AI/ML News & Innovations Hub

Comparing differences between discrete variables in two datasets of unequal size

I have two datasets, X, Y, containing discrete variables which both have a different number of instances. I want to analyse the difference between the datasets for each of these variables. When these discrete variables are binary, I can use the Chi2 test (as suggested in Chi square test when sample sizes are different? ) or Fishers Exact test (as in Compare means of two datasets of binary data ). However, one of these discrete features isn't binary and I'm struggling to find a test for discrete variables where the datasets are of unequal length. This discrete variable is ordinal and has values 0, 1 or 2. Maybe one approach could be to treat this as a continuous variable and use a non-parametric test for continuous distributions like Kolmogorov-Smirnov test, but this gives me suspiciously low p values (e.g. $1\times 10 ^{-60}$ ). Another approach might be to bin the variable by setting 1 to 0 or 2, and then use one of the above tests, but this would skew the data and lose information. Is there an appropriate test for this situation?

Understanding Deep Learning Algorithms that Leverage Unlabeled Data, Part 1: Self-training

Deep models require a lot of training examples, but labeled data is difficult to obtain. This motivates an important line of research on leveraging unlabeled data, which is often more readily available. For example, large quantities of unlabeled image data can be obtained by crawling the web, whereas labeled datasets such as ImageNet require expensive labeling procedures. In recent empirical developments, models trained with unlabeled data have begun to approach fully-supervised performance (e.g., Chen et al., 2020 , Sohn et al., 2020 ). This series of blog posts will discuss our theoretical work which seeks to analyze recent empirical methods which use unlabeled data. In this first post, we’ll analyze self-training , which is a very impactful algorithmic paradigm for semi-supervised learning and domain adaptation . In Part 2, we will use related theoretical ideas to analyze self-supervised contrastive learning algorithms, which have been very effective for unsupervised representation learning . Background: self-training We will first provide a basic overview of self-training algorithms, which are the main focus of this blog post. The core idea is to use some pre-existing classifier \(F_{pl}\) (referred to as the “pseudo-labeler”) to make predictions (referred to as “pseudo-labels”) on a large unlabeled dataset, and then retrain a new model with the pseudo-labels. For example, in semi-supervised learning, the pseudo-labeler is obtained from training on a small labeled datase…

Comment on Parenting in the age of digi-schools in Kenya by Dr. Florence Kisirkoi

This is good research. True. There is disconnect between parents and children in digital world

Generative model that satisfies certain algebraic constraints

Disclaimer : I need guidance and help with where to start looking for solution of the problem I have described below. My background is in optimization and I am new to statistical methods, so there is a good chance that I am asking the wrong question or/and used wrong terminologies (please correct if that is the case). Below I setup my problem: Given: the set of $n\times n$ matrices. two functoins, $f: \mathbb{R}^{n^2} \rightarrow\mathbb{R}$ and ${\bf{g}}: \mathbb{R}^{n^2}\rightarrow Symm.(n\times n)$ Two constraints as follows: $$1 - \epsilon_1 $$-\epsilon_2 for ${\bf{M}} \in \mathbb{R}^{n^2}$ and $\epsilon_1$ and $\epsilon_2$ both fixed small positive numbers. Here are my questions: Can I find a model, or a distribution, which when I sample from it, it produces $n\times n$ matrices that satisfy the above two constraints (most of the time)? The sampled data needs to be close the real distribution in order to be representative. Is question (1) a well formulated question? If the answer to (2) is yes, what method(s) should I look into in order to work towards a solution? For those who are interested in more concrete realizations of functions $f$ and $\bf{g}$ , $f({\bf{M}})=\mathrm{det}({\bf{M}})$ and ${{\bf{g}}}({\bf{M}})=\frac{1}{2}({\bf{M}}^T{\bf{M}} - {\bf{I}})$ . I appreciate any hint or help with this problem. Tags mentioned below are speculative.

Simulating Multiple Time Series with Relationships

Suppose I have several time series (these are financial series, prices, indicators) with the same time. There may be two or more. I don’t know what relationships there are between the series, correlations, cointegrations, non-linear relationships, or there are no relationships at all ... I would like to automatically find out if there are connections and what I would like to simulate these series with connections UPD======= Here is an example of the data I am using library(quantmod) getFX("EUR/USD") getFX("GBP/USD") eu $EUR.USD) gb GBP.USD) library(TTR) roll.cor mydata eu gb roll.cor rsi [1,] 1.177237 1.372932 NA NA [2,] 1.179440 1.376770 NA NA [3,] 1.179474 1.376679 NA NA [4,] 1.179910 1.376033 NA NA [5,] 1.181616 1.376859 0.83911455 NA [6,] 1.182466 1.376244 -0.15342956 100.000000 [7,] 1.185494 1.380270 0.84880800 100.000000 [8,] 1.188066 1.384948 0.95413564 100.000000 [9,] 1.187910 1.386290 0.97020360 97.715547 [10,] 1.187945 1.386288 0.98141920 97.730090 [11,] 1.186888 1.384056 0.95570036 78.794634 [12,] 1.186068 1.380795 0.93082598 66.331768 [13,] 1.182712 1.377056 0.97051123 36.664213 [14,] 1.182410 1.381155 0.82146557 34.907988 I have a trading strategy that works on this historical data, I would like to simulate many hundreds of years of similar data in order to test the strategy better and also find more optimal parameters... I'm wondering if this can be done and how to do it

