A growing number of consumer devices, including smart speakers, headphones, and watches, use speech as the primary means of user input. As a result, voice trigger detection systems—a mechanism that uses voice recognition technology to control access to a particular device or feature—have become an important component of the user interaction pipeline as they signal the start of an interaction between the user and a device. Since these systems are deployed entirely on-device, several considerations inform their design, like privacy, latency, accuracy, and power consumption.

Voice Trigger System for Siri

The Photos app learns about significant people, places, and events based on the user’s library, and then presents Memories: curated collections of photos and videos set to music. The key photo for a Memory is influenced by the popularity of iconic scenes learned from iOS users—with DP assurance.

<Figure id="figure1" title="Figure 1" className="inline" caption="Figure 1: This feature automatically picks one photo from the Memory as the key image. The algorithm is powered, among other signals, by iconic scenes crowd-learned with differential privacy.">
<Video src="https://mlr.cdn-apple.com/video/Fig_1_screens_10bd45bc93.mp4" autoPlay loop />
</Figure>

We prioritized three key aspects of the machine learning research powering this feature: 

1. How to accommodate data changes given its sensitivity
2. How to navigate the tradeoffs between local and central differential privacy
3. How to accommodate non-uniform data density

**Data changes.** Identifying iconic scenes—like popular locations such as Central Park in New York City or the Golden Gate Bridge in San Francisco—and the type of photos people take at those locations allows for better key photo selection. Even so, we want to make Memories a compelling feature for device owners no matter where they are located. In addition, iconic scenes can change over time or depending on the season. For example, if a new dog park opens near the Golden Gate Bridge, people might start taking more photos of dogs in that location. Or, if it is winter, people might be taking pictures of the ice skating rink in Central Park. Identifying a new category of photos in a popular location is a perfect task for a data-driven solution. To preserve user privacy, we want to keep information about where people go and what they see there on their device. 

**Local versus central differential privacy.** DP was designed to learn statistics about data with strong assurances of not leaking personally identifiable information. With these considerations in mind, we started with local DP solutions, as seen in the paper [Learning with Privacy at Scale,](https://docs-assets.developer.apple.com/ml-research/papers/learning-with-privacy-at-scale.pdf) that implement sufficient protections in the operating system so that researchers could verify our claims by reverse-engineering the code. However, we soon discovered that this approach was limited since local DP requires a significant amount of noise to be added, allowing us to discover only the most prominent signal. We needed a better solution that could provide a transparent and verifiable DP assurance while also improving the utility of the learned histograms.

**Nonuniform data density.** Data gathered across the world is not uniformly distributed, so the privacy challenges differ across regions. For example, suppose many users take photos in a particular location. This makes learning detailed statistics with high precision and privacy assurances easier and provides more useful results. If fewer users visit other locations and take fewer photos, we have less data to work from. This makes it more difficult to glean a meaningful signal without jeopardizing user privacy. In high-density areas, we could learn better precision (with the same privacy assurance) because there is a bigger crowd to hide in. But we don’t know the distribution (that is, which areas are high or low density).

## Balancing Privacy with Utility

We combined local noise addition with a technique called *secure aggregation* to address these concerns of balancing privacy with utility. To understand how balancing privacy with utility works, let’s dive into a Photos Memories use case. 

A user takes a photo in a place they visit. The photo is annotated with common categories, such as recreation, person, sky, and so on. The model that assigns those categories runs locally on the user’s device, as described in [A Multi-Task Neural Architecture for On-Device Scene Analysis.](https://machinelearning.apple.com/research/on-device-scene-analysis) 

If the user opts in to the Analytics & Improvements feature, and enables Location Services and precise location stored on photos, we randomly pick one location-category pair (like *Central Park (New York)* and *person*) and encode it into a one-hot vector, as described in [Figure 2](#figure2).

Now, we take this one-hot vector and flip each bit with some probability. The noise introduced by flipped bits provides a local DP assurance, which is later amplified through secure aggregation.

<Figure id="figure2" theme="dark-gray" className="full-bleed" title="Figure 2" caption="Figure 2: On the user’s device, one randomly selected location-category pair gets encoded into a one-hot vector. Each tile on the map shown here corresponds to a segment within the vector. Each segment contains one bit representing each of the categories. Then each of the bits is randomly flipped with some probability to satisfy the DP assurance.">
 <Figure2 title="Encoding Location-Category as One-Hot Vector" />
</Figure>

Next, we split this binary vector (one-hot vector with random noise) into two shares. Each share on its own is meaningless noise, but when combined with the other, we get the noised vector back. Each share is encrypted with a different public key and uploaded to the server.

