Events & Conferences

Alexa enters the “age of self”

Published

4 years ago

June 2, 2021

Alexa launched in 2014, and in the more than six years since, we’ve been making good on our promise to make Alexa smarter every day. In addition to foundational improvements in Alexa’s core AI technologies, such as speech recognition and natural-language-understanding systems, Alexa scientists have developed technologies that continue to delight our customers, such as whispered speech and Alexa’s new live translation service.

Prem Natarajan, Alexa AI vice president of natural understanding, at a conference in 2018.

But some of the technologies we’ve begun to introduce, together with others we’re now investigating, are harbingers of a step change in Alexa’s development — and in the field of AI itself. Collectively, these technologies will bring a new level of generalizability and autonomy to both the Alexa voice service and the tools available to Alexa developers, ushering in what I like to think of as a new “age of self” in artificial intelligence, an age in which AI systems such as Alexa become more self-aware and more self-learning, and in which they lend themselves to self-service by experienced developers and even end users.

By self-awareness, I mean the ability to maintain an awareness of ambient state (e.g., time of day, thermostat readings, and recent actions) and to employ commonsense reasoning to make inferences that reflect that awareness and prior/world knowledge. Alexa hunches can already recognize anomalies in customers’ daily routines and suggest corrections — noticing that a light was left on at night and offering to turn it off, for instance. Powered by commonsense reasoning, self-awareness goes further: for instance, if a customer turns on the television five minutes before the kids’ soccer practice is scheduled to end, an AI of the future might infer that the customer needs a reminder about pickup.

In the “age of self”, AIs will be able to infer customers’ implicit intentions from observable temporal patterns, such as interactions with smart-home devices like thermostats, door locks, and lights.

Self-learning is Alexa’s ability to improve and expand its capabilities without human intervention. And like self-awareness, self-learning employs reasoning: for example, does the customer’s response to an action indicate dissatisfaction with that action? Similarly, when a customer issues an unfamiliar command, a truly self-learning Alexa would be able to infer what it might mean — perhaps by searching the web or exploring a knowledge base — and suggest possibilities.

Self-service means, essentially, the democratization of AI. Alexa customers with no programming experience should be able to customize Alexa’s services and even create new Alexa capabilities, and skill developers without machine learning experience should be able to build complex yet robust conversational skills. Colloquially, these are the conversational-AI equivalents of no-code and low-code development environments.

To be clear, the age of self is not yet upon us, and its dawning will require the maturation of technologies still under development, at Amazon and elsewhere. But some of Alexa’s recently launched capabilities herald a lightening in the Eastern sky.

Self-awareness

In 2018, we launched Alexa hunches for the smart home, with Alexa suggesting actions to take in response to anomalous sensor data. By early 2021, the science has advanced adequately for us to launch an opt-in service in which Alexa can take action immediately and automatically. In the meantime, we’ve also been working to expand hunches to Alexa services other than the smart home.

Technologies will bring a new level of generalizability and autonomy to both the Alexa voice service and the tools available to Alexa developers, ushering in what I like to think of as a new ‘age of self’ in artificial intelligence.

But commonsense reasoning requires something more — the ability to infer customers’ implicit intentions from observable temporal patterns. For instance, what does it mean if the customer turns down the thermostat, turns out the lights, locks the front door, and opens the garage? What if the customer initiates an interaction with a query like “Alexa, what’s playing at Rolling Hills Cine Plaza?”

In 2020, we took steps toward commonsense reasoning with a new Alexa function that can infer a customer’s latent goal— the ultimate aim that lies behind a sequence of requests. When a customer asks for the weather at the beach, for instance, Alexa might use that query, in combination with other contextual information, to infer that the customer may be interested in a trip to the beach. Alexa could then offer the current driving time to the beach.

To retrieve that information, Alexa has to know to map the location of the weather request to the destination variable in the route-planning function. This illustrates another aspect of self-awareness: the ability to track information across contexts.