Stanford AI Lab Papers and Talks at AAAI 2022

The 36th AAAI Conference on Artificial Intelligence (AAAI 2022) is being hosted virtually from February 22th - March 1st. 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 Partner-Aware Algorithms in Decentralized Cooperative Bandit Teams Authors : Erdem Bıyık, Anusha Lalitha, Rajarshi Saha, Andrea Goldsmith, Dorsa Sadigh Contact : ebiyik@stanford.edu Links: Paper | Video | 2nd Video | Website Keywords : bandits, multi-agent systems, collaboration, human-robot interaction, partner-awareness Constraint Sampling Reinforcement Learning: Incorporating Expertise For Faster Learning Authors : Tong Mu, Georgios Theocharous, David Arbour, Emma Brunskill Contact : tongm@stanford.edu Links: Paper Keywords : reinforcement learning, constraints IS-Count: Large-scale Object Counting from Satellite Images with Covariate-based Importance Sampling Authors : Chenlin Meng*, Enci Liu*, Willie Neiswanger, Jiaming Song, Marshall Burke, David Lobell, Stefano Ermon Contact : jesslec@stanford.edu Award nominations: Oral presentation Links: Paper | Blog Post | Website Keywords : remote sensing, sampling PantheonRL Authors : Bidipta Sarkar, Aditi Talati, Andy Shih, Dorsa Sadigh Contact : bidiptas@stanford.edu Links: Paper | Video | Website Keywords : multiagent reinforcement learning; soft…

Learning with not Enough Data Part 2: Active Learning

This is part 2 of what to do when facing a limited amount of labeled data for supervised learning tasks. This time we will get some amount of human labeling work involved, but within a budget limit, and therefore we need to be smart when selecting which samples to label.

Multi-level / Hierarchical Machine Learning

I am trying to tackle a problem that involves binary classification. However, my data is multi-level or hierarchical in its structure. This example illustrates its structure: Time Group Response F1 F2 F3 F4 1 1 a 0 0.89 0.01 10 2 2 2 a 1 0.82 0.58 21 4 3 3 a 1 0.34 0.39 37 6 4 1 b 0 0.32 0.26 10 2 5 2 b 0 0.32 0.04 21 4 6 3 b 0 0.00 0.20 37 6 7 1 c 1 0.80 0.18 10 2 8 2 c 0 0.34 0.71 21 4 9 3 c 1 0.55 0.52 37 6 We have multiple groups which have a binary response (0, 1) that is measured over time. The are an array of features associated with each response in each group. Some of these features take unique values for each time point in each group (F1, F2) , while others (F3, F4) represent the environmental "state" that was common across groups at each time point. All feature variables are continuous and numeric and I also do not tell the model that they belong to different groups as I want the model to predict the response based on the continuous features. This is important because over time some groups go extinct and disappear from the data set and other new groups come in so I want the model to be agnostic to group ID. I use a rolling window walk-forward validation scheme to account for the temporal nature of the data. Until now, I have been running this data through a random forest model via the h2o package in R with y = Response and x = c(F1:F4) . The results have been reasonable, however I am noticing that the environmental features (F3, F4) have the most importance and th…

Mask R-CNN: How are the computed masks projected back to the input image?

The computed masks by Mask R-CNN are of fixed size $m \times m$ each. How are they projected back to the image?

Multiclass Ensemble Methods with weak classifiers under 50%

Normally, when using an ensemble method, such as baggin or boosting, in binary classification, there is a reqauirment that each weak classifier have accuracy better than 50%. In the multiclass claaification setting, this is often infeasible. Is there a way to improve upon multiclass classification with ensembles. For an example to make this concrete: Say I have a problem with 1000 classes, and I train 50 models, each with 10% accuracy, which is 100x better than random guessing. Is there a way to combine these models to form a better classification algorithm?

Comment on Digital taxation: Can it contribute to more just taxation? by Andrew Rens

An important and timely paper. It would benefit from global historical context. That context would include the ways in which the most developed countries allowed the current technology multinationals to structure their affairs to avoid paying tax. For example an subsidiary of Apple in the 2010's paid no tax anywhere in the world for five years despite billions in revenue. It is only more recently that developed countries have begun to impose taxes on technology companies. The impact of this on global competitiveness shouldn't be underestimated. Similarly the preference in economics and tax literature for residence over source based taxation has tended to favour developed countries even prior to digital services. The e-commerce customs moratorium at the WTO would also repay greater attention. The purported rationale for this, that digital economies were not yet mature or well understood is no longer tenable. In this context the suggestion that "It is time to start considering the lifting of this moratorium, which the Africa cohort within the WTO are in favour of doing" is an invitation for delay since even a strong demand that it be lifted immediately would be met with strenuous resistance and lengthy delays. It would also be helpful to explain more clearly why revenue flow from an African country to a developed country is not itself a sufficient nexus for taxation.

How to scale data for model retraining on production?