<Figure id="figure3" className="inline" title="Figure 3" caption="Figure 3: The secure aggregation protocol that enforces the DP assurance. Each binary vector is split into two shares, and each share is encrypted with a different public key. As long as no single entity gets access to both private keys, nobody can see any individual vectors, only the aggregate, which satisfies the desired DP assurance.">
 <Image 
 src="https://mlr.cdn-apple.com/media/Fig_3_pipeline_38eaf7f2d0.png"
 />
</Figure>

On the server side, there are two main components, as shown in [Figure 3](#figure3): 

- **Leader** has one private key to decrypt, and it aggregates the first shares from the given cohort of devices. 
- **Helper** has the other private key to decrypt, and it aggregates the second shares from the given cohort of devices. 

Both the helper and leader components publish their aggregate only if it satisfies a minimal cohort size (the minimal number of devices contributing their shares). As long as no single entity has access to both private keys, nobody can see the original one-hot vector from any single device. They would see only the aggregates, which satisfies the DP assurance. This DP assurance is enforced through local noise (added on the user device), which gets amplified by secure aggregation, as described in [Private Federated Statistics in an Interactive Setting](https://arxiv.org/abs/2211.10082).

Finally, both aggregates from the helper and leader functions are combined to get a noisy aggregate vector from all the devices within the cohort. This vector is decoded back into the location and photo categories, which can then be visualized on a map, as shown in [Figure 4.](#figure4)

We needed a solution to address one more complication: Detecting malicious updates would be difficult because nobody can see any single vector. For example, a malicious user might submit vectors to the server that would poison the final histogram. We surmounted this obstacle using Prio validation, as discussed in the work [Prio: Private, Robust, and Scalable Computation of Aggregate Statistics.](https://arxiv.org/abs/1703.06255) 

<Figure id="figure4" className="" title="Figure 4" caption="Figure 4: Map showing final normalized histograms of top categories for photos taken near the Brooklyn Bridge in New York City.">
 <Image 
 src="https://mlr.cdn-apple.com/media/Fig_4_histograms_c0049fb8b0.jpeg" rounded
 darkgray
 />
</Figure>

# Providing Iconic Scenes Frequencies 

We discovered frequencies for 4.5 million location-category pairs for 1.5 million unique locations and 100 categories using this approach. These frequencies have been powering ML selections of Memories key photos for millions of device owners worldwide since iOS&nbsp;16, as well as a ranking of photos and locations in Places Map in iOS&nbsp;17.

The learned histograms satisfy DP $epsilon=1, delta=1.5e^{-7}$. We used a cohort size of 150,000. To avoid dropping histograms with smaller populations, we add additional noise in the secure aggregation protocol to achieve a DP of $epsilon=1$. To ensure our strong privacy assurance holds without requiring that users trust either the helper or the leader component, this noise is added independently by both the helper and leader components.

# Building on Our Wins and Looking Forward

Privacy-preserving ML research has already improved user experiences, and this is just the beginning. There are many opportunities to improve our system further, including:

- Privacy accounting and transparency
- Better algorithms
- Other data science tools

**Privacy accounting and transparency.** The current approach provides a formal DP assurance for a given histogram. However, DP is composable, and that’s an opportunity to provide an even more precise privacy assurance—one that is not specific to a single task or user data. One of the next steps is exploring formal accounting methods for precise accounting across multiple histograms. 

The current approach is compatible with the [Distributed Aggregation Protocol](https://datatracker.ietf.org/doc/draft-ietf-ppm-dap/) developed by the IETF working group, which is publicly available. We would like to open source key components of the system to make it easier, to verify the correctness of our implementation.

**Better algorithms.** The data is not uniformly distributed. For instance, many people might take pictures in one location, and very few might take photos in other locations. We are limited to fixed precision to provide a consistent privacy assurance with our current approach, which can’t be optimal for all locations. An iterative approach of gradually traversing the data and drilling into more popular places with higher precision will likely bring better results. 

Each device has hundreds or thousands of photos with multiple categories, but we pick only one random category from one randomly sampled historical photo. There is an opportunity to develop efficient algorithms for encoding and aggregating multiple location-category pairs to produce more precise histograms while maintaining the same privacy assurance.

