Ensuring that Models do not Treat Individuals Unfairly due to Biases in the Data and Model Inaccuracies
Last year, October 2019, a group of researchers in the US found that a health care risk-prediction algorithm for identifying which patients would benefit from “high-risk care management, favored white patients over black patients. Before that in 2015, Amazon realized that an algorithm they used for hiring employees was biased against women. These are classical examples of bias in ML as a result of relying on a faulty metric.
ML-based systems are reaching society at large and in many aspects of everyday life. The examples above show the importance of creating non-biased algorithms. It is relevant from an ethical and legal point of view to ensure that these algorithms do not discriminate based on sensitive social matters.
This new research aims to extend progress towards achieving fairness in ML systems. Researchers have discussed an emerging area of machine learning that studies the development of techniques for ensuring that models do not treat individuals unfairly due to biases in the data and model inaccuracies. According to them, data that are used has to be bias-free and the engineers that are creating these algorithms need to make sure they’re not leaking any of their own biases in the algorithms. Statistical tools and methods are then required to detect and eliminate such potential biases. The paper also discusses legal restrictions with the use of sensitive attributes and introduces an in-processing approach that does not require the use of sensitive attributes during the deployment of the model.
Two Methods that Can Be Combined to Produce Speedier and More Accurate Language Models
Researchers with Facebook AI in collaboration with Allen Institute for AI challenge the conventional wisdom that scaling transformer language models to longer sequences improves results in this paper. They also introduce new techniques based on shorter input subsequences that improve both efficiency and perplexity.
While previous methods require computationally expensive relative position embeddings; their work introduces simple methodologies involving the addition of absolute position embeddings to queries and keys instead of to word embeddings, which efficiently produces superior results.
By combining the proposed techniques, they manage to increase training speed by 65%, make generation nine times faster.
Neural Machine Translation: A Review of Methods, Resources, and Tools
Machine translation (MT) is an important sub-field of natural language processing that aims to translate natural languages using computers. Now Neural Machine Translation, an approach that uses an artificial neural network to predict the likelihood of a sequence of words, typically modeling entire sentences in a single integrated model has become the dominant approach to machine translation in both research and practice.
If you are into Neural Machine Translation, this paper reviews the widely used methods in NMT, including modeling, decoding, data augmentation, interpretation, as well as evaluation. Researchers then summarize the resources and tools that are useful for NMT research and current problems that need to be explored for researchers in the field.
Maximizing Online Model Performance without Going Online
Learning-to-rank (LTR) has become a key technology in E-commerce applications. Currently, most existing LTR approaches follow a supervised learning paradigm from offline labeled data collected from the online system. They achieve a good validation performance over offline validation data but have a poor online performance, implying a possible large inconsistency between the offline and online evaluation.
This paper presents an Evaluator-Generator Reranking (EG-Rerank) framework for E-commerce LTR systems to address the offline-online inconsistency problem while avoiding the pitfalls of direct online interaction-based evaluation. EG-Rerank consists of an evaluator and a generator. The evaluator aims at modeling user feedback (i.e. predicting the probabilities of purchase given a list).
In AliExpress Search, EG-Rerank+ consistently improves the conversion rate by 2% over the fine-tuned industrial-level re-ranking model in online A/B tests, which translates into a significant improvement of business profits.
This New Deep Transfer Learning Model is your Best Bet for Real-time Robotic Applications According to Research
It’s easy for humans to adapt to different environments dynamically by watching and learning about new object categories. Conversely, migrating a robot to a new environment often requires one to completely re-program its knowledge base. This explains the reason open-ended object category learning ( the ability to learn new object categories sequentially without forgetting the previously learned categories) increasingly becoming important. While there’s a notable improvement in the field, challenges still remain.
Researchers recently proposed OrthographicNet, a new model they claim can achieve performance better than the selected state-of-the-art and is suited for real-time robotic applications.
Specifically, OrthographicNet is a Convolutional Neural Network (CNN)-based model for 3D object recognition in open-ended domains. It generates a global rotation- and scale-invariant representation for a given 3D object, enabling robots to recognize the same or similar objects seen from different perspectives.
Why it matters? Experimental results show that the new approach yields significant improvements over the previous state-of-the-art approaches concerning object recognition performance and scalability in open-ended scenarios. The approach also provides a good tradeoff between descriptiveness, computation time, and memory usage, allowing concurrent object recognition and pose estimation.