Let's say I have a basic regression model being used in production and now I want to implement periodical model retraining (i.e. once a month) where I take a batch of new data from last month and fit old model on this new batch with one epoch only. Assuming that model is using MinMaxScaler as feature normalization mechanism, how should I proceed with scaling during such automated periodical retraining? Should I scale the data with old scaler, that was fitted on the initial training set or should I somehow fit the scaler again but if so, on what data? Only the newest batch or newest concatenated with the old initial training set?

#LearnAI Stories – Ankur Shukla

cuDNN isn't found FWD algo for convolution. How to TRAIN DARKNET ON GE FORCE GTX 1650

ISSUE: while training Darknet with GE FORCE GTX 1650 using following: CUDA 11.0 cuDNN 8.0.5 OPENCV 4.5 Model starts training with config file details as below for [net] section: [net] # Testing #batch=1 #subdivisions=1 # Training batch=64 subdivisions=16 width=416 height=416 channels=3 momentum=0.949 decay=0.0005 angle=0 saturation = 1.5 exposure = 1.5 hue=.1 learning_rate=0.001 burn_in=1000 max_batches = 6000 policy=steps steps=4800,5400 scales=.1,.1 #cutmix=1 mosaic=1 #:104x104 54:52x52 85:26x26 104:13x13 for 416 When I change the batch from 64 to 32 (reducing it ) coupled with subdivisions increased from 16 to as high as 128, it keeps on going well but then stops after some time as reported above: Error: cuDNN isn't found FWD algo for convolution.

How to Improve User Experience (and Behavior): Three Papers from Stanford's Alexa Prize Team

Introduction In 2019, Stanford entered the Alexa Prize Socialbot Grand Challenge 3 for the first time, with its bot Chirpy Cardinal , which went on to win 2nd place in the competition. In our previous post , we discussed the technical structure of our socialbot and how developers can use our open-source code to develop their own. In this post we share further research conducted while developing Chirpy Cardinal to discover common pain points that users encounter when interacting with socialbots, and strategies for addressing them. The Alexa Prize is a unique research setting, as it allows researchers to study how users interact with a bot when doing so solely for their own motivations. During the competition, US-based Alexa users can say the phrase “let’s chat” to speak in English to an anonymous and randomly-selected competing bot. They are free to end the conversation at any time. Since Alexa Prize socialbots are intended to create as natural an experience as possible, they should be capable of long, open-domain social conversations with high coverage of topics. We observed that Chirpy users were interested in many different subjects, from current events (e.g., the coronavirus) to pop culture (e.g., the movie Frozen 2 ) to personal interests (e.g,. their pets). Chirpy achieves its coverage of these diverse topics by using a modular design that combines both neural generation and scripted dialogue, as described in our previous post . We used this setting to study three quest…

Why does OpenAI's PPO algorithm not follow the discounting method used in Sutton & Barto?

As discussed in this question , the policy gradient algorithms given in Reinforcement Learning: An Introduction use the gradient \begin{align*} \gamma^t \hat A_t \nabla_{\theta} \log \pi(a_t \, | \, s_t, \theta) \end{align*} where $\hat A_t$ is the advantage estimate for step $t$ . For example, $\hat A_t = r_t + \gamma V(s_{t+1}) - V(s_t)$ in the one-step actor-critic algorithm given in section 13.5. In the answers to the linked question, it is claimed that the extra discounting is "correct", which implies that it should be included. If I look in the literature to a seminal paper such as Proximal Policy Optimization Algorithms by OpenAI, they do not include the extra discounting factor, i.e. they use a gradient defined as \begin{align*} \hat A_t \dfrac{\nabla_{\theta}\pi(a_t \, | \, s_t, \theta)}{\pi(a_t \, | \,s_t, \theta_{\rm old})} \end{align*} which does not include the discounting factor (of course, it's dealing with the off-policy case, but I don't see how that would make a difference in terms of the discounting). OpenAI's implementation of PPO also does not include the extra discounting factor. So, how am I supposed to interpret this discrepancy? I agree that the extra discounting factor should be present, from a theoretical standpoint. Then, why is it not in the OpenAI code or paper?

How do you solve strictly constrained optimization problems with pytorch?

I am trying to solve the following problem using pytorch: given a six sided die whose average roll is known to be 4.5, what is the maximum entropy distribution for the faces? (Note: I know a bunch of non-pytorch techniques for solving problems of this sort - my goal here is really to be better understand how to solve constrained optimization problems in general with pytorch. In real life I'm working on a much harder constrained optimization problem involving a neural model implemented in pytorch, and I'm hoping that if I can solve this problem then it will help with the harder problem.) In principle it should be possible to handle this by looking for critical points of the Lagrangian: $$L(p) = -\sum_i p_i \log p_i + \lambda\left(\sum_i p_i - 1\right) + \mu\left(\sum_i i p_i - 4.5\right)$$ Here's my attempt to do this with pytorch: class MaxEntropyDice(torch.nn.Module): def __init__(self, num_faces=6, mean_constraint=3.5): super().__init__() self.num_faces = num_faces self.mean_constraint = mean_constraint self.p = torch.nn.Parameter(F.normalize(torch.rand(num_faces), p=1, dim=0)) self.probability_multiplier = torch.nn.Parameter(torch.rand(1)) self.mean_multiplier = torch.nn.Parameter(torch.rand(1)) def forward(self): entropy = -torch.sum(self.p * torch.log(self.p)) probability_term = self.probability_multiplier * (torch.sum(self.p) - 1) mean_term = self.mean_multiplier * ( torch.sum(torch.tensor(range(1, self.num_faces + 1)) * self.p) - self.mean_constraint ) lagrangian = en…