**Other data science tools.** Histograms are very powerful, but to build even better experiences with privacy-preserving ML, we must provide a full set of data science tools. For example, for tasks that have good supervision, gradient-based algorithms might be powerful, along with tools for rigorous statistical testing to compare different solutions. And for tasks where no supervision is available, unsupervised algorithms such as k-means clustering might be needed. Many of those algorithms can be implemented using histograms, but custom solutions often bring significantly better utility, which is key for building better user experiences.

# Conclusion

Our goal is to provide an interpretable privacy assurance to our users, in a transparent way. As more users worldwide contribute their photos of iconic scenes in more locations, the resulting datasets will better represent the world’s population, enabling us to build more inclusive experiences.

In this post, we described a step toward that goal, how we learned frequencies of iconic scenes with formal DP assurance. This enabled us to improve key photo selection for Memories in iOS&nbsp;16, and Places in iOS&nbsp;17. This approach of applying privacy-preserving machine learning research to real-world problems fuels ML innovation and helps our users keep their data private.

# Acknowledgments

Many people contributed to this research, including Mike Chatzidakis, Junye Chen, Oliver Chick, Eric Circlaeays, Sowmya Gopalan, Yusuf Goren, Kristine Guo, Michael Hesse, Omid Javidbakht, Vojta Jina, Kalu Kalu, Anil Katti, Albert Liu, Richard Low, Audra McMillan, Joey Meyer, Steve Myers, Alex Palmer, David Park, Gianni Parsa, Paul Pelzl, Rehan Rishi, Michael Scaria, Chiraag Sumanth, Kunal Talwar, Karl Tarbe, Shan Wang, and Mayank Yadav.

# Apple Resources   

Apple. “Learning with Privacy at Scale." n.d. [[link.]](https://docs-assets.developer.apple.com/ml-research/papers/learning-with-privacy-at-scale.pdf)

Apple Support. “Use Memories in Photos on your iPhone, iPad, or iPod touch.” 2023. [[link.]](https://support.apple.com/en-us/HT207023)

McMillan, Audra, Omid Javidbakht, Kunal Talwar, Elliot Briggs, Mike Chatzidakis, Junye Chen, John Duchi, et al. 2022. “Private Federated Statistics in an Interactive Setting.” [[link.]](https://arxiv.org/abs/2211.10082)

“A Multi-Task Neural Architecture for On-Device Scene Analysis." 2022. Apple Machine Learning Research. [[link.]](https://machinelearning.apple.com/research/on-device-scene-analysis)

# External References  

Corrigan-Gibbs, Henry, and Dan Boneh. 2017. “Prio: Private, Robust, and Scalable Computation of Aggregate Statistics.” March 18, 2017. [[link.]](https://arxiv.org/abs/1703.06255)

Geoghegan, Tim, Christopher Patton, Eric Rescorla, and Christopher A. Wood. 2023. “Distributed Aggregation Protocol for Privacy Preserving Measurement.” IETF. July 10, 2023. [[link.]](https://datatracker.ietf.org/doc/draft-ietf-ppm-dap/)

In this article, we share how we apply differential privacy (DP) to learn about the kinds of photos people take at frequently visited locations (iconic scenes) without personally identifiable data leaving their device. This approach is used in several features in Photos, including choosing key photos for <a href="https://support.apple.com/en-us/HT207023" class="more">Memories,</a> and selecting key photos for locations in Places in iOS 17.

Learning Iconic Scenes with Differential Privacy 

<div class="callout">

## Schedule

### Saturday, August&nbsp;19
<ul class="links-stacked">
<li>WORKSHOP</li>
 <li><a href=https://sites.google.com/view/yfrsw-2023/>Young Female Researchers in Speech</a></li>
 <li>9:00 AM - 5:00 PM LT at the Trinity College Dublin O'Reilly Institute</li>
</ul>

### Monday, August&nbsp;21
<ul class="links-stacked">

<li>ORAL</li>
 <li><a href=/research/latent-phrase-matching>Latent Phrase Matching for Dysarthric Speech</a></li>
 <li>11:00 AM - 1:00 PM LT</li>
 <li>Colin Lea, Dianna Yee, Jaya Narain, Zifang Huang, Lauren Tooley, Jeffrey P. Bigham, Leah Findlater</li>
<li>POSTER</li>
 <li><a href=/research/unified-query-rewriting>5IDER: Unified Query Rewriting for 5teering, Intent Carryover, Disfluencies, Entity Carryover and Repair</a></li>
 <li>2:30 - 4:30 PM LT</li>
 <li>Jiarui Lu, Bo-Hsiang Tseng, Joel Ruben Antony Moniz, Site Li, Xueyun Zhu, Hong Yu, Murat Akbacak</li>
<li>POSTER</li>
 <li><a href=/research/nearest-neighbor-phrase>Approximate Nearest Neighbour Phrase Mining for Contextual Speech Recognition</a></li>
 <li>2:30 - 4:30 PM LT</li>
 <li>Maurits Bleeker, Pawel Swietojanski, Stefan Braun, Xiaodan Zhuang</li>
</ul>