That ability is at the core of the night-out experience we’ve developed, which engages the customer in a multiturn conversation to plan a complete night out, from buying movie tickets to making restaurant and ride-share reservations. The night-out experience tracks times and locations across skills, revising them on the fly as customers evaluate different options. To build the experience, we leveraged the machinery of Alexa Conversations, a service that enables developers to quickly and easily create dialogue-driven skills, and we drew on our growing body of research on dialogue state tracking.

_{Dialogue states at several successive dialogue turns}

Self-awareness, however, includes an understanding not only of the conversational context but also of the customer’s physical context. In 2020, we demonstrated natural turn-taking on Alexa-enabled devices with cameras. When multiple speakers are engaging with Alexa, Alexa can use visual cues to distinguish between speech the customers are directing at each other and speech they’re directing at Alexa. In ongoing work, we’re working to expand this functionality to devices without cameras, by relying solely on acoustic and linguistic signals.

Finally, self-awareness also entails the capacity for self-explanation. Today, most machine learning models are black boxes; even their creators have no idea how they’re doing what they do. That uncertainty has turned explainable or interpretable AI into a popular research topic.

Amazon actively publishes on explainable-AI topics. In addition, the Alexa Fund, an Amazon venture capital investment program, invested in fiddler.ai, a startup that uses techniques based on the game-theoretical concept of Shapley values to do explainable AI.

Self-learning

Historically, the AI development cycle has involved collection of data, annotation of that data, and retraining of models on the newly annotated data — all of which add up to a laborious process.

In 2019, we launched Alexa’s self-learning system, which automatically learns to correct errors — both customer errors and errors in Alexa’s language-understanding models — without human involvement. The system relies on implicit signals that a request was improperly handled, as when a customer interrupts a response and rephrases the same request.

Alexa’s self-learning system models customer interactions with Alexa as sequences of states; different customer utterances (u₀, u₁, u₂) can correspond to the same state (h₀). The final state of a sequence, known as the “absorbing state”, indicates the success (checkmark) or failure (X) of a transaction.
Stacy Reilly

Currently, that fully automatic system is correcting 15% of defects. But those are defects that occur across a spectrum of users; only when enough people implicitly identify the same flaw does the system address it. We are working to adapt the same machinery to individual customers’ preferences — so that, for instance, Alexa can learn that when a particular customer asks for the song “Wow”, she means not the Post Malone hit from 2019 but the 1978 Kate Bush song.

Customers today also have the option of explicitly teaching Alexa their preferences. In the fall of 2020, we launched interactive teaching by customers, a capability that enables customers to instruct Alexa how they want certain requests to be handled. For instance, the customer can teach Alexa that the command “reading mode” means lights turned all the way up, while “movie mode” means only twenty percent up.

Self-service

Interactive teaching is also an early example of how Alexa is enabling more self-service. It extends prior Alexa features, like blueprints, which let customers build their own simple skills from preexisting templates, and routines, which let customers chain together sequences of actions under individual commands.

In March 2021, we announced the public release of Alexa Conversations, which allows developers to create dialogue-driven skills by uploading sample dialogues. Alexa Conversations’ sophisticated machine learning models use those dialogues as templates for generating larger corpora of synthetic training data. From that data, Alexa Conversations automatically trains a machine learning model.

Alexa Conversations does, however, require the developer to specify the set of entities that the new model should act upon and an application programming interface for the skill. So while it requires little familiarity with machine learning, it assumes some programming experience.

An Alexa feature known as catalogue value suggestions suggests entity names to skill developers on the basis of their “embeddings”, or locations in a representational space. If the embeddings of values (such as bird, dog, or cat) for a particular entity type are close enough (dotted circles) to their averages (solid circle and square), the system suggests new entity names; otherwise, it concludes that suggestions would be unproductive.

We are steadily chipping away at even that requirement, by making development for Alexa easier and more intuitive. As Alexa’s repertory of skills grows, for instance, entities are frequently reused, and we already have systems that can inform developers about entity types that they might not have thought to add to their skills. This is a step toward a self-service model in which developers no longer have to provide exhaustive lists of entities — or, in some cases, any entities at all.