#LearnAI Stories – Wan Rong

Reward Isn't Free: Supervising Robot Learning with Language and Video from the Web

This work was conducted as part of SAIL and CRFM . Deep learning has enabled improvements in the capabilities of robots on a range of problems such as grasping 1 and locomotion 2 in recent years. However, building the quintessential home robot that can perform a range of interactive tasks, from cooking to cleaning, in novel environments has remained elusive. While a number of hardware and software challenges remain, a necessary component is robots that can generalize their prior knowledge to new environments, tasks, and objects in a zero or few shot manner. For example, a home robot tasked with setting the dining table cannot afford lengthy re-training for every new dish, piece of cutlery, or dining room it may need to interact with. A natural way to enable such generalization in our robots is to train them on rich data sources that contain a wide range of different environments, tasks, and objects. Indeed, this recipe of massive, diverse datasets combined with scalable offline learning algorithms (e.g. self-supervised or cheaply supervised learning) has been the backbone of the many recent successes of foundation models 3 in NLP 4 5 6 7 8 9 and vision 10 11 12 . Replicating these impressive generalization and adaptation capabilities in robot learning algorithms would certainly be a step toward robots that can be used in unstructured, real world environments. However, directly extending this recipe to robotics is nontrivial, as we neither have sufficiently large and diverse…

The Illustrated Retrieval Transformer

Discussion: Discussion Thread for comments, corrections, or any feedback. Translations: Korean, Russian Summary: The latest batch of language models can be much smaller yet achieve GPT-3 like performance by being able to query a database or search the web for information. A key indication is that building larger and larger models is not the only way to improve performance. Video The last few years saw the rise of Large Language Models (LLMs) – machine learning models that rapidly improve how machines process and generate language. Some of the highlights since 2017 include: The original Transformer breaks previous performance records for machine translation. BERT popularizes the pre-training then finetuning process, as well as Transformer-based contextualized word embeddings. It then rapidly starts to power Google Search and Bing Search. GPT-2 demonstrates the machine’s ability to write as well as humans do. First T5, then T0 push the boundaries of transfer learning (training a model on one task, and then having it do well on other adjacent tasks) and posing a lot of different tasks as text-to-text tasks. GPT-3 showed that massive scaling of generative models can lead to shocking emergent applications (the industry continues to train larger models like Gopher, MT-NLG…etc). For a while, it seemed like scaling larger and larger models is the main way to improve performance. Recent developments in the field, like DeepMind’s RETRO Transformer and OpenAI’s WebGPT, reverse this tre…

BanditPAM: Almost Linear-Time k-medoids Clustering via Multi-Armed Bandits

TL;DR Want something better than \(k\)-means? Our state-of-the-art \(k\)-medoids algorithm from NeurIPS, BanditPAM, is now publicly available! \(\texttt{pip install banditpam}\) and you're good to go! Like the \(k\)-means problem, the \(k\)-medoids problem is a clustering problem in which our objective is to partition a dataset into disjoint subsets. In \(k\)-medoids, however, we require that the cluster centers must be actual datapoints, which permits greater interpretability of the cluster centers. \(k\)-medoids also works better with arbitrary distance metrics, so your clustering can be more robust to outliers if you're using metrics like \(L_1\). Despite these advantages, most people don't use \(k\)-medoids because prior algorithms were too slow. In our NeurIPS paper, BanditPAM, we sped up the best known algorithm from \(O(n^2)\) to \(O(n\text{log}n)\). We've released our implementation, which is pip-installable. It's written in C++ for speed and supports parallelization and intelligent caching, at no extra complexity to end users. Its interface also matches the \(\texttt{sklearn.cluster.KMeans}\) interface, so minimal changes are necessary to existing code. Useful Links: 3-minute video summary PyPI Github Repository Full Paper \(k\)-means vs. \(k\)-medoids If you're an ML practitioner, you're probably familiar with the \(k\)-means problem. In fact, you may know some of the common algorithms for the \(k\)-means problem. You're much less likely, however, familiar with the…