### Tuesday, August&nbsp;22
<ul class="links-stacked">
<li>ORAL</li>
 <li><a href=/research/matching-latent-encoding>Matching Latent Encoding for Audio-Text based Keyword Spotting</a></li>
 <li>1:30 - 3:30 PM LT</li>
 <li>Kumari Nishu, Minsik Cho, Devang Naik</li>
<li>POSTER</li>
 <li><a href=/research/efficient-multimodal-neural-networks>Efficient Multimodal Neural Networks for Trigger-less Voice Assistants</a></li>
 <li>4:00 - 6:00 PM LT</li>
 <li>Sai Srujana Buddi, Utkarsh Oggy Sarawgi, Tashweena Heeramun, Karan Sawnhey, Ed Yanosik, Saravana Rathinam, Saurabh Adya</li>
</ul>

### Wednesday, August&nbsp;23
<ul class="links-stacked">
<li>POSTER</li>
 <li><a href=/research/spatial-librispeech>Spatial LibriSpeech: An Augmented Dataset for Spatial Audio Learning</a></li>
 <li>1:30 - 3:30 PM LT</li>
 <li>Miguel Sarabia, Elena Menyaylenko, Alessandro Toso, Skyler Seto, Zakaria Aldeneh, Shadi Pirhosseinloo, Luca Zappella, Barry-John Theobald, Nicholas Apostoloff, Jonathan Sheaffer</li>
</ul>
</div>

## Accepted Papers

[5IDER: Unified Query Rewriting for 5teering, Intent Carryover, Disfluencies, Entity Carryover and Repair](/research/unified-query-rewriting)

*Jiarui Lu, Bo-Hsiang Tseng, Joel Ruben Antony Moniz, Site Li, Xueyun Zhu, Hong Yu, Murat Akbacak*

[Approximate Nearest Neighbour Phrase Mining for Contextual Speech Recognition](/research/nearest-neighbor-phrase)

*Maurits Bleeker, Pawel Swietojanski, Stefan Braun, Xiaodan Zhuang*

[Efficient Multimodal Neural Networks for Trigger-less Voice Assistants](/research/efficient-multimodal-neural-networks)

*Sai Srujana Buddi, Utkarsh Oggy Sarawgi, Tashweena Heeramun, Karan Sawnhey, Ed Yanosik, Saravana Rathinam, Saurabh Adya*

[Latent Phrase Matching for Dysarthric Speech](/research/latent-phrase-matching)

*Colin Lea, Dianna Yee, Jaya Narain, Zifang Huang, Lauren Tooley, Jeffrey P. Bigham, Leah Findlater*

[Matching Latent Encoding for Audio-Text based Keyword Spotting](/research/matching-latent-encoding)

*Kumari Nishu, Minsik Cho, Devang Naik*

[Spatial LibriSpeech: An Augmented Dataset for Spatial Audio Learning](/research/spatial-librispeech)

*Miguel Sarabia, Elena Menyaylenko, Alessandro Toso, Skyler Seto, Zakaria Aldeneh, Shadi Pirhosseinloo, Luca Zappella, Barry-John Theobald, Nicholas Apostoloff, Jonathan Sheaffer*

## Acknowledgements

Zak Aldeneh, Panos Georgiou, Tuomo Raitio, Vineet Garg, and Dipjyoti Paul are reviewers for Interspeech 2023.

Apple is sponsoring the Interspeech Conference, which will take place in person from August 20 to 24 in Dublin, Ireland. Interspeech is the world’s largest and most comprehensive conference on the science and technology of spoken language processing. Below is the schedule of Apple-sponsored workshops and events at Interspeech 2023.