Another technique that makes it easier to build machine learning models is few-shot learning, in which an existing model is generalized to a related task using only a handful of new training examples. This is an active area of research at Alexa: earlier this year, for example, we presented a paper at the Spoken Language Technologies conference that described a new approach to few-shot learning for natural-language-understanding tasks. Compared to its predecessors, our approach reduced the error rate on certain natural-language-understanding tasks by up to 12.4%, when each model was trained on only 10 examples.

These advances, along with the others reported on Amazon Science, demonstrate that the Alexa AI team continues to accelerate its pace of invention. More exciting announcements lie just over the horizon. I’ll be stopping back here every once in a while to update you on Alexa’s journey into the age of self.

Source link

Events & Conferences

A New Ranking Framework for Better Notification Quality on Instagram

Published

2 days ago

September 2, 2025

Xian Sun

We’re sharing how Meta is applying machine learning (ML) and diversity algorithms to improve notification quality and user experience.
We’ve introduced a diversity-aware notification ranking framework to reduce uniformity and deliver a more varied and engaging mix of notifications.
This new framework reduces the volume of notifications and drives higher engagement rates through more diverse outreach.

Notifications are one of the most powerful tools for bringing people back to Instagram and enhancing engagement. Whether it’s a friend liking your photo, another close friend posting a story, or a suggestion for a reel you might enjoy, notifications help surface moments that matter in real time.

Instagram leverages machine learning (ML) models to decide who should get a notification, when to send it, and what content to include. These models are trained to optimize for user positive engagement such as click-through-rate (CTR) – the probability of a user clicking a notification – as well as other metrics like time spent.

However, while engagement-optimized models are effective at driving interactions, there’s a risk that they might overprioritize the product types and authors someone has previously engaged with. This can lead to overexposure to the same creators or the same product types while overlooking other valuable and diverse experiences.

This means people could miss out on content that would give them a more balanced, satisfying, and enriched experience. Over time, this can make notifications feel spammy and increase the likelihood that people will disable them altogether.

The real challenge lies in finding the right balance: How can we introduce meaningful diversity into the notification experience without sacrificing the personalization and relevance people on Instagram have come to expect?

To tackle this, we’ve introduced a diversity-aware notification ranking framework that helps deliver more diverse, better curated, and less repetitive notifications. This framework has significantly reduced daily notification volume while improving CTR. It also introduces several benefits:

The extensibility of incorporating customized soft penalty (demotion) logic for each dimension, enabling more adaptive and sophisticated diversity strategies.
The flexibility of tuning demotion strength across dimensions like content, author, and product type via adjustable weights.
The integration of balancing personalization and diversity, ensuring notifications remain both relevant and varied.

The Risks of Notifications without Diversity

The issue of overexposure in notifications often shows up in two major ways:

Overexposure to the same author: People might receive notifications that are mostly about the same friend. For example, if someone often interacts with content from a particular friend, the system may continue surfacing notifications from that person alone – ignoring other friends they also engage with. This can feel repetitive and one-dimensional, reducing the overall value of notifications.

Overexposure to the same product surface: People might mostly receive notifications from the same product surface such as Stories, even when Feed or Reels could provide value. For example, someone may be interested in both reel and story notifications but has recently interacted more often with stories. Because the system heavily prioritizes past engagement, it sends only story notifications, overlooking the person’s broader interests.

Introducing Instagram’s Diversity-Aware Notification Ranking Framework

Instagram’s diversity-aware notification ranking framework is designed to enhance the notification experience by balancing the predicted potential for user engagement with the need for content diversity. This framework introduces a diversity layer on top of the existing engagement ML models, applying multiplicative penalties to the candidate scores generated by these models, as figure1, below, shows.

The diversity layer evaluates each notification candidate’s similarity to recently sent notifications across multiple dimensions such as content, author, notification type, and product surface. It then applies carefully calibrated penalties—expressed as multiplicative demotion factors—to downrank candidates that are too similar or repetitive. The adjusted scores are used to re-rank the candidates, enabling the system to select notifications that maintain high engagement potential while introducing meaningful diversity. In the end, the quality bar selects the top-ranked candidate that passes both the ranking and diversity criteria.