Stanford AI Lab Papers and Talks at NeurIPS 2021

The thirty-fifth Conference on Neural Information Processing Systems (NeurIPS) 2021 is being hosted virtually from Dec 6th - 14th. We’re excited to share all the work from SAIL that’s being presented at the main conference , at the Datasets and Benchmarks track and the various workshops , and you’ll find links to papers, videos and blogs below. Some of the members in our SAIL community also serve as co-organizers of several exciting workshops that will take place on Dec 13-14, so we hope you will check them out! Feel free to reach out to the contact authors and the workshop organizers directly to learn more about the work that’s happening at Stanford! Main Conference Improving Compositionality of Neural Networks by Decoding Representations to Inputs Authors : Mike Wu, Noah Goodman, Stefano Ermon Contact : wumike@stanford.edu Links: Paper Keywords : generative models, compositionality, decoder Reverse engineering recurrent neural networks with Jacobian switching linear dynamical systems Authors : Jimmy T.H. Smith, Scott W. Linderman, David Sussillo Contact : jsmith14@stanford.edu Links: Paper | Website Keywords : recurrent neural networks, switching linear dynamical systems, interpretability, fixed points Compositional Transformers for Scene Generation Authors : Drew A. Hudson, C. Lawrence Zitnick Contact : dorarad@cs.stanford.edu Links: Paper | Github Keywords : GANs, transformers, compositionality, scene synthesis Combining Recurrent, Convolutional, and Continuous-time Mode…

Learning with not Enough Data Part 1: Semi-Supervised Learning

When facing a limited amount of labeled data for supervised learning tasks, four approaches are commonly discussed.

How to obtain conditional use accuracy equality with communities with different real positive rates?

The short version is that I would like to know what the confusion matrices (numbers of true positives, false positives, true negatives, and false negatives) should be to achieve conditional use accuracy equality among two communities, one with 40% real positives and one with 60% real positives. Here is the long version... I am trying to understand the difference fairness metrics described in Understanding Fairness . It contains an interactive pair of pie charts representing two communities' confusion matrices. Blue portions represent real positives (RP); striped portions represent predicted positives. To the right of the charts are different fairness metrics and how well the proportions in the pie charts satisfy them. Here is the original configuration: The small red and blue circles are handles for adjusting the sizes of regions. As shown in green, these fairness criteria are achieved: Group fairness Equalized Odds Overall Accuracy Equality I have been unable to come up with adjustments that satisfy Conditional Use Accuracy Equality. As described in the document, the two communities should have the same: positive predictive value (PPV) or precision, i.e. TP / Predicted Positive, and negative predictive value (NPV), i.e. TN / Predicted Negatives Can Conditional Use Accuracy Equality be obtained in this scenario?

Linear Regression - Data Subset with Lowest Mean-Squared Error?

Short version: given a linear regression dataset and an integer $K$ , what data subset of size $K$ results in linear parameters with the lowest mean squared error on the entire dataset? Long version: Suppose I have a supervised dataset $D_n := \{x_n, y_n\}_{n=1}^N$ and I want to perform ordinary linear regression by minimizing $$ L_{OLS, N}(w) = ||Y_N - X_N w||_2^2$$ where $Y_N$ is the matrix formed by stacking $y_n$ as row vectors and $X_N$ is the matrix formed by stacking $x_n$ as row vectors. I use the subscript $N$ to remind us that we used all the data to compute the parameter estimates. Alternatively, I might want to perform ridge linear regression by minimizing $$ L_{Ridge, N}(w) = ||Y_N - X_N w||_2^2 + c ||w||_2^2$$ We know that the parameters that minimize the above losses are given by $$w_{OLS, N} = (X_N^T X_N)^{-1} X_N^T Y_N$$ $$w_{Ridge, N} = (X_N^T X_N + c I)^{-1} X_N^T Y_N$$ What I'm curious to know is: if we're forced to choose a subset of only $K data $D_K := \{(x_k, y_k)\}_{k=1}^K \subset D_N$ and we fit parameters $$w_{OLS, K} = (X_K^T X_K)^{-1} X_K^T Y_K$$ $$w_{Ridge, K} = (X_K^T X_K + c I)^{-1} X_K^T Y_K$$ what subset of $K$ data results in parameters $w_{OLS, K}$ and $w_{Ridge, K}$ such that $L_{OLS, N}(w)$ and $L_{Ridge, N}(w)$ (respectively) are minimized?

Why is there a 1 in complexity formula of uniform-cost search?

I am reading the book titled Artificial Intelligence: A Modern Approach 4th ed by Stuart Russell and Peter Norvig. According to the book, the complexity of uniform-cost search is as $$ O(b^{1+\lfloor{C^*/\epsilon}\rfloor}), $$ where $b$ is the branching factor (i.e. the number of available actions in each state), $C^*$ is the cost of the optimal solution, and $\epsilon > 0$ is a lower bound of the cost of each action. My question is: Why is there is a 1 in the formula? For example, suppose in the following tree, the red node is the initial state and the green one is the goal state, and two actions are needed to reach the goal state from the initial state. If the cost of both actions is equal to $\epsilon = 1$ , so, $C^*$ will be $2$ . Therefore, the complexity will be $O(b^{2})$ . But, from the above formula, the complexity will be $O(b^{3})$ . PS. I know there is a similar question in stackoverflow and have read the answer. But there is a disagreement between the answers about the 1.

