a separate family of algorithms. I think both GPI and Actor-Critic describe the iterative process of policy evaluation (critic) and policy improvement (actor), while Q-learning is only bootstrapping using the Bellman optimality equation. To elaborate on my understanding: Policy evaluation (critic) is done via Monte-Carlo or temporal difference methods, including function approximation if necessary, and policy improvement (actor) can be done by being greedy (tabular case) or using the policy gradient theorem (large state space). Q-learning is not doing any of these. It's just trying to estimate the $Q^\ast$ using the Bellman optimality equation by iteratively fitting the Bellman optimality equation for Q value. I would appreciate it if anyone can confirm whether my understanding is correct or give a more systematic/precise summary of the taxonomy of online RL algorithms.

Extrinsic Hallucinations in LLMs

Hallucination in large language models usually refers to the model generating unfaithful, fabricated, inconsistent, or nonsensical content. As a term, hallucination has been somewhat generalized to cases when the model makes mistakes. Here, I would like to narrow down the problem of hallucination to cases where the model output is fabricated and not grounded by either the provided context or world knowledge. There are two types of hallucination: In-context hallucination: The model output should be consistent with the source content in context. Extrinsic hallucination: The model output should be grounded by the pre-training dataset. However, given the size of the pre-training dataset, it is too expensive to retrieve and identify conflicts per generation. If we consider the pre-training data corpus as a proxy for world knowledge, we essentially try to ensure the model output is factual and verifiable by external world knowledge. Equally importantly, when the model does not know about a fact, it should say so. This post focuses on extrinsic hallucination. To avoid hallucination, LLMs need to be (1) factual and (2) acknowledge not knowing the answer when applicable.

How do you save a stable diffusion model locally for later us?

I am new to ML and plan to use KerasCV stabledifussion model to generate images from text. The example on the KerasCV website is straightforward but I could not find a way to save the model locally for later use. I also noticed that the library connects to hugging face to download encoder and diffusion model. Could you please point me to the right direction to do this locally? I would like all the model and its parameters to be local and I will be using it in a server. Also, if you have experience running such a model/server on the could, I would appreciate your guidance on the best approach wrt costs. Should I upload everything and store the whole data on the cloud or load it from hugging face? Which one would make more sense for cloud applications?

New paper: AI agents that matter

Rethinking AI agent benchmarking and evaluation

An Introduction to the Code of Practice for General-Purpose AI

Last updated: 14 August 2025. As AI Act implementation gradually unfolds, it is important to understand the different mechanisms of enforcement included in the Regulation. One of the most important is the general-purpose AI Code of Practice, which was developed by the AI Office and a wide range of stakeholders. This summary, detailing the Code […]

Should I use raw score or z-score for CFA?

I am running confirmatory factor analyses and my variables have different scales. I wonder whether I should use raw score and report factor loadings of std.all column OR use z-score to run CFA.

Predicting Values with Bayesian Neural Network

I want to use a Bayesian Neural Network for a regression task. To do that I converted a BNN from this paper to Python 3. The provided training script runs and I receive a pickle file, which I want to use to predict a value for my regression. Event though the loss for the training doesn't really go down, but this seems also be the case for the models used in the paper. The input of my network is a vector with 59 features, where one feature is my target variable. It's pretty similar to the boston housing dataset. The MLP inside the BNN returns two vales: $\mu$ and $\sigma$ . If I understand BNNs correctly $\mu$ and $\sigma$ refer to the mean and standard deviation of my input and I should use $\mu$ to predict my values. But when I use input vector I always receive a value that is roughly around the mean of my target variable. So lets say the mean is 62. When I insert a vector from my test set, that should return the target variable 142, I will get 61. When I do the same with a vector, that should return 21, I will get 63. So my predicted values are always around the mean of the target variable. What do I have to do, to get real predictions from my input data? I never worked with BNNs before and it's quite hard to find any resources on how I should model my network. Maybe someone has a nice tutorial to learn how I can use this model to predict my target variable. I have to use this class and can't use Pyro or any other framework currently, because otherwise I can't use the othe…

