AI Scholar Weekly - Issue #46



Subscribe to our newsletter

By subscribing, you agree with Revue’s Terms of Service and Privacy Policy and understand that AI Scholar Weekly will receive your email address.

AI Scholar Weekly - Issue #46
By Educate AI  • Issue #46 • View online
Top 6 Cheat Sheets Novice ML Engineers Need; Transformer Models in Computer Vision; How to Detect Facial Skin Problems; A Unified, Modularized, and Extensible Framework for Text Generation; Google AI ~ Using ML to Better Emulate Professional-looking Portraits (and more)

AI Success Stories
Company Implements AI, Generates $2M in Sales and more …. in Less than 8 Months
Goal: Pinpoint new prospects as well as re-engage qualified but stale leads to increase monthly revenue  
Challenge: Marketers know how well it can be a pain point when you get leads but they end up being stale leads. That was SecureAuth’s greatest challenge! Most of the leads they got tended to be very early stage and went stale even when we tweaked in a lead nurturing program. Even more challenging is that the leads they thought were qualified because they achieved a certain activity-based score were getting kicked back to Marketing because prospects weren’t fully engaging with their inside Sales Reps who couldn’t get them to take a phone call or respond to their emails. They needed to find a way to ensure consistent, continuous qualification and engagement of their leads. 
Results: With AI working ‘miracles ’ in many successful businesses, the company followed suit and worked with a team that helped them design and implement an Automated AI assistant that engages potential customers in natural, human-like interactions. The results?  
  • Generated $2M in the new pipeline in less than 8 months
  • Created 25 new opportunities worth $1.5M in the pipeline from re-engaged prospects
  • Closed $241K in new business
PS: If you have a similar business need or want similar success, don’t hesitate, Sign up for a 30-minute consultation and learn how AI can upgrade your business today. It’s eye-opening, impactful, empowering, and worth your time! 
Top AI Research This Week
#1 How to classify any video in the wild as fake or real (Proposed model achieves 89.79% accuracy on FaceForensics++, 80.0 % on Deep Fakes, and 88.35% on CelebDF datasets) 
Due to the popularity and access of deep fake technology (synthetic media produced by AI), it’s hard to determine whether specific audio you have listened to or a video you have just watched is real or, in fact, fake. Recognizing such media to the untrained human eye. is becoming increasingly problematic, if not impossible. Researchers have been working on creating deepfake detectors that tell the difference between fake and real media. 
In this paper, AI researchers conducted the first of its kind and extensive deep fake gaze analysis. It is the first approach to build a detector solely based on holistic eye and gaze features. They evaluate the suggested method on various datasets, compare it against other architectures, and experiment with several ablation studies (removing some “feature” of a model to see how that affects performance).  
They compile features into signatures, analyze and compare real and fake videos, and formulating geometric, visual, metric, temporal, and spectral variations. 
They also generalize this formulation to a deepfake detection problem by a deep neural network to classify any video in the wild as fake or real. 
The new approach achieves 89.79% accuracy on FaceForensics++, 80.0% on Deep Fakes (in the wild), and 88.35% on CelebDF datasets. 
Why this approach is useful: Research findings can be used to create better deep fakes with consistent fake gazes, which, with increased photorealism, can be useful for virtual/augmented reality applications, avatars, data augmentation, transfer learning, and controlled learning and testing environments for eye-tracking research. 
#2 Just How Effective are Alternative Data Sources for Developing Credit Risk Models? 
In this work, scholars present the effect of alternative data that originates from an app-based marketplace, in contrast to traditional bureau data, upon credit scoring models.
These alternative data sources have shown themselves to be immensely powerful in predicting borrower behavior in segments traditionally underserved by banks and financial institutions. Obtained results show that these new sources of data are particularly useful for predicting financial behavior in low-wealth and young individuals, who are also the most likely to engage with alternative lenders.
Moreover, using the TreeSHAP method for Stochastic Gradient Boosting interpretation, results revealed interesting non-linear trends in the variables originating from the app, which would not normally be available to traditional banks. Results suggest that these apps are useful contributions to financial inclusion, therefore, regulatory efforts should also proceed in this direction.
Why it matters? It represents an opportunity for technology companies to disrupt traditional banking by correctly identifying alternative data sources and handling this new information properly. At the same time, alternative data must be carefully validated to overcome regulatory hurdles across diverse jurisdictions.
#3 A Comprehensive Overview of the Transformer Models in Computer Vision
Transformer models have recently demonstrated exemplary performance on a broad range of language tasks including text classification, machine translation, and question answering.
Assuming little to no prior background in the field, this paper aims to provide a comprehensive overview of the transformer models in the computer vision discipline.
Researchers start with an introduction to fundamental concepts behind the success of transformer models i.e., self-supervision and self-attention.
Attention has played a key role in delivering efficient and accurate computer vision systems, while simultaneously providing insights into the function of deep neural networks
In the study, the researchers describe the state of the art self-attention models for image recognition, object detection, semantic and instance segmentation, video analysis, and classification, visual question answering, visual commonsense reasoning, image captioning, vision language navigation, clustering, few-shot learning, and 3D data analysis. They systematically highlight the key strengths and limitations of the existing methods and particularly elaborate on the important future research directions.