Interspeech Conference 2023

<div class="callout">

## Schedule

### Sunday, July&nbsp;23
<ul class="links-stacked">
<li>EXPO TALK</li>
<li><a href=https://icml.cc/Expo/Conferences/2023/talk%20panel/25687>Learning Iconic Scenes with Differential Privacy</a></li>
<li>3:00 - 4:00 PM LT in Theater 320 Emaline</li>
<li>Vojta Jina</li>
</ul>

### Monday, July&nbsp;24
<ul class="links-stacked">
 <li>WORKSHOP</li>
 <li><a href=https://www.latinxinai.org/icml-2023>LatinX in AI</a></li>
<li>8:00 AM - 6:00 PM LT in Meeting Room 315</li>	
<li>Tatiana Likhomanenko will participate in the mentoring hour during the workshop as a mentor.</li>
<li>TUTORIAL</li>
 <li><a href=https://icml.cc/virtual/2023/tutorial/21559>Optimal Transport in Learning, Control, and Dynamical Systems</a></li>
 <li>9:30 - 10:30 AM and 11:00 AM - 12:00 PM LT in Meeting Room 316 A-C</li>
 <li>Charlotte Bunne, Marco Cuturi</li>
</ul>

### Tuesday, July&nbsp;25
<ul class="links-stacked">
<li>POSTER</li>
 <li><a href=/research/nerf>NerfDiff: Single-image View Synthesis with NeRF-guided Distillation from 3D-aware Diffusion</a></li>
 <li>11:00 AM - 1:30 PM LT in Exhibit Hall 1 #231</li>
 <li>Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, Ravi Ramamoorthi</li>
 <li>POSTER</li>
 <li><a href=https://icml.cc/virtual/2023/poster/23737>The Monge Gap: A Regularizer to Learn All Transport Maps</a></li>
 <li>11:00 AM - 1:30 PM LT in Exhibit Hall 1 #333</li>
 <li>Théo Uscidda, Marco Cuturi</li>
<li>POSTER</li>
 <li><a href=/research/duet>DUET: 2D Structured and Approximately Equivariant Representations</a></li>
 <li>2:00 - 3:30 PM LT in Exhibit Hall 1 #235</li>
 <li>Xavi Suau, Federico Danieli, T. Anderson Keller, Arno Blaas, Chen Huang, Jason Ramapuram, Dan Busbridge, Luca Zappella</li>
<li>POSTER</li>
 <li><a href=https://icml.cc/virtual/2023/poster/24935>Stabilizing Transformer Training by Preventing Attention Entropy Collapse</a></li>
 <li>2:00 - 3:30 PM LT in Exhibit Hall 1 #503</li>
 <li>Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, Joshua M Susskind</li>
<li>POSTER</li>
 <li><a href=https://icml.cc/virtual/2023/poster/24269>Monge, Bregman and Occam: Interpretable Optimal Transport in High-Dimensions with Feature-Sparse Maps</a></li>
 <li>2:00 - 3:30 PM LT in Exhibit Hall 1 #431</li>
 <li>Marco Cuturi, Michal Klein, Pierre Ablin</li>
 <li>ORAL PRESENTATION</li>
 <li><a href=https://icml.cc/virtual/2023/oral/25520>Generalization on the Unseen, Logic Reasoning and Degree Curriculum</a></li>
 <li>6:50 PM LT in Meeting Room 316 A-C</li>
 <li>Emmanuel Abbe, Samy Bengio, Aryo Lotfi, Kevin Rizk</li>
</ul>

### Wednesday, July&nbsp;26
<ul class="links-stacked"><li>POSTER</li>
 <li><a href=/research/unconstrained-channel-pruning>UPSCALE: Unconstrained Channel Pruning</a></li>
 <li>11:00 AM - 12:30 PM LT in Exhibit Hall 1 #729</li>
 <li>Alvin Wan, Hanxiang Hao, Kaushik Patnaik, Yueyang Xu, Omer Hadad, David Güera, Zhile Ren, Qi Shan</li>
 <li>POSTER</li>
 <li><a href=/research/conformalization-sparse-generalized>Conformalization of Sparse Generalized Linear Models</a></li>
 <li>11:00 AM - 12:30 PM LT in Exhibit Hall 1 #707</li>
 <li>Etash Guha, Eugene Ndiaye, Xiaoming Huo</li>
<li>POSTER</li>
 <li><a href=https://icml.cc/virtual/2023/poster/23735>Near-Optimal Algorithms for Private Online Optimization in the Realizable Regime</a></li>
 <li>2:00 - 3:30 PM LT in Exhibit Hall 1 #801</li>
 <li>Hilal Asi, Vitaly Feldman, Tomer Koren, Kunal Talwar</li>
 <li>POSTER</li>
 <li><a href=https://icml.cc/virtual/2023/poster/24523>Generalization on the Unseen, Logic Reasoning and Degree Curriculum</a></li>
 <li>2:00 - 3:30 PM LT in Exhibit Hall 1 #433</li>
 <li>Emmanuel Abbe, Samy Bengio, Aryo Lotfi, Kevin Rizk</li>
<li>SOCIAL</li>
 <li><a href=https://icml.cc/virtual/2023/social/28470>Black in AI</a></li>
 <li>5:45 - 7:30 PM LT in Meeting Room 313</li>
</ul>