Forecasting time series using simulations

Suppose we have a stationary time series $x_{1}, x_{2}, ..., x_{T}$ . Goal is to forecast up to $T+h$ , i.e., forecast $x_{T+1}, x_{T+2}, ..., x_{T+h}$ . Forecasting methodology: Using econometric techniques one can try to fit a model, which describes the data generating process of $\{x\}_{t}$ (e.g., ARIMA or other model). Then, having estimated parameters of the model one can simulate using forward Monte Carlo technique many paths up to $T+h$ . Using simulated paths, one can construct an empirical PDF for each point in time in the future, i.e., $f_{T+1}(x), f_{T+2}(x), ..., f_{T+h}(x)$ . As a forecast for $x_{T+1}, x_{T+2}, ..., x_{T+h}$ , take mean of $f_{T+1}(x), f_{T+2}(x), ..., f_{T+h}(x)$ , respectively. Question: Is this kind of method acceptable or widely used for stationary process forecasting? What are the main problems and assumptions of this method? Also, I assume that forecasts will be the same as the current value, i.e., $E(x_{T+i}|x_T)=x_{t}$ for any $i$ . Is this assumption correct?

How do transformer models handle negation in sentiment analysis

I'm trying to understand how transformer models, such as BERT or GPT, handle negation in sentiment analysis. Specifically, I'm curious about how these models manage to correctly interpret sentences where negation changes the sentiment, such as "The movie is not good." A simple model using word embeddings + global averaging fails to handle negation properly. Intuitively, for example, if "good" has a positive sentiment score and "bad" has a negative sentiment score, a model might misinterpret "not good" by simply averaging the scores of "not" and "good". Example Without Negation Consider the following sentences with sentiment words: "The movie is good." "The movie is awesome." "The movie is terrible." Suppose we have the following word embeddings representing sentiment scores: "good" = [10] "awesome" = [12] "terrible" = [-10] Neutral words (assuming embeddings around 0): "the" = [0] "movie" = [0] "is" = [0] For these sentences, a simple global average of the sentiment scores works well: "The movie is good" = average([0, 0, 0, 10]) = 10 / 4 = 2.5 (positive sentiment) "The movie is awesome" = average([0, 0, 0, 12]) = 12 / 4 = 3 (positive sentiment) "The movie is terrible" = average([0, 0, 0, -10]) = -10 / 4 = -2.5 (negative sentiment) Example With Negation Now, consider the sentence "The movie is not good." In this case, the sentiment should be negative due to the presence of "not." However, averaging the scores naively might not handle this correctly. For example: "The movie is…

Recap: Square Unboxed 2024

Top highlight's from this year's event

Announcing the Winners of the Square Local Community Hackathon

Building apps to connect businesses with their local community

How to actually use the Empirical Influence Function for BCa Bootstrap Intervals?

In the course of seeking reassurance for another part of a hobby analysis* I found a stack answer which mentioned The Jackknife, the Bootstrap and Other Resampling Plans (Efron, 1980). Having managed to find the paper online, the bit about non-parametric bias and skew adjustments to the bounds of bootstrap CIs caught my eye. The problem is that I don't really understand how to do it as described in that paper . Efron (1980) as I am reading it suggests that the core ingredients of the BCa CI method are the $U_i$ and $a$ values. The derivation of the latter is simple given you know the $U_i$ with formula (pg. 21): $$ a \doteq \frac{1}{6}\frac{\sum\limits_{i=0}^n U_i^3}{\left( \sum\limits_{i=0}^n U_i^2 \right)^{\frac{3}{2}}} $$ The $U_i$ have a more complicated formula, coming from what's termed the empirical influence function (also pg. 21): It is this formula for the $U_i$ that I don't know how to work with. Clicking around in some more stack questions, I found a reference to Bootstrap Methods and their Application (Davison and Hinkley, 1997) but I couldn't find that online. Nevertheless, I did find a slideshow by Davison on ResearchGate based on it which proposes that the $a$ value can be jackknifed like so (I've changed the subscript $j$ to $i$ to match Efron, 1980): $$ l_i \approx l_{jack,i} = (n-1)(\hat{\theta} - \hat{\theta}_i)$$ where $n$ is the number of observations in the original sample. I have tested my understanding of the jackknife estimate of $a$ against the $U_…

End-to-end LLM Workflows Guide

Is it possible to use attention layer as a sort of filter of input data before passing it further to the network?