Figure.1: Instagram’s diversity-aware ranking framework where the diversity layer sits on top of the existing modeling layer and penalizes notifications that are too similar to recently sent ones.

Mathematical Formulation

Within the diversity layer, we apply a multiplicative demotion factor to the base relevance score of each candidate. Given a notification candidate 𝑐, we compute its final score as the product of its base ranking score and a diversity demotion multiplier:

$\text{Score}(c) = R(c) \times D(c)$

where R(c) represents the candidate’s base relevance score, and D(c) ∈ [0,1] is a penalty factor that reduces the score based on similarity to recently sent notifications. We define a set of semantic dimensions (e.g., author, product type) along which we want to promote diversity. For each dimension i, we compute a similarity signal p_i(c) between candidate c and the set of historical notifications H, using a maximal marginal relevance (MMR) approach:

$p_i(c) = \mathrm{max}_{h \in H}\mathrm{sim}_i(c, h)$

where sim_i(·,·) is a predefined similarity function for dimension i. In our baseline implementation, p_i(c) is binary: it equals 1 if the similarity exceeds a threshold 𝜏_i and 0 otherwise.

The final demotion multiplier is defined as:

$D(c) = \prod_{i=1}^{m} \left( 1 - w_i \cdot p_i(c) \right)$

where each w_i∈ [0,1] controls the strength of demotion for its respective dimension. This formulation ensures that candidates similar to previously delivered notifications along one or more dimensions are proportionally down-weighted, reducing redundancy and promoting content variation. The use of a multiplicative penalty allows for flexible control across multiple dimensions, while still preserving high-relevance candidates.

The Future of Diversity-Aware Ranking

As we continue evolving our notification diversity-aware ranking system, a next step is to introduce more adaptive, dynamic demotion strategies. Instead of relying on static rules, we plan to make demotion strength responsive to notification volume and delivery timing. For example, as a user receives more notifications—especially of similar type or in rapid succession—the system progressively applies stronger penalties to new notification candidates, effectively mitigating overwhelming experiences caused by high notification volume or tightly spaced deliveries.

Longer term, we see an opportunity to bring large language models (LLMs) into the diversity pipeline. LLMs can help us go beyond surface-level rules by understanding semantic similarity between messages and rephrasing content in more varied, user-friendly ways. This would allow us to personalize notification experiences with richer language and improved relevance while maintaining diversity across topics, tone, and timing.

Source link

Events & Conferences

Simplifying book discovery with ML-powered visual autocomplete suggestions

Published

2 days ago

September 2, 2025

Mao Sheng Liu

Every day, millions of customers search for books in various formats (audiobooks, e-books, and physical books) across Amazon and Audible. Traditional keyword autocomplete suggestions, while helpful, usually require several steps before customers find their desired content. Audible took on the challenge of making book discovery more intuitive and personalized while reducing the number of steps to purchase.

We developed an instant visual autocomplete system that enhances the search experience across Amazon and Audible. As the user begins typing a query, our solution provides visual previews with book covers, enabling direct navigation to relevant landing pages instead of the search result page. It also delivers real-time personalized format recommendations and incorporates multiple searchable entities, such as book pages, author pages, and series pages.

1 of 2

Audible’s visual-autocomplete experience.

2 of 2

Amazon’s visual-autocomplete experience.

Our system needed to understand user intent from just a few keystrokes and determine the most relevant books to display, all while maintaining low latency for millions of queries. Using historical search data, we match keystrokes to products, transforming partial inputs into meaningful search suggestions. To ensure quality, we implemented confidence-based filtering mechanisms, which are particularly important for distinguishing between general queries like “mystery” and specific title searches. To reflect customers’ most recent interests, the system applies time-decay functions to long historical user interaction data.

Events & Conferences

Revolutionizing warehouse automation with scientific simulation

Published

1 week ago

August 26, 2025

Deniz Akyildiz

Modern warehouses rely on complex networks of sensors to enable safe and efficient operations. These sensors must detect everything from packages and containers to robots and vehicles, often in changing environments with varying lighting conditions. More important for Amazon, we need to be able to detect barcodes in an efficient way.