### Thursday, July&nbsp;27
<ul class="links-stacked"><li>POSTER</li>
 <li><a href=https://icml.cc/virtual/2023/poster/24314>Robustness in Multimodal Learning under Train-Test Modality Mismatch</a></li>
 <li>10:30 AM - 12:00 PM LT in Exhibit Hall 1 #717</li>
 <li>Brandon McKinzie, Vaishaal Shankar, Joseph Cheng, Yinfei Yang, Jonathon Shlens, Alexander Toshev</li>
<li>POSTER</li>
 <li><a href=/research/entropy-reconstruction>The Role of Entropy and Reconstruction for Multi-View Self-Supervised Learning</a></li>
 <li>10:30 AM - 12:00 PM LT in Exhibit Hall 1 #220</li>
 <li>Borja Rodríguez Gálvez, Arno Blaas, Pau Rodriguez, Adam Golinski, Xavi Suau, Jason Ramapuram, Dan Busbridge, Luca Zappella</li>
<li>POSTER</li>
 <li><a href=https://icml.cc/virtual/2023/poster/23804>Federated Online and Bandit Convex Optimization</a></li>
 <li>10:30 AM - 12:00 PM LT in Exhibit Hall 1 #440</li>
 <li>Kumar Kshitij Patel, Lingxiao Wang, Aadirupa Saha, Nati Srebro</li>
<li>POSTER</li>
 <li><a href=https://icml.cc/virtual/2023/poster/24344>From Robustness to Privacy and Back</a></li>
 <li>1:30 - 3:00 PM LT in Exhibit Hall 1 #438</li>
 <li>Hilal Asi, Jonathan Ullman, Lydia Zakynthinou</li>
</ul>

### Friday, July&nbsp;28
<ul class="links-stacked">
 <li>WORKSHOP</li>
 <li><a href=https://sites.google.com/view/imlh2023/>Interpretable Machine Learning in Healthcare</a></li>
 <li>9:00 AM - 5:00 PM LT in Ballroom C</li>
 <li>Agni Kumar is a co-organizer of this workshop.</li>

 <li>WORKSHOP</li>
 <li><a href=https://syns-ml.github.io/2023/>Synergy of Scientific and Machine Learning Modeling</a></li>
 <li>9:00 AM - 5:00 PM LT in Meeting Room 320</li>
 <li>Jörn-Henrik Jacobsen and Antoine Wehenkel are co-organizers of this workshop, Andrew C. Miller is an invited speaker/panelist for this workshop, and Tatiana Likhomanenko is a reviewer for this workshop.</li>

 <li>WORKSHOP</li>
 <li><a href=https://differentiable.xyz>Differentiable Almost Everything: Differentiable Relaxations, Algorithms, Operators, and Simulators</a></li>
 <li>9:00 AM - 5:00 PM LT in Meeting Room 310</li>
 <li>Marco Cuturi is a co-organizer of this workshop.</li>
 <li>WORKSHOP</li>
 <li><a href=https://klr-icml2023.github.io>Knowledge and Logical Reasoning in the Era of Data-driven Learning</a></li>
 <li>9:00 AM - 5:00 PM LT in Meeting Room 301</li>
 <li>Samy Bengio is an invited speaker</li>
 <li>WORKSHOP</li>
 <li><a href=https://sites.google.com/wimlworkshop.org/wiml-unworkshop-2023/home>Women in Machine Learning</a></li>
 <li>8:15 AM - 4:45 PM LT in Meeting Room 316C</li>
 <li>Rishika Agarwal and Ivy Zhang will host a roundtable on mentorships during the workshop.</li>
 <li>WORKSHOP</li>
 <li><a href=https://fl-icml2023.github.io>Federated Learning and Analytics in Practice: Algorithms, Systems, Applications, and Opportunities</a></li>
 <li>9:00 AM - 5:00 PM LT in Meeting Room 311</li>
<li>Jianyu Wang is a co-organizer of this workshop. Vojta Jina is an invited speaker.</li>
<li><a href=/research/population-expansion>Accepted Paper: Population Expansion for Training Language Models with Private Federated Learning</a></li>
<li>Tatsuki Koga, Congzheng Song, Martin Pelikan, Mona Chitnis</li>
 <li><a href=/research/differentially-private-heavy>Accepted Paper: Differentially Private Heavy Hitters using Federated Analytics</a></li>
<li>Karan Chadha, Junye Chen, John Duchi, Vitaly Feldman, Hanieh Hashemi, Omid Javidbakht, Audra McMillan, Kunal Talwar</li>
</ul>