Is it possible to use attention layer as a sort of filter of input data before passing it further to the network? Is it possible to use it to reduce the dimension of the input (similar as PCA, for example - only attention layer would be trained with the network) and therefore reduce the network architecture (for example we can use network which accepts smaller input dimension)? I can't find decisive sources on that matter.

How are perplexities over multiple instance aggregated?

The perplexity of the $i^{th}$ token in the $k^{th}$ sequence is $$ P_{ki} = \frac{1}{p(t_{ki})} $$ The perplexity aggregated for the $k^{th}$ sequence is then $$ P_{k} = \left(\prod_{i=1}^N P_{ki}\right)^{1/N} \\ = \left(\prod_{i=1}^N \frac{1}{p(t_{ki})} \right)^{1/N} $$ which is the geometric mean of the perplexities of the tokens. This makes sense as we are essentially taking the multiplicative inverse of the probability that the model got the whole sequence correct. Now my question is how to aggregate the perplexities of several sequences. It seems from various places, including the Hugging Face Tutorial , I see that the prescription is to take the arithmetic mean of the perplexities of sequences $$ P = \frac{1}{m} \sum_{k=1}^m P_k $$ I am not quite understanding what it means to take the average of 1/probabilities. What is this actually capturing?

Experiments in Weak-to-Strong Generalization

Writing up results from a recent project

Moderator Analysis in Cox Regression

Imagine an RCT with two groups with a time-to-event endpoint. The (pre-specified) strategy for analysing this trial is a Cox Regression using common covariate adjustment to reduce outcome heterogeneity. So far, so good. Now I want to know whether my treatment effect is moderated by level of education (3 levels). If I used the pre-specified cox regression with covariate adjustment and integrated the treatment*education interaction, the SEs are getting very large. What may the reasons be? Is my sample size (120 per group) too little and should I reduce the covariates for which I adjust? Edit: I controlled for 4 covariates. There were 76 events in the IG and 91 events in the CG. More specifically: Education level1: IG: 13 CG: 18. Education level2: IG: 32 CG: 28. Education level3: IG: 31 CG: 45

Free Form Least-Squares Concept Erasure Without Oracle Concept Labels

Achieving even more surgical edits than LEACE without concept labels at inference time.

DDPG model outputting a fixed action at every timestep

I am trying to create a Car Following model, for which i am using DDPG. My action is acceleration bounded in a range of [-3,3] m/s2. While training the model, for every state it gives a single acceleration value i.e. 3 (or sometimes -3). It can be clearly seen that my actor is performing really bad. What can be done to resolve this issue?

Z-scores (heatmap) and statistics with replicates and outliers

My group performed a biological assay.The assay was performed for 6 different cell culture conditions (columns) and for 6 genes (rows). However, it was also performed 5 times (5 replicates) per each gene. An example of our data: Genes Condition 1 Condition 2 Gene 1 replicate 1 0.22 120 Gene 1 replicate 2 0.34 122 Gene 1 replicate 3 0.45 119 Gene 1 replicate 4 0.33 34 Gene1 replicate 5 0.55 90 Gene 2 replicate 1 12 0.1 Gene 2 replicate 2 19.3 0.45 Gene 2 replicate 3 23.4 NA Gene 2 replicate 4 11.7 0.89 Gene 2 replicate 5 12.4 0.1 And so on...for 6 genes and 6 conditions. Unfortunately we also have 2 NA values (for 2 conditions in gene 5 replicate 3, which doesn't mean the value is 0, it's just undetectable by the machine (below detectable value). I have already posted a question on stackoverflow about how to perform the analysis (z-score and heatmap) with R, and a very friendly user also suggested to ask here for a more statistical insight. The final heatmap will represent the changes of these genes expression in the different conditions. Basically in the end we want to compare the conditions based on their gene expressions, and maybe a pattern. In the end we would like to state "this cell culture condition has this gene pattern, which is different from the one of that cell culture condition". Our problem number 1 is the outlier. Replicate number 4 is in every condition an outlier. Should we exclude it? If so we would only have 4 replicates for the statistical analysis (ex. c…

Can I use Prophet on a timeseries with a nonstationary daily variation?

