NEWS COMMENTARY

Researchers at MIT and IBM assemble new object recognition dataset that highlights the limitations of current AI techniques

Published:
December 16, 2019
Coverage:
Digital Transformation More...
Activities:
Research
Average importance

The dataset consisted of an assortment of objects shot at unusual angles and surrounded by clutter. While computer vision performance has increased significantly over the past decade, a top-performing algorithm's accuracy rates fell from a high of 97% on more simple datasets to just 50% to 55% on the new dataset. This study highlights a major challenge for AI, namely, that as the environment becomes more complex and unpredictable, machine learning performance drops off considerably. While clients could wait for the algorithms to become more robust before deploying a solution, a more near-term solution to this challenge is to consider constraining a problem to make it a better fit for today's tools.

For the original news article, click here .


Further Reading

The AI Hardware Landscape Update

State of the Market Report | June 25, 2020

Technical, financial, and societal factors have driven innovation activity in the AI hardware space to an all‑time high. Such drivers include new applications with low latency demands, a massive increase in compute requirements to train and run modern AI models, limited network bandwidth, and ... Not part of subscription

GPT-3 is overhyped for practical applications, but large-scale models point to the future of AI

Analyst Insight | August 04, 2020

In late May, OpenAI released a paper describing GPT‑3, a much larger version of its previous natural language processing (NLP) language models, GPT and GPT‑2, that was trained on a large archive of text data found on the internet. As we've previously described, what was notable about the model was ... Not part of subscription

Strategies for managing the inaccuracy of AI and machine learning systems

Analyst Insight | October 14, 2020

One feature common to almost all AI and machine learning software is inaccuracy. Unlike traditional software, which is rule‑based and deterministic, modern machine learning‑based AI is probabilistic and can be unpredictable, leading to systems that provide inaccurate results a nontrivial percentage ... Not part of subscription