### Saturday, July&nbsp;29
<ul class="links-stacked">
 <li>WORKSHOP</li>
 <li><a href=https://indigenousinai.org>Indigenous in AI</a></li>
 <li>9:00 AM - 5:00 PM LT in Meeting Room 316C</li>
 <li>This is a joint workshop with Queer in AI.</li>
<li>Sinead Williamson is a co-organizer of this workshop.</li>
</ul>

</div>

## Accepted Papers

[Conformalization of Sparse Generalized Linear Models](/research/conformalization-sparse-generalized)

*Etash Guha, Eugene Ndiaye, Xiaoming Huo*

[DUET: 2D Structured and Approximately Equivariant Representations](/research/duet)

*Xavi Suau, Federico Danieli, T. Anderson Keller, Arno Blaas, Chen Huang, Jason Ramapuram, Dan Busbridge, Luca Zappella*

[From Robustness to Privacy and Back](/research/from-privacy-robustness)

*Hilal Asi, Jonathan Ullman, Lydia Zakynthinou*

[Generalization on the Unseen, Logic Reasoning and Degree Curriculum](/research/generalization-on-the-unseen)

*Emmanuel Abbe, Samy Bengio, Aryo Lotfi, Kevin Rizk*

*Congratulations to the team for receiving the Outstanding Paper Award at ICML 2023.*

[Monge, Bregman and Occam: Interpretable Optimal Transport in High-Dimensions with Feature-Sparse Maps](/research/monge-bregman-occam)

*Marco Cuturi, Michal Klein, Pierre Ablin*

[Near-Optimal Algorithms for Private Online Optimization in the Realizable Regime](/research/near-optimal-algorithms)

*Hilal Asi, Vitaly Feldman, Tomer Koren, Kunal Talwar*

[NerfDiff: Single-image View Synthesis with NeRF-guided Distillation from 3D-aware Diffusion](/research/nerf)

*Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, Ravi Ramamoorthi*

[Robustness in Multimodal Learning under Train-Test Modality Mismatch](/research/robustness-multimodal-learning)

*Brandon McKinzie, Vaishaal Shankar, Joseph Cheng, Yinfei Yang, Jonathon Shlens, Alexander Toshev*

[Stabilizing Transformer Training by Preventing Attention Entropy Collapse](/research/stabilizing-transformer-training)

*Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, Joshua M Susskind*

[The Monge Gap: A Regularizer to Learn All Transport Maps](/research/monge-map)

*Théo Uscidda, Marco Cuturi*

[The Role of Entropy and Reconstruction for Multi-View Self-Supervised Learning](/research/entropy-reconstruction)

*Borja Rodríguez Gálvez, Arno Blaas, Pau Rodriguez, Adam Golinski, Xavier Suau, Jason Ramapuram, Dan Busbridge, Luca Zappella*

[UPSCALE: Unconstrained Channel Pruning](/research/unconstrained-channel-pruning)

*Alvin Wan, Hanxiang Hao, Kaushik Patnaik, Yueyang Xu, Omer Hadad, David Güera, Zhile Ren, Qi Shan*

## Workshop Accepted Papers

[Differentially Private Heavy Hitters using Federated Analytics](/research/differentially-private-heavy)

*Karan Chadha, Junye Chen, John Duchi, Vitaly Feldman, Hanieh Hashemi, Omid Javidbakht, Audra McMillan, Kunal Talwar*

[Population Expansion for Training Language Models with Private Federated Learning](/research/population-expansion)

*Tatsuki Koga, Congzheng Song, Martin Pelikan, Mona Chitnis*

## Demos 

**Learning Iconic Scenes with Differential Privacy**

This demo allows users to browse heat maps of popular locations and histograms of the photos taken at each location. Learn more about the research behind the application by reading the paper, [Private Federated Statistics in an Interactive Setting](https://arxiv.org/abs/2211.10082).

