The Vienna Tourist Board supports you with a wide range of services both before and during your trip to Vienna.

This talk has been postponed. We will update the schedule when we have a new time. We will also push a message to the Whova app.

In the rapidly evolving field of speech generative models, there is a pressing need to ensure audio authenticity against the risks of voice cloning. We present AudioSeal, the first audio watermarking technique designed specifically for localized detection of AI-generated speech. AudioSeal employs a generator/detector architecture trained jointly with a localization loss to enable localized watermark detection up to the sample level, and a novel perceptual loss inspired by auditory masking, that enables AudioSeal to achieve a state of art imperceptibility and robustness.

LLM Evaluators, RAG, and Quantum ML in Financial Services: An insider view of how AI is set up to change the banking industry

Tomorrow’s bank is digital. It’s why we value innovative and creative thinkers in technology who can help us deliver the next generation of banking services, for HSBC and our customers around the world.

With recent advancements in Generative AI and Quantum Computing, we have been actively testing applications of these technologies in banking scenarios. In this talk, you’ll hear the latest results from AI and technology leadership, concluding with a panel discussion on the future of AI in the sector.

Quantum Machine Learning (ML) solutions for financial classification tasks will be discussed alongside approaches for testing large language models (LLMs) on document-focused tasks within the bank, including LLM evaluators in the financial services space.

Responsible AI is at the core of our ambitions to Digitise at Scale. Our panel discussion will also cover the latest from our research and reporting on the impact of AI.

This presentation will cover a variety of work at the intersection of graph representation learning and AI being done at Google. It will provide a general overview of graph neural networks & LLMs and then go into 3 areas that we think will be of interest to a general ML audience, including:

Encoding of Graphs as Text for GenAI models [1]. This will cover insights on how best to encode structured data, such as graphs, for LLMs and other GenAI models. Will cover results Graph Structure Learning [2,3]. Will cover work on learning the best graph structure for a given dataset. Graph Foundation Models [4,5]. Will cover more complex models, such as structure encoding functions, which can learn the best representation of data for LLMs Theoretical Connections between GNNs and Transformer [6]. Will briefly cover our results on the complexity of graph algorithms in Transformer architecture & the insights derived from this.

References:

[1] Talk Like a Graph: Encoding Graphs for Large Language Models https://arxiv.org/pdf/2310.04560.pdf

[2] Grale: Designing networks for graph learning https://arxiv.org/pdf/2007.12002.pdf

[3] UGSL: A unified framework for benchmarking graph structure learning https://arxiv.org/pdf/2308.10737.pdf

[4] Let Your Graph Do the Talking: Encoding Structured Data for LLMs https://arxiv.org/pdf/2402.05862.pdf

[5] Don't Forget to Connect! Improving RAG with Graph-based Reranking https://arxiv.org/abs/2405.18414

[6] Understanding Transformer Reasoning Capabilities via Graph Algorithms https://arxiv.org/abs/2405.18512

AutoGluon, the most popular open-source library published by the Amazon Science team, is a state-of-the-art toolkit designed to make automated machine learning (AutoML) accessible and powerful across diverse data types, including tabular, text, image, and time series powered by foundational models. During this presentation, we introduce the latest advancements in AutoGluon, highlighting the Chronos model (3.4MM downloads in 1 month), foundation models for forecasting. Chronos significantly enhances the accuracy and efficiency of time series predictions with a 60% win-rate improvement over the previous version. AutoGluon 1.1 also brings major improvements to deep learning model automation, ease of use, and performance, particularly for large datasets. Rather than diving deep into the mechanisms underlining each individual ML models, we emphasize on how you can take advantage of a diverse collection of models to build an automated ML pipeline. Join us to explore how AutoGluon can solve real-world problems with just three lines of code and discover the cutting-edge techniques that make it a leading AutoML toolkit for researchers and practitioners.

Private Federated Learning (PFL) is an approach to collaboratively train a machine learning model between edge devices with coordination by a central server, whilst preserving the privacy of the data on each edge device. PFL is an emerging field, with exponential growth in the number of papers published over the past few years and several big tech companies invest heavily in the practical applications of PFL. Researchers commonly perform experiments in a simulation environment to quickly iterate on PFL ideas. However, previous open-source tools do not offer the efficiency required to simulate FL on larger and more realistic FL datasets. We introduce pfl-research (https://github.com/apple/pfl-research), a fast, modular, and easy-to-use Python framework for simulating FL and PFL, 7-72x faster than alternative open-source frameworks. In this talk, we will start by briefly introducing Federated Learning as a subject and describe techniques to preserve the privacy of participating edge devices. We will quickly dive deeper into the unique challenges encountered in PFL, explore research problems that persist today. Then we will introduce pfl-research, its features and performance. We demonstrate how researchers can use pfl-research to significantly boost their productivity with intuitive interfaces and fast distributed simulations, including fine-tuning LLMs in PFL.