Why is this useful? The survey provides a unique view of the recent progress in self-attention and transformer-based methods with regards to computer vision tasks. Researchers hope this effort will drive further interest in the vision community to leverage the potential of transformer models and improve on their current limitations e.g., reducing their carbon footprint.
#4 Object-by-Object Learning for Detecting Facial Skin Problems
Semantic segmentation has been one of the fundamental and active topics in computer vision for a long time. This topic is of wide interest for real-world applications of autonomous driving, robotics, and a range of medical imaging applications.
To this end, this paper proposes successfully an efficient network architecture to address detecting multi-type facial skin lesions by a novel object-by-object learning technique.
Scholars introduce a new concept named “object-by-object” learning, where an object can be identified by looking at other objects. They propose a new residual block called REthinker modules that support the ‘object-by-object’ technique by capturing contextual relationships between object classes. In connection to that, they developed a new RethNet architecture that detects skin lesions with higher accuracy than the recent state-of-the-art segmentation approaches.
Results: show that the proposed model outperformed state-of-the-art segmentation networks by a high gap in the MSLD dataset. Furthermore, the new model takes promising results on the ISIC 2018 segmentation task. In the future, they plan to consider the time complexity of Rethinker blocks and try to design more lightweight models.
#5 A Unified, Modularized, and Extensible Framework for Text Generation
To facilitate the building of the next text generation models, a few remarkable open-source libraries have been developed.
TextBox is a new open-source library, which provides a unified, modularized, and extensible text generation framework. TextBox aims at enhancing the reproducibility of existing models, standardize the implementation and evaluation protocol of text generation algorithms, and ease the development process of new algorithms.
It implements 15 text generation models, including VAE-based, GAN-based, RNNbased Transformer-based models and pre-trained language models, and 6 benchmark datasets for unconditional and conditional text generation tasks. Moreover, the TextBox is modularized to easily plug-in or swap out components, and extensible to support seamless incorporation of other external modules. In the future, we would further add more models and datasets and consider more functions for covering more text generation tasks.
TextBox is implemented based on PyTorch and can be used to support several real-world applications in the field of text generation.
Why is this important? It is especially suitable for researchers and practitioners to efficiently reproduce baseline models and develop new models.
Read the paper: Introducing TextBox
Other Great AI Research Papers
Google AI: Photographers can now use machine learning to better emulate professional-looking portraits with Portrait Light, a new post-capture feature for the Pixel Camera and Google Photos apps that add a simulated directional light source to portraits, with the directionality and intensity set to complement the lighting from the original photo. Read more
SmartDeal is a new model that can effectively resolve the practical bottleneck of high-cost memory storage/access and can aggressively trim down energy cost and model size while incurring minimal accuracy drops, for both inference and training. Read more
Objects in a 3D world do not follow any particular orientation, and box-based detectors have difficulties enumerating all orientations or fitting an axis-aligned bounding box to rotated objects. To address the challenge, researchers propose to represent, detect, and track 3D objects as points. The resulting detection and tracking algorithm is simple, efficient, effective, and outperforms previous models. The code and pre-trained models are available on Github.
State-of-the-art question answering models are unreasonably heavy and need expensive GPU computing to answer questions in a reasonable time. Now researchers have unveiled a new model that beats the previous models by 1.3 points. Read more
This new CNN-based method that unifies gesture recognition and prediction of fingertip position in a single-step process. The proposed algorithm is significantly fast in computation and can play a significant role in the HCI, VR, and MR applications. Read more
AI Resources
As a Machine Learning engineer, cheat sheets go a long way. These six cheat sheets have quick, useful references for libraries and functions that you can use on a daily basis as a machine learning engineer. Go here
Many AI programs do not generate business gains. Here are three things you need to develop successful AI programs. Read more
Top AI News
Researchers recently found out that machine learning models still struggle to detect hate speech. One of the reasons is because harmful speeches come in many different forms, and models must learn to differentiate each one from innocuous turns of phrase. Read story
OpenAI recently launches DALL-E an AI app that can create an image out of nearly any text or CLIP description. DALL-E has significant potential to disrupt fields like stock photography and illustration, with all the good and bad. Read more
Working together with an international team, researchers at the University of Münster show that so-called photonic processors, with which data is processed by means of light, can process information much more rapidly and in parallel – something electronic chips are incapable of doing. Read more
AI Scholar Weekly
 Thanks for reading! If this newsletter lights you up, create a ripple effect by sharing it with someone else, and they will also be lit up! And, if you have suggestions, comments, or other thoughts, we would love to hear from you, email me at, tweet at @cdossman, like us on Facebook, or connect with me on LinkedIn
Did you enjoy this issue?
Educate AI

AI Scholar Weekly brings you everything new and exciting in the world of Artificial Intelligence, Machine Learning, and Deep Learning every week for free.

In order to unsubscribe, click here.
If you were forwarded this newsletter and you like it, you can subscribe here.
Powered by Revue