**Object Capture**

Object Capture allows users to generate high‑quality 3D models from a series of photos taken on an iPhone or iPad. Using CVML algorithms, Object Capture can analyze the overlap area between different images to match up landmarks, and then produce a 3D model of a photographed object. These models can then be viewed in AR Quick Look, seamlessly integrated into an Xcode project, or used in professional 3D content workflows.

All ICML attendees are invited to stop by the Apple booth (booth number 101, located in Hall III of the Hawai’i Convention Center) to experience these demos in person.

## Acknowledgements

Samy Bengio is a Senior Area Chair for ICML 2023.

Ronan Collobert and Navdeep Jaitly are Area Chairs for ICML 2023.	

Sinead Williamson is a Diversity and Inclusion Chair for ICML 2023.

Marco Cuturi is a Communications Chair for ICML 2023.

Tatiana Likhomanenko, Rin Susa, Barry Theobald, Pau Rodriguez, Vimal Thilak, Pierre Ablin, Hilal Asi, Samira Abnar, Chen Huang, Vaishaal Shankar, Miguel Sarabia, Antoine Wehenkel, Agni Kumar, Bogdan Mazoure, Preetum Nakkiran, Rishika Agarwal, Yinfei Yang, Amir Shafiq, Jierui Lin, and Xavi Suau are reviewers for ICML 2023.

Apple is sponsoring the International Conference on Machine Learning (ICML), which will take place in person from July 23 to 29 in Honolulu, Hawai'i. ICML is a conference dedicated to advancing artificial intelligence, with over 6,000 attendees. Below is the schedule of Apple-sponsored workshops and events at ICML 2023.

International Conference on Machine Learning (ICML) 2023 

Apple machine learning teams are engaged in state of the art research in machine learning and artificial intelligence. Learn about the latest advancements.

Overview

We present a novel differentiable rendering framework for joint geometry, material, and lighting estimation from multi-view images. In contrast to previous methods which assume a simplified environment map or co-located flashlights, in this work, we formulate the lighting of a static scene as one neural incident light field (NeILF) and one outgoing neural radiance field (NeRF). The key insight of the proposed method is the union of the incident and outgoing light fields through physically-based rendering and inter-reflections between surfaces, making it possible to disentangle the scene geometry, material, and lighting from image observations in a physically-based manner. The proposed incident light and inter-reflection framework can be easily applied to other NeRF systems. We show that our method can not only decompose the outgoing radiance into incident lights and surface materials, but also serve as a surface refinement module that further improves the reconstruction detail of the neural surface. We demonstrate on several datasets that the proposed method is able to achieve state-of-the-art results in terms of the geometry reconstruction quality, material estimation accuracy, and the fidelity of novel view rendering.

NeILF++: Inter-Reflectable Light Fields for Geometry and Material Estimation

Large Language Models (LLMs) have made substantial progress in the past several months, shattering state-of-the-art benchmarks in many domains. This paper investigates LLMs' behavior with respect to gender stereotypes, a known stumbling block for prior models. We propose a simple paradigm to test the presence of gender bias, building on but differing from WinoBias, a commonly used gender bias dataset which is likely to be included in the training data of current LLMs. We test four recently published LLMs and demonstrate that they express biased assumptions about men and women, specifically those aligned with people's perceptions, rather than those grounded in fact. We additionally study the explanations provided by the models for their choices. In addition to explanations that are explicitly grounded in stereotypes, we find that a significant proportion of explanations are factually inaccurate and likely obscure the true reason behind the models' choices. This highlights a key property of these models: LLMs are trained on unbalanced datasets; as such, even with reinforcement learning with human feedback, they tend to reflect those imbalances back at us. As with other types of societal biases, we suggest that LLMs must be carefully tested to ensure that they treat minoritized individuals and communities equitably.

Machine Learning
Research at Apple

Recent research

NeILF++: Inter-Reflectable Light Fields for Geometry and Material Estimation

Gender Bias in LLMs

Research highlights

Voice Trigger System for Siri

Learning Iconic Scenes with Differential Privacy

All events

Interspeech Conference 2023

International Conference on Machine Learning (ICML) 2023

Discover opportunities in Machine Learning.

Machine LearningResearch at Apple

Recent research

NeILF++: Inter-Reflectable Light Fields for Geometry and Material Estimation

Gender Bias in LLMs

Research highlights

Voice Trigger System for Siri

Learning Iconic Scenes with Differential Privacy

All events

Interspeech Conference 2023

International Conference on Machine Learning (ICML) 2023

Discover opportunities in Machine Learning.

Machine Learning
Research at Apple