Reading Minds and Machines

Introduction

Can we decode the training data of a Deep Neural Network (DNN) directly from its parameters? Training data of DNNs are assumed safe. Recent findings by us and by others indicate that this is not the case, with severe implications on data privacy. Yet, such findings shed light on why DNNs perform so well.

Exploring Brain Activity

On a different front: Can we decode what a person saw/heard/thinks directly from their brain activity? This may have huge benefits, such as:

• Communicating with “locked-in” patients
• Exploring dreams
• Developing man-machine interfaces
• Enhancing our understanding of the human brain

There is no risk of violating human privacy here, as thoughts do not “float” in the air and a person’s collaboration is required.

Interconnected Questions

Each of those two questions is intriguing on its own, with far-reaching implications. Despite the inherent differences between human brains and DNNs, they also have much in common. Exploring the two in tandem can lead to significant breakthroughs in both fields.

Recent Advancements

Recent advancements in both areas, with the incorporation of Deep Learning (DL) tools to analyze brain activity, open the door to explore the two jointly. Our expertise in both domains will enable explicit encoding/decoding between brain activity and DNN activations, allowing us to directly learn/infer from one about the other.

Ambitious Goals

Initial explorations indicate that our proposed goals, although ambitious, are within reach. Our intermediate goals in each domain are worthwhile on their own, forming a strong safety net.

Expected Outcomes

Expected outcomes include:

1. Deep data privacy
2. Insights on DNNs, their generalization, and vulnerabilities
3. Insights on “what is encoded where” in the brain
4. New scientific tools for brain scientists to explore the brain
5. Allowing “locked-in” (ALS) patients to communicate their thoughts/needs
6. Using brain scanning to improve DNNs and DNNs to improve brain scanning

Project Requirements

Our project requires no human subjects nor brain science expertise; only publicly available datasets. All methods lie in computer vision and DL, with impact on both DL and brain science.