Recent advances in natural language processing (NLP) have led to the development of Large Language Models (LLMs), which can generate text virtually indistinguishable from human written texts. These models are adept at interpreting and executing human commands, facilitating their incorporation into a broad spectrum of mainstream applications, including summarization systems, interactive chatbots, and virtual assistants. Amazon, for instance, integrates LLMs to enhance customer experiences, utilizing them to summarize customer reviews on its website and to power innovative shopping chatbots like Rufus. Moreover, Amazon Web Services (AWS) has introduced Amazon Q, a LLM-powered assistant for commercial applications. However, studies have revealed that text generated by LLMs can frequently exhibit factual inconsistencies such as contradictions with the provided input or hallucinations that are irrelevant to the context at hand. Identifying inconsistencies in text produced by LLMs poses a significant challenge because these errors often align with the task's overarching structure and theme, rendering them subtle and hard to detect. Thus, there is a need for an objective evaluation of text generated from LLMs. Human evaluation is considered the gold standard, but it requires subject expertise and is time-consuming, making it non-scalable. This underscores the need for automated metrics to efficiently evaluate LLM-generated texts. In this talk, we will explore different automated evaluation strategies from the literature. These methods range from simple text similarity based methods to using LLMs for evaluation. We will also discuss various ways to benchmark evaluate these automated evaluation methods and cover publicly available datasets for evaluation.

"On-device machine learning has been challenging, as it not only has more constraints (power, memory, thermal) compared to the cloud server, but the quality and efficiency bar can be relatively high. With the recent trend of larger language models, it further pushes the limit of the hardware.

ExecuTorch, a native on-device solution from PyTorch, is designed in such a way that researchers and ML engineers can fully optimize the models to the target hardware within the PyTorch ecosystem. The design principles are:

The Executor is focused on execution. Whereas obvious, this means we won’t tag any other intent to the program representation nor to the runtime’s APIs, such as human readability of the program or recomputing a memory plan. Any extra concerns are thus implemented as support libraries and not on the executable program. The program is unidirectional and the runtime’s program immutable. The program starts from the source code, and any change to the runtime program has to start again from the source or some other higher level representation. It basically works as a binary. The runtime’s behavior is entirely defined by (a small) instruction set. Any and all capabilities are determined by the runtime’s instruction set. The runtime makes no decisions. Its behavior is pre-defined by the instructions it supports and the program it runs. Debugging and program understanding is handled by specific tools to analyze the executable program according to its original source (i.e. the pytorch source).

We already enable a list of models (See appendix 1) in one or more backends. And we showcase SOTA CPU performance for llama2 on device, and early support with promising numbers for llama3 (See appendix 2). Moving forward, we target to enable more models, more LLMs, multi modality, and further accelerate LLMs by lowering to hardware accelerators.

We’d like to walk through the colab notebook to demonstrate how to export the model and deploy on device."

"Generative AI, combined with traditional statistical and causal techniques, is transforming the monitoring and observability domain by enabling proactive and predictive system management. These methods enhance anomaly detection, automate root cause analysis, and extract intelligent insights from vast monitoring data. This talk will detail how Amazon is developing solutions that merge the robustness of traditional ML techniques with the flexibility of modern generative methods to address monitoring challenges. The talk will cover three key areas: 1) How advanced statistical and forecasting methods improve anomaly detection in timeseries and log data, 2) How causal techniques help isolate the root causes of application issues, and 3) How generative AI can derive insights to anticipate problems before they affect users, ensuring smoother operations and better performance in complex, dynamic environments.

The talk will be based on published work in recent ICML, NeurIPS, and AISTATS conferences"

"Over a period of just about 5 years, the use of AI-based tools for software engineering has gone from being a very promising research investigation to indispensable features in modern developer environments. At Google, we have been developing and deploying AI-based tools to surfaces where Google engineers spend the majority of their time, including inner loop activities such as code authoring, review and search, as well as outer loop ones such as bug management and planning. Improvements in these surfaces are monitored carefully for productivity and developer satisfaction.

This talk will present AI-powered improvements and continuing transformation of Google’s internal software development. We will touch upon the challenges in how to align our internal efforts with the very fast moving field of LLMs, and what challenges we have faced to bridge the gap from research to real products with usage at scale. We need to constantly make judgment calls on technical feasibility, the possibility of iterative improvement and the measurability of impact as we decide what ideas to pursue for production level adaptation and adoption. The talk will go into several examples of this that we have gone through in the recent past, and what we have learned in the process.

We will demo some of the generative AI based projects for software engineering at Google. In particular, we will show how we weave a conversational agent in an IDE, which combines convenient workflows for information gathering, conversations about code, and code transformation abilities.

We will conclude the talk with a discussion of opportunities we see for the next five years and some thoughts on how the community can collaborate better by focusing on good benchmarks."