Improving Singapore’s AI Literacy With AI for Everyone

AI Readiness Index – Where Are You on the AI Journey

Building Platforms and Products at AI Singapore

Improving Fraudulent Claims Detection with AI

Get on the AI Transformation Journey With AISG

PMET Transitions – Passion, Pain and Plan

Stanford AI Lab Papers at CoRL 2021

The Conference on Robot Learning (CoRL 2021) will take place next week. We’re excited to share all the work from SAIL that will be 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 LILA: Language-Informed Latent Actions Authors : Siddharth Karamcheti*, Megha Srivastava*, Percy Liang, Dorsa Sadigh Contact : skaramcheti@cs.stanford.edu, megha@cs.stanford.edu Keywords : natural language, shared autonomy, human-robot interaction BEHAVIOR: Benchmark for Everyday Household Activities in Virtual, Interactive, and Ecological Environments Authors : Sanjana Srivastava*, Chengshu Li*, Michael Lingelbach*, Roberto Martín-Martín*, Fei Xia, Kent Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, C. Karen Liu, Silvio Savarese, Hyowon Gweon, Jiajun Wu, Li Fei-Fei Contact : sanjana2@stanford.edu Links: Paper | Website Keywords : embodied ai, benchmarking, household activities Co-GAIL: Learning Diverse Strategies for Human-Robot Collaboration Authors : Chen Wang, Claudia Pérez-D’Arpino, Danfei Xu, Li Fei-Fei, C. Karen Liu, Silvio Savarese Contact : chenwj@stanford.edu Links: Paper | Website Keywords : learning for human-robot collaboration, imitation learning DiffImpact: Differentiable Rendering and Identification of Impact Sounds Authors : Samuel Clarke, Negin Heravi, Mark Rau, Ruohan Gao, Jiajun Wu, Doug James, Jeannette Bohg Contact : spclar…

Stanford AI Lab Papers at EMNLP/CoNLL 2021

The 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP 2021) will take place next week, colocated with CoNLL 2021. We’re excited to share all the work from SAIL that will be 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 Calibrate your listeners! Robust communication-based training for pragmatic speakers Authors : Rose E. Wang, Julia White, Jesse Mu, Noah D. Goodman Contact : rewang@stanford.edu Links: Paper | Video Keywords : language generation, pragmatics, communication-based training, calibration, uncertainty Cross-Domain Data Integration for Named Entity Disambiguation in Biomedical Text Authors : Maya Varma, Laurel Orr, Sen Wu, Megan Leszczynski, Xiao Ling, Christopher Ré Contact : mvarma2@stanford.edu Links: Paper | Video Keywords : named entity disambiguation, biomedical text, rare entities, data integration ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts Authors : Yuta Koreeda, Christopher D. Manning Contact : koreeda@stanford.edu Links: Paper | Website Keywords : natural language inference, contract, law, legal, dataset Venue : The Findings of EMNLP 2021 The Emergence of the Shape Bias Results from Communicative Efficiency Authors : Eva Portelance, Michael C. Frank, Dan Jurafsky, Alessandro Sordoni, Romain Laroche Contact : portelan@stanford.edu Links…

Obtain a between-class similarity. And is the way to do it through PCA valid?

Context: I have a dataset containing instances labeled into different classes, and for all the classes, I have the same set of features. My research question is to identify classes that are more similar to each other. My initial thought was to compare these classes by estimating the pairwise similarity. And, by pairwise similarity, I mean the similarity matrix between all the classes considered. As bellow: Similarity matrix for classes A, B, C, D : A B C D A 1.0, 0.3, 0.7, 0.8 B 0.3, 1.0, 0.2, 0.4 C 0.7, 0.2, 1.0, 0.9 D 0.8, 0.4, 0.9, 1.0 Example: For simplicity, let's consider the iris dataset . And my goal is to find if iris setosa is more similar to iris virginica or to iris versicolor. I want to compute the similarity for each possible pair (a,b) for a,b in (setosa, virginica, and versicolor) . Assume that I have standardized all the features between 0 and 1 universally. Only after standardizing, I separated the iris labeled instances into 3 subsets ( X_setosa , X_virginica , X_versicolor ), according to their classes. Then, I have generated 3 PCs ( PC_setosa , PC_virginica , and PC_versicolor ), one for each set s as bellow: pca_s = PCA(n_components=2) pca.fit_transform(X_s) PC_s = pca_s.components_ My questions are: Does that idea of comparing the PCs (eigenvectors) as a proxy for classes similarity make sense? How could I compare the PCs structures using the cosine similarity? After some googling, I don't know if its better to compare the loadings or the eigenvectors…

Selective Classification Can Magnify Disparities Across Groups

Selective classification, where models are allowed to “abstain” when they are uncertain about a prediction, is a useful approach for deploying models in settings where errors are costly. For example, in medicine, model errors can have life-or-death ramifications, but abstentions can be easily handled by backing off to a doctor, who then makes a diagnosis. Across a range of applications from vision 1 2 3 and NLP 4 5 , even simple selective classifiers, relying only on model logits, routinely and often dramatically improve accuracy by abstaining. This makes selective classification a compelling tool for ML practitioners 6 7 . However, in our recent ICLR paper, we find that despite reliably improving average accuracy, selective classification can fail to improve and even hurt the accuracy over certain subpopulations of the data . As a motivating example, consider the task of diagnosing pleural effusion, or fluid in the lungs, from chest X-rays. Pleural effusion is often treated with a chest drain, so many pleural effusion cases also have chest drains, while most cases without pleural effusion do not have chest drains 8 . While selective classification improves average accuracy for this task, we find that it does not appreciably improve accuracy on the most clinically relevant subgroup, or subpopulation, of the data: those that have pleural effusion but don’t yet have a chest drain, i.e. those that have pleural effusion but have not yet been treated for it. Practitioners, thus,…