I have some temperature data with gaps that I need to interpolate onto a uniform time series (and extrapolate a bit too). I am trying to use FB Prophet and it seems almost perfect. And yet.... If you look at the graph you can see that if the missing data is far from a daily extreme then Prophet is fine. But when the missing data is near an daily max/min then it ignores the measured data. Reading the documentation it seems that Prophet anticipates data of the form a(t) + F(H) where t is daily time, a(t) is slowly varying but nonstationary, H is the hourly time and F(H) is the periodic variation for that day, but it does not anticipate a(t) + b(t) F(H) where b is also slowly varying but nonstationary. I see in Adaptive Normalization for Non-stationary Time Series Forecasting: that those authors decompose each day into slices and then apply normalization by removing and restores the non-stationary factors a(t) and b(t) . Would that be the correct approach here? It doesnt seem completely trivial to implement because they use machine learning to do the splicing and dicing. Alternatively, I do see that Prophet allows introduction of arbitrary regressors. If I somehow estimated b(t) (e.g. by averaging each day), or whatever, then is there a way to tell Prophet about b ? How else do people interpolate gaps in time series? This seems like it should be a completely standard problem to solve, but I'm stumbling. df_p = df[["Time","Temperature"]].copy() df_p = df_p.rename(columns={"Time"…

Why work at the EU AI Office?

It's probably not for everyone, but there are a lot of great reasons to consider, including the potential to have an impact on AI governance worldwide, leveraging the first-mover advantage, and more.

Would the DDPG algorithm still function effectively if some transitions stored in its replay buffer are generated by a completely unrelated policy?

Let's hypothesize a scenario where some of the records ( s i , a i , r i , s i+1 ) in the replay buffer are generated by another completely unrelated random policy. If the DDPG algorithm still samples random minibatches from this buffer for learning as usual, would the learning process proceed successfully? Actually, there's a pre-processing stage before training DDPG in my online learning application, where another module learns the safe action range. I wonder if the transition records obtained during this stage can be used to pre-train the DDPG agent.

Q&A: Discover what’s new in the Mobile Payments SDK beta

Learn about how KIOS, a kiosk software, conducted early stage testing of the SDK and their integration experience.

Cross-modal Search for E-commerce: Building and Scaling a Cross-Modal Image Retrieval App

Bayes rule and terms with expectation

I am reading the following paper in economics; link On page 495, authors give an expression with Bayes rule. As an example, say that there is a random variable $\beta$ which can be either $\beta_L$ or $\beta_H$ . So, there is an information that you receive, say $T$ such that $$H_t =\beta X_t+\epsilon_t$$ Let the probability of having a $\beta_H$ is $\pi_t$ . Then, with the Bayes rule, we can have $$\pi_{t+1}=\frac{\pi_{t}exp\left(-p_{\varepsilon}\left(H_{t}-\beta_{H}X_{t}\right)^{2}\right)}{\pi_{t}exp\left(-p_{\varepsilon}\left(H_{t}-\beta_{H}X_{t}\right)^{2}\right)+\left(1-\pi_{t}\right)exp\left(-p_{\varepsilon}\left(H_{t}-\beta_{L}X_{t}\right)^{2}\right)}$$ My question is; is it possible to take the expectation of the probability $\pi$ . Normally, is not that $\pi_{t+1}$ is the conditional expectation of $\pi$ on the signal $H$ ? I ask this question because authors of the paper take the expectation of this probability on page 516. Also, for Bayesian updating (as in Kalman filter), the updated mean (posterior) is the expectation of the random variable, conditional on the signal. Thanks a lot in advance for any hints/suggestions.

