May 2019 Summaries

The 2019 O’Reilly AI Conference in New York City highlighted significant discussions on AI's evolving role in business, with notable contributions from experts like Matt Zeiler, CEO of Clarifai. Zeiler's talk addressed the challenge of AI systems losing accuracy over time, emphasizing the importance of implementing feedback loops to maintain and enhance AI performance. Feedback loops involve reusing an AI model's predicted outputs as training data for new model versions, thereby reinforcing the model's learning and accuracy. This process is likened to a teacher correcting students' work, where both auto-labeled and human-verified data are fed back into the system. The feedback loop ensures AI models continuously learn and adapt, preventing performance stagnation by aligning training data with real-world conditions that matter to customers.

Deep Learning (DL) is a subset of Machine Learning that involves using hierarchical neural networks to extract complex features from data, with networks like Deep Belief Networks, Convolutional Neural Networks, and Recurrent Neural Networks being notable examples. Unlike traditional "shallow" learning, where features are manually selected, DL systems automatically learn features through processes such as back propagation across multiple layers, enabling them to recognize high-level patterns more effectively. Although theoretically, a single-layer perceptron can solve any problem, DL's advantage lies in its ability to use fewer neurons across multiple layers to achieve the same results, avoiding the impracticality of using an exponentially large single layer. However, DL presents challenges such as the need for large datasets, significant computational power, and difficulties in training and optimization. Vendors like Clarifai can help overcome these hurdles by providing access to extensive datasets, powerful cloud-based computational resources, and expert teams to enhance model performance, thereby simplifying the process of building effective DL models.

Terrabeasts is a mobile game developed by a team of passionate engineers and designers who met at the NASA Space Hacks Hackathon in 2017, utilizing Clarifai's image recognition technology to encourage ecological eating habits through gamification. The game features digital alien pets that evolve based on the carbon footprint of the food players photograph and "feed" them, teaching users about the environmental impact of their eating choices, such as the higher carbon footprint of steak compared to chicken or fish. The team, comprising software engineers Alex Zaman, Skanda Mohan, Tim Shin, strategist Gloria Chow, and designer Michelle Ng, created Terrabeasts to address global warming concerns by engaging players in a fun and educational way, inspired by virtual pets like Tamagotchis. The Clarifai Predict API was crucial in the app's development, enabling seamless food recognition and reducing the need for manual logging, making the game more enjoyable. The team envisions expanding Terrabeasts to integrate with social media and potentially develop a wearable app that promotes healthy habits beneficial for both users and the planet.