Duke Computer Scientist Wins $1 Million Artificial Intelligence Prize, A ‘New Nobel’

Duke professor Cynthia Rudin is the second recipient of the AAAI Squirrel AI Award for pioneering socially responsible AI. She is being cited for “pioneering scientific work in the area of interpretable and transparent AI systems in real-world deployments, the advocacy for these features in highly sensitive areas such as social justice and medical diagnosis, and serving as a role model for researchers and practitioners.” The post Duke Computer Scientist Wins $1 Million Artificial Intelligence Prize, A ‘New Nobel’ appeared first on AAAI .

Closed networks vs Networks with a removed delay to predict new data

I've come across two types of neural networks to predict, both from Matlab, the closed structure and the net that removes one delay to find new data. From Matlab's app generated scripts we see: % Closed Loop Network % Use this network to do multi-step prediction. % The function CLOSELOOP replaces the feedback input with a direct % connection from the output layer. netc = closeloop(net); netc.name = [net.name ' - Closed Loop']; view(netc) [xc,xic,aic,tc] = preparets(netc,{},{},T); yc = netc(xc,xic,aic); closedLoopPerformance = perform(net,tc,yc) % Step-Ahead Prediction Network % For some applications it helps to get the prediction a timestep early. % The original network returns predicted y(t+1) at the same time it is % given y(t+1). For some applications such as decision making, it would % help to have predicted y(t+1) once y(t) is available, but before the % actual y(t+1) occurs. The network can be made to return its output a % timestep early by removing one delay so that its minimal tap delay is now % 0 instead of 1. The new network returns the same outputs as the original % network, but outputs are shifted left one timestep. nets = removedelay(net); nets.name = [net.name ' - Predict One Step Ahead']; view(nets) [xs,xis,ais,ts] = preparets(nets,{},{},T); ys = nets(xs,xis,ais); stepAheadPerformance = perform(nets,ts,ys) My question is: What is the real difference between them? Can one uses them equivalently? If yes, why? I mean, even tho the structure or how they are equipp…

Stanford AI Lab Papers at ICCV 2021

The International Conference on Computer Vision (ICCV 2021) will be hosted virtually next week. We’re excited to share all the work from SAIL that will be 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 GLoRIA: A Multimodal Global-Local Representation Learning Framework for Label-efficient Medical Image Recognition Authors : Mars Huang Contact : mschuang@stanford.edu Keywords : medical image, self-supervised learning, multimodal fusion 3D Shape Generation and Completion Through Point-Voxel Diffusion Authors : Linqi Zhou, Yilun Du, Jiajun Wu Contact : linqizhou@stanford.edu Links: Paper | Video | Website Keywords : diffusion, shape generation CAPTRA: CAtegory-level Pose Tracking for Rigid and Articulated Objects from Point Clouds Authors : Yijia Weng*, He Wang*, Qiang Zhou, Yuzhe Qin, Yueqi Duan, Qingnan Fan, Baoquan Chen, Hao Su, Leonidas J. Guibas Contact : yijiaw@stanford.edu Award nominations: Oral Presentation Links: Paper | Video | Website Keywords : category-level object pose tracking, articulated objects Detecting Human-Object Relationships in Videos Authors : Jingwei Ji, Rishi Desai, Juan Carlos Niebles Contact : jingweij@cs.stanford.edu Links: Paper Keywords : human-object relationships, video, detection, transformer, spatio-temporal reasoning Geography-Aware Self-Supervised Learning Authors : Kumar Ayush, Bura…

R package to solve Gaussian MLE under conditional independence constraints

Is there any R package or function to solve Gaussian MLE under conditional independence constraints ? Suppose we have $y_i\overset{i.i.d}{\sim}\mathcal{N}(0,\Sigma_{p\times p})$ , $i = 1,2,\ldots,n$ . We know that $(\Sigma^{-1})_{ij} = 0$ , for some $i,j\in\{1,2,\ldots,p\}$ , i.e. $X_i$ and $X_j$ are conditionally independent given the rest of the variables. We would like to find MLE of $\Sigma$ under the conditional independence constraints. I also tried to implement it according to section 5.1 of the above paper, but I couldn't do it successfully. I wonder if there is any R implementation to find Gaussian MLE under conditional independence constraints?

How to Train Really Large Models on Many GPUs?

[Updated on 2022-03-13: add expert choice routing .] [Updated on 2022-06-10]: Greg and I wrote a shorted and upgraded version of this post, published on OpenAI Blog: “Techniques for Training Large Neural Networks”

Regressor contribution in OLS regression

Assume I have the following model, estimated using OLS: $Y_{it}=β0+β1∗X1_{it}+β2∗X2_{it}+β3∗X3_{it}+ϵ_{it}$ I know that some methods exist to compute the relative contribution of each variable to the overall goodness of fit (Shapley decomposition for instance). But is there any easy way to measure a time-varying contribution? I.e. the evolution over time of the contribution of each factor? Probably silly question: what does $(β1∗X1_{it})/Y_{it}$ represent? Can this represent some kind of contribution of $X_1$ to $Y$ ? Many thanks

How to understand the common practices followed for writing a "bounding box" for an image in datasets?