Minimum Survey Respondents for Hypotheses I want to answer

I am an eager statistics noob, so apologies on the basic questions. I am looking to send out a survey to a population that is large (people with diabetes who take insulin in the US) (~ 8.4 million) Based on this I am aiming for 96+ survey responses overall (with a margin of error of +- 10%), with a goal of getting closer to 300 responses. I have some hypotheses I want to look at in the survey results and will be doing some statistical testing One hypothesis is "Hypothesis: There is a significant difference in rating of importance of (xyz user need, just giving a fake example - being able to quickly calculate amount of insulin for a meal) based on time the user has been using insulin" This would be a demographic question I collect (How long have you been using insulin for?), with a range of responses in months. Since importance is a rating on a likert scale (ordinal), the groups are different, and I want to understand difference, I was thinking of doing the Mann-Whitney test. I specifically would like to understand if longer time using insulin would impact rating of importance on this one user need. My question: Would I need a minimum of 96 respondents who answer in each bucket (for example 1 year or longer and 1 year or shorter) to aim for a stat sig response or would I just need to be looking at 96 overall for all respondents who answer this q. Also, if you have feedback on my approach feel free to weigh in.

Is reinforcement learning suitable for application automation?

I have basically automatised the use of an app through the use of OCR and computer vision. So basically when a word or an image is detected it will perform a certain action. When that action is successfully completed it will go to the next state. Now I want to try basically with a more "heuristic" approach and I thought about reinforcement learning. Why? Because I am aiming to build a tool that basically understand automatically what actions to perform in a certain state. But I have a doubt. Even though I don't need to declare an association like this (it would beat the purpose of deep reinforcement learning or deep learning in general): if(state.MENU_VIEW) clickManager.clickOnFolder(); ... I still need to define the states, the actions and the reward. Meaning I would need to instruct my app that when the OCR result is "Open Folder" it means the state I am in is MENU_VIEW. I simply wouldn't tell my app what action to perform in a that state. Am I correct? What I am trying to say is: how exactly could I make it so that the states (and maybe also the actions?) are generated automatically? The reward in this case scenario would be basically the folder being opened successfully.

Probability interpretation of attention mechanism in Seq2Seq

I have ready many explanations of the seq2seq model. In my opinion, however, it is really like a robot that might say something correctly, but doesn't really understand it, just as is true with an LLM generally. In my opinion, the correct way to describe Seq2Seq and similar NLP models should start from a probability view. My probability view is very simple; the output of the encoder is a representation of the probability distribution of the next word. In each step of the Decoder, it just modifies the distribution based on each word it predicted from the distribution and outputs the modified distribution. It then does this repeatedly. Assuming this probability view is correct, how could we explain the attention mechanism used in Seq2seq?

How to evaluate the relevance of a search made with an AI Agent?

I’m building an AI agent designed to find the best results for with a search tool. The agent is essentially a Large Language Model (LLM) connected to a Search API via function calling. The goal is for this agent to use the Search API to identify the most relevant results. Let's imagine that I have an evaluation function f(x) that allow to score a specific result. How can I evaluate the performance of my agent ? For instance, how can I tell between two versions of my agent (v1 and v2), which one is performing better ? Here is a specific constraint due to the fact that in a real world search, I don’t have access to the entire database of results due to the sheer volume (millions of results) and the cost involved. Since I can’t purchase all these results, it’s difficult to know if a specific record should be included in the search results. So I can't use recall like metrics. Given this constraint, how can I effectively evaluate the performance of my AI agent?

Providing lawyers with a leg-up in navigating AI for legal practice

AISG and the Singapore Academy of Law (SAL) will work together to support the legal profession as it faces a time of transformation, precipitated by the rise of Generative...

IBM and AI Singapore Ink MOU to Collaborate on First Large Language Model developed with Southeast Asian context

Ahead of ATx Enterprise 2024, IBM announced the signing of a memorandum of understanding (MOU) with AI Singapore (AISG) on 28 May 2024 to collaborate on the testing of...