Efficient Convolutional Learning and Presenting Data to Neural Networks

1min Snip

00:00

Play full episode

Key takeaways

Transcript

Episode notes

Neural networks are like a brain and need to be presented with simple data first before complex data.
Sequence length warm up can help in presenting smaller sequences and learning primitives before presenting bigger data.
Combination of presenting data in specific ways can make convolutional learning more efficient.
There are still interesting things that the brain does that neural networks don't do today.

There are many speculations about the future of artificial intelligence (AI), and in this episode, we hear the opinions and predictions of a player in the inner folds of the AI space. Naveen Rao is the CEO and Co-Founder of the machine learning (ML) training platform, MosaicML, and the former CEO and Co-Founder of Nervana Systems. Naveen shares insight into the thesis behind Mosaic and the practical applications of Large Language Models (LLMs), as well as the Generative Pre-trained Transformer-2 (GPT-2) to GPT-3 transition and the challenge of training models with the constraint of data limits. He predicts the evolution of AI models in terms of quantity, size, and function, and the future of computers in general. If you’re curious as to whether it makes sense to build smaller models or own your own model, this episode is for you. Tune in to hear Naveen’s opinions on the impact of AI on the economy, the danger of centralizing resources, what constitutes sentience, and much more.

“What we’re doing at Mosaic is building tools to enable more people to have access to these technologies. When it’s all centralized in one or two or three players, that creates a huge power dynamic.” — @NaveenGRao Key Points From This Episode:

Naveen Rao’s educational background and interest in synthetic intelligence.
What led him to start his first AI company, Nervana Systems.
The thesis behind MosaicML.
What MosaicML offers customers.
What Naveen considers to be 2022’s most exciting breakthrough in AI.
The innovation of ChatGPT.
The GPT-2 to GPT-3 transition.
The challenge of training models with the constraint of data limits.
Naveen explains the concept of synthetic data.
He predicts the evolution of AI models in terms of quantity, size, and function.
Why it makes sense to build smaller models and own your own model where possible.
Data as a moat component.
Practical applications of LLMs.
Naveen’s opinion on whether AI will disrupt the economy or increase Gross Domestic Product (GDP).
The danger of the centralization of resources.
How MosaicML is making trading more efficient given the limits AI is facing.
The efficiency improvements MosaicML customers are seeking out.
Naveen’s prediction for the rate of cost decline.
The history of the computer password.
The future of computers.
The question of what constitutes sentience.
Naveen recounts the acquisition process of selling Nervana Systems to Intel.
How the innovator’s dilemma will play out among competitors in the AI space.
Naveen’s advice for his past self.