For the image datasets, there may be a bounding box for each image at the dataset. It is an annotation for an image. It is a rectangular box intended for focusing on something inside the image. I read about the following two types of representations for a bounding box. using two points $(x_1, y_1)$ and $(x_2, y_2)$ . $$ $$ Using a point $(x_1, y_1)$ , width, and height. $$ $$ How do understand both the representations, Specifically, does the point $(x_1, y_1)$ used to denote the top right or top left or bottom right or bottom left in both cases?

How to perform ANOVA for data measured with unequal time interval?

So I have a set of patient data divided into 3 groups based upon their vision. I also have data on few variables that I believe to be relevent which are measured pre-surgery, 1 day post-surgery, and 7-day post-surgery. I am trying to figure out if those variables differ from each other statistically within and between those 3 groups, and a simple F-test should do the trick. However, since the measured time period had uneven time interval (pre-surgery - 1 day post-surger had shorter time interval than 1 day to 7 day). Is there any other test I should perform? Also I am using R. I have read somewhere on how to deal with similar problem using SPSS, but the author was vague and didn't put any explanations on the reason behind. So if you guys can provide a simple procedure on how to perfom it using R it would be great ! Thanks a lot in advance !

A Gentle Introduction to Graph Neural Networks

What components are needed for building learning algorithms that leverage the structure and properties of graphs?

Understanding Convolutions on Graphs

Understanding the building blocks and design choices of graph neural networks.

What are the Calculus books recommended for beginner to advanced researchers in artificial intelligence?

Calculus is a branch of mathematics that primarily deals with the rate of change of outputs of a function w.r.t the inputs. It contains several concepts including limits, first-order derivatives, higher-order derivatives, chain rule, derivatives of special and standard functions, definite integrals, indefinite integrals, derivative tests, gradients, higher-order gradients, and so on... Calculus has been heavily used in optimization and maybe in several other aspects of artificial intelligence. What are the Calculus textbook(s) recommended that cover all the concepts required for a researcher in artificial intelligence?

Missing data in MLR RStudio

I'm really new to coding and to R so struggling to find an answer I can understand, or recognise if it is an answer. I'm hoping someone(s) can help. I'm trying to conduct a multiple linear regression in R. The data is from a survey of 561 people. Here's a sample of the first 12 lines of data. I want to look for the significanc between a positive attitude, which is a mean of 4 Likert-scale questions (dependent variable), and the explanatory variable of age, gender, user and format. I carried out the code all in one: mlrpositive.lm I noticed it was missing explanatory variables, for example in age it was missing '18-24', gender was missing 'female', user was missing 'no', and format missing 'Audio'. I carried out each one individually and got the same result. positive.gender.lm I don't know how to correct this.

Can you use recursive least squares (RLS) for mini batches?

For my application, I am considering a learning problem where I simulate a bunch of episodes, say ' $n$ ' first, and than carry out the recursive least squares update. Similar to $TD(1)$ . I know that RLS can be used to update parameters being learned as they arrive. This can be done efficiently for a single data point and the derivations are easily available online and also easy to understand. However, for my case, I am looking for the same equations when data arrives as a mini batch and not a single data point at a time. I could not find any material regarding RLS for mini batches. According to my understanding, the same equations can be also used by appropriately considering matrix dimensions. However, I do not know if this is valid. What are the alternatives to be used?

Smoothed CDF to calculate asymptotic normality

If we have the following estimator: $\hat{F_Z}(z)=\frac{1}{N}\sum_{i=1}^N1\{Z_i\leq z\}$ . The CDF of $Z$ is defined as $F_Z(z)=Pr(Z\leq z)$ . $Z_1, ..., Z_N$ is i.i.d. data. What would be the steps to show that $\hat{F}$ is consistent and asymptotically normal and to find the asymptotic variance at a given point $z$ . My thought was to get the pdf but apparently $\hat{F}$ is not very useful for estimating the PDF. That I don't understand why? I am not sure but this is what I got so far: $$\hat{F_Z}(z)=\frac{1}{N}\sum_{i=1}^N1\{Z_i\leq z\}\xrightarrow{LLN}E[1(Z_i\leq z)]=Pr(Z\leq z)=F_Z(z)$$ $$\sqrt{N}(N^{-1}\sum 1(Z_i\leq z)-E[1(Z_i\leq z))\xrightarrow{CLT}N(0, Var(1(Z\leq z)))$$ The questions are related so I merged them together but if needed I can post a new one. If we would consider now a random variable U independent of Z with CDF $F_U(\cdot)$ , and a symmetric PDF, and consider some $h > 0$ . We now consider a different estimator $$\tilde{F}(z)=\frac{1}{N}\sum_{i=1}^NF_U[\frac{z-Z_i}{h}]$$ Is $\tilde{F}$ consistent for $F_Z$ ? For this, I guess we would be using the nonparametric approach. Would we follow the same steps as with the previous one or is it a different thing?