Original article was published by Jesus Rodriguez on Artificial Intelligence on Medium
The Sequence Scope: The Friction Between Privacy and Learning
Weekly newsletter that discusses impactful ML research papers, cool tech releases, the money in AI, and real-life implementations.
The Sequence Scope is a summary of the most important published research papers, released technology and startup news in the AI ecosystem in the last week. This compendium is part of TheSequence newsletter. Data scientists, scholars, and developers from Microsoft Research, Intel Corporation, Linux Foundation AI, Google, Lockheed Martin, Cardiff University, Mellon College of Science, Warsaw University of Technology, Universitat Politècnica de València and other companies and universities are already subscribed to TheSequence.
In the current issues of TheSequence Edge, we have been covering different topics related to security in machine learning models. Security and privacy are the aspects of machine learning solutions that are often ignored until they become a problem. In some contexts, nobody can dispute the importance of preserving privacy in training datasets in machine learning models. However, it is important to realize that, very often, introducing privacy methods creates friction in the learning process of machine learning models.
The friction between privacy and learning is conceptually trivial to understand. We shouldn’t expect a model trained in a clear dataset to perform identically to a model trained using processes such as differential privacy or secured multi-party computations. Those techniques require very unique architectures in order to enforce privacy without affecting the performance of the target machine learning model. Furthermore, frameworks for private machine learning are still in very early stages, requiring high degrees of expertise in order to be applied correctly. From a practical experience standpoint, the only way to build effective private machine learning solutions is to start from day one with privacy as a first-class component of your neural network architecture.
Otherwise, you will always be balancing the friction between privacy and learning.
Did you come across problems with security and privacy in machine learning? What solutions can you suggest?
Leave a comment
🔺🔻TheSequence Scope — our Sunday edition with the industry’s development overview — is free. To receive high-quality educational content every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻
🗓 Next week in TheSequence Edge:
Edge#31: the concept of differential privacy; Apple’s research about differential privacy at scale, the TensorFlow Privacy framework.
Edge#32: the concept of adversarial attacks; OpenAI’s metric for the robustness against adversarial attacks; IBM’s adversarial robustness toolbox to protect neural networks against security attacks.
Now, let’s review the most important developments in the AI industry this week.
🔎 ML Research
DeepMind published a paper proposing a faster approach to reinforcement learning, by decomposing a complex task into smaller tasks with individual reward functions ->read more on DeepMind blog
Gendered Correlations in Language Models
Google Research published a paper proposing best practices to derive correlations between topics based on gender ->read more on Google Research blog
Singing Synthesis System
Amazon Research published a paper proposing an attention-based technique for synthesizing speech->read more on Amazon Research blog
🤖 Cool AI Tech Releases
Facebook published an insightful blog post detailing the architecture behind Nemo, a platform that powers data discovery at Facebook while leveraging natural language processing ->read more on the Facebook engineering team’s blog
LinkedIn engineering published a blog post outlining the architecture behind Pensieve, its feature embedding platform ->read more on their blog
Spotify open-sourced Klio, a framework for audio data processing->read more on Klio’s GitHub page
💸 Money in AI
- Telemedicine startup 98point6 raised $118 million in a venture funding round. The platform pairs AI and machine learning with the expertise of certified physicians. The majority of conversations with patients is held by a digital assistant enabled with natural language processing. It decides what information is the most important to be gathered from a patient. The platform accumulates data from each visit and constantly improves the perception and reaction of the assistant. The company underlines that diagnosis is always made by a human, and not by AI.
- Real estate tech company Snapdocs raised $60 million in a funding round. All activities are performed in one cloud-based service, and it allows Snapdocs to use AI for prompt analysis of the massive amounts of information and gain insights, such as compliance risks, team workflows performance and document errors. They also automate everything related to the signing process.
- AI-based machine health solution provider Augury raised $55 million in a funding round. They substitute the traditional time-based processes of machine checking and maintenance with AI tools that “look” into machine functions and give accurate insights about actual conditions of a given machine.
- AI-powered hearing aid company Whisper raised $35 million in a Series B round. The difference from other hearing aid devices is in an external pocket-sized box called Whisper Brain. It works wirelessly and is equipped with a proprietary engine that leverages AI for sound separation.
- AI-powered edge intelligence startup Infiot came out from stealth mode with $15 million in funding. Infiot leverages distributed processing and AI to locally process data, gain actionable insights, and deliver real-time decisions, improving availability while enabling superior data governance of Internet of Things devices.
- AI-powered data annotation and data lifecycle management startup Dataloop raised an additional $11 million to the previous round of $5 million. The proprietary engine helps manage the entire data life cycle for ML&AI projects, including annotating the datasets, automating data ops, customizing production pipelines and weaving the human-in-the-loop for data validation.
- Privacy-preserving news recommendation app River raised $10.4 million in a funding round. Using algorithms, the platform aggregates events and news from credible sources around the globe and delivers real-time recommendations based on trending stories. The main difference? Anonymity. No sign-ups, no access to social networks is required. The team claims that the app also leverages “content understanding” technology to gain insight into content and publishers, and “understands what each piece of content is about, who created it, and how it fits into the broader landscape.”
- Conversational platform Balto raised$10 million in seed funding. It leverages AI solutions by listening to both sides of a call and delivering critical insights and feedback to sales representatives in real-time. Data helps managers understand instantly what works better and